WO2022183335A1

WO2022183335A1 - Image encoding and decoding methods, encoder, decoder, and storage medium

Info

Publication number: WO2022183335A1
Application number: PCT/CN2021/078522
Authority: WO
Inventors: 虞露; 周胜辉; 邵宇超; 于化龙; 戴震宇
Original assignee: 浙江大学; Oppo广东移动通信有限公司
Priority date: 2021-03-01
Filing date: 2021-03-01
Publication date: 2022-09-09
Also published as: CN116982082A

Abstract

The present application provides image encoding and decoding methods, an encoder, a decoder, and a storage medium. The image encoding method comprises: acquiring a current image to be encoded; inputting the current image into a neural network to obtain feature data of the current image, the feature data of the current image comprising feature data of N channels; quantizing the feature data of at least one of the N channels; and encoding the quantized feature data of the at least one channel to obtain a code stream, the code stream comprising first information, and the first information being used for instructing a decoding point to perform inverse quantization on the feature data of at least one of the N channels. In this way, the quantization of feature data output by an intermediate layer of a neural network is achieved, so that the technology in the existing video and image encoding and decoding standards can be reused to encode the feature data, thereby improving the encoding efficiency.

Description

Image coding and decoding method, encoder, decoder and storage medium

technical field

The present application relates to the technical field of video encoding and decoding, and in particular, to an image encoding and decoding method, an encoder, a decoder, and a storage medium.

Background technique

Digital video technology can be incorporated into a variety of video devices, such as digital televisions, smartphones, computers, e-readers or video players, and the like. With the development of video technology, the amount of data included in video data is relatively large. In order to facilitate the transmission of video data, video devices implement video compression technology to enable more efficient transmission or storage of video data.

With the rapid development of visual analysis technology, a video coding framework for machine vision is proposed by combining neural network technology with image and video compression technology.

However, the current video coding framework for machine vision has low coding efficiency.

SUMMARY OF THE INVENTION

Embodiments of the present application provide an image encoding and decoding method, an encoder, a decoder, and a storage medium, so as to improve encoding efficiency.

In a first aspect, the present application provides an image encoding method, including:

Get the current image to be encoded;

Inputting the current image into a neural network to obtain feature data of the current image, where the feature data of the current image includes feature data of N channels, and N is a positive integer;

quantifying the feature data of at least one channel in the N channels;

Encoding the quantized feature data of the at least one channel to obtain a code stream, where the code stream includes first information, and the first information is used to indicate the characteristics of at least one channel in the N channels Data is dequantified.

In a second aspect, an embodiment of the present application provides an image decoding method, including:

Decoding the code stream to obtain the feature data of the current image, where the feature data of the current image includes the feature data of N channels, and the N is a positive integer;

Decoding the code stream to obtain first information, where the first information is used to instruct to perform inverse quantization on the feature data of at least one channel in the N channels;

According to the first information, inverse quantization is performed on the feature data of the at least one channel.

In a third aspect, the present application provides a video encoder for performing the method in the first aspect or each of its implementations. Specifically, the encoder includes a functional unit for executing the method in the above-mentioned first aspect or each of its implementations.

In a fourth aspect, the present application provides a video decoder for executing the method in the second aspect or each of its implementations. Specifically, the decoder includes functional units for performing the methods in the second aspect or the respective implementations thereof.

In a fifth aspect, a video encoder is provided, including a processor and a memory. The memory is used for storing a computer program, and the processor is used for calling and running the computer program stored in the memory, so as to execute the method in the above-mentioned first aspect or each implementation manner thereof.

In a sixth aspect, a video decoder is provided, including a processor and a memory. The memory is used for storing a computer program, and the processor is used for calling and running the computer program stored in the memory, so as to execute the method in the above-mentioned second aspect or each implementation manner thereof.

In a seventh aspect, a video encoding and decoding system is provided, including a video encoder and a video decoder. A video encoder is used to perform the method in the first aspect or each of its implementations, and a video decoder is used to perform the method in the above-mentioned second aspect or its implementations.

In an eighth aspect, a chip is provided for implementing any one of the above-mentioned first aspect to the second aspect or the method in each implementation manner thereof. Specifically, the chip includes: a processor for invoking and running a computer program from a memory, so that a device on which the chip is installed executes any one of the above-mentioned first to second aspects or each of its implementations method.

In a ninth aspect, a computer-readable storage medium is provided for storing a computer program, the computer program causing a computer to execute the method in any one of the above-mentioned first aspect to the second aspect or each of its implementations.

In a tenth aspect, a computer program product is provided, comprising computer program instructions, the computer program instructions causing a computer to perform the method in any one of the above-mentioned first to second aspects or the implementations thereof.

In an eleventh aspect, there is provided a computer program which, when run on a computer, causes the computer to perform the method in any one of the above-mentioned first to second aspects or the respective implementations thereof.

Based on the above technical solutions, by acquiring the current image to be encoded, inputting the current image into a neural network to obtain feature data of the current image, the feature data of the current image includes the feature data of N channels; Perform quantization; encode the quantized feature data of at least one channel to obtain a code stream, where the code stream includes first information, and the first information is used to instruct the decoding point to perform the quantization on the feature data of at least one channel in the N channels Inverse quantization. In this way, the feature data output by the middle layer of the neural network is quantized, so that the technologies in the existing video and image encoding and decoding standards can be reused to encode the feature data, and the encoding efficiency is improved.

Description of drawings

1 is a schematic diagram of an encoding and decoding framework for image pre-analysis and recompression involved in an embodiment of the present application;

Fig. 2 is a schematic diagram of an MPEG-VCM potential encoding process;

FIG. 3 is a schematic flowchart of an image encoding method 300 provided by an embodiment of the present application;

FIG. 4 is a schematic flowchart of an image encoding method 400 provided by an embodiment of the present application;

FIG. 5 is a schematic flowchart of an image encoding method 500 provided by an embodiment of the present application;

FIG. 6 is a schematic flowchart of an image encoding method 600 provided by an embodiment of the present application;

FIG. 7 is a schematic flowchart of an image decoding method 700 provided by an embodiment of the present application;

FIG. 8 is a schematic flowchart of an image decoding method 800 provided by an embodiment of the present application;

FIG. 9 is a schematic flowchart of an image decoding method 900 provided by an embodiment of the present application;

FIG. 10 is a schematic flowchart of an image decoding method 1000 provided by an embodiment of the present application;

FIG. 11 is a schematic block diagram of a video encoder 10 provided by an embodiment of the present application;

12 is a schematic block diagram of a video decoder 20 provided by an embodiment of the present application;

FIG. 13 is a schematic block diagram of an electronic device 30 provided by an embodiment of the present application;

FIG. 14 is a schematic block diagram of a video encoding and decoding system 40 provided by an embodiment of the present application.

Detailed ways

This application can be applied to various video encoding and decoding fields for machine vision and human-machine hybrid vision, combining 5G, AI, deep learning, feature extraction and video analysis technologies with existing video processing and encoding technologies. The 5G era has spawned a large number of machine-oriented applications, such as the Internet of Vehicles, unmanned driving, industrial Internet, smart and safe cities, wearables, video surveillance and other machine vision content. Compared with the increasingly saturated human-oriented video, the application scenarios are more extensive. , video encoding for machine vision will become one of the main sources of incremental traffic in the 5G and post-5G era.

For example, the solution of the present application can be combined with audio video coding standard (audio video coding standard, AVS for short), for example, H.264/audio video coding (audio video coding, AVC for short) standard, H.265/High Efficiency Video Coding ( High efficiency video coding, referred to as HEVC) standard and H.266/versatile video coding (versatile video coding, referred to as VVC) standard. Alternatively, the schemes of the present application may operate in conjunction with other proprietary or industry standards including ITU-TH.261, ISO/IECMPEG-1 Visual, ITU-TH.262 or ISO/IECMPEG-2 Visual, ITU-TH.263 , ISO/IECMPEG-4Visual, ITU-TH.264 (also known as ISO/IECMPEG-4AVC), including Scalable Video Codec (SVC) and Multi-View Video Codec (MVC) extensions. It should be understood that the techniques of this application are not limited to any particular codec standard or technique.

FIG. 1 is a schematic diagram of an encoding/decoding framework for image pre-analysis and recompression according to an embodiment of the present application.

In intelligent analysis-oriented application scenarios, in addition to high-quality viewing of videos and images, videos and images are also used to analyze and understand the semantic information in them. In view of the more unique analysis requirements of video and image coding in intelligent analysis tasks, many researchers now switch from traditional direct compression coding of images to compression coding of feature data output by the middle layer of the intelligent analysis task network.

As shown in Figure 1, end-side devices such as cameras first use task networks to pre-analyze the original video and image data collected or input, such as input task network A, task network B, and task network B, and extract enough cloud analysis. feature data, and compress, encode and transmit these feature data. After the cloud device receives the corresponding code stream, it reconstructs the corresponding feature data according to the syntax information of the code stream, and inputs it into the specific task network for further analysis. Under the coding and decoding framework shown in Figure 1, there is a large amount of feature data transmission between the terminal device and the cloud device. The purpose of feature data compression is to compress and encode the feature data extracted from the existing task network in a recoverable manner. , for further intelligent analysis and processing in the cloud.

Aiming at the problem of efficient video and image coding for intelligent analysis task scenarios shown in Figure 1, the current international ISO/IEC HTC 1/SC 29 (Audio, Image Coding, Multimedia and Hypermedia Information Subcommittee) under the MPEG ( The Moving Picture Experts Group (formerly WG11) International Standards Organization has established the Video Coding for Machines (VCM) standard working group at its 127th meeting in July 2019 to study technologies in this area, aiming at compressing video or video The feature information extracted from the data defines a code stream, so that it can use the same code stream to perform multiple intelligent analysis tasks without significantly reducing the performance of intelligent task analysis. At the same time, the decompressed information is more friendly to intelligent analysis tasks. The performance loss of intelligent analysis tasks is smaller at the bit rate. At the same time, the standard working meeting of the Multimedia Sub-Committee under the National Information Technology Standardization Technical Committee held the first working group meeting in Hangzhou, Zhejiang Province in January 2020, and correspondingly established the data coding for machine intelligence (Data Compression for Machines, DCM) standard working group to study the technical application of this aspect, aiming to support the involved machine intelligence applications or human-machine hybrid intelligent applications through efficient data representation and compression.

FIG. 2 is a schematic diagram of a potential coding flow of MPEG-VCM. At present, the VCM standard working group has designed a potential coding flow chart as shown in Figure 2, in order to improve the coding efficiency of video and images under intelligent analysis tasks. The video and image can directly pass through the video and image encoder optimized for the task, or use network pre-analysis to extract feature data and encode it, and then input the decoded feature data into the subsequent network for further analysis. If it is necessary to multiplex the existing video and image coding standards to compress the extracted feature data, it is necessary to perform fixed-point processing on the feature data represented by the floating point type.

The image coding method involved in the embodiments of the present application will be described in detail below with reference to specific examples.

First, the encoding process is introduced by taking the encoding end as an example.

FIG. 3 is a schematic flowchart of an image encoding method 300 provided by an embodiment of the present application. The execution body of the embodiment of the present application can be understood as the encoder shown in FIG. 2 , as shown in FIG. 3 , including:

S301. Acquire a current image to be encoded.

S302, input the current image into a neural network to obtain feature data of the current image, where the feature data of the current image includes the feature data of N channels, where N is a positive integer;

S303, quantify the characteristic data of at least one channel in the N channels;

S304. Encode the quantized feature data of the at least one channel to obtain a code stream, where the code stream includes first information, where the first information is used to indicate that the feature data of at least one channel among the N channels is to be encoded. Inverse quantization.

The current image in this application can be understood as a frame of image to be encoded in the video process or a part of the image in the frame; or, the current image can be understood as a single image to be encoded or a part of the image in the image to be encoded.

The neural network in this application is any task network, for example, a classification network, a target detection network, a semantic segmentation gateway, etc. The type of the neural network is not limited in this application.

Input the current image into the neural network to obtain the feature data output by the middle layer of the neural network. In some embodiments, the feature data is a floating-point number type. In order to reuse the existing video coding framework, it is necessary to Feature data is quantified.

In some embodiments, most existing video coding frameworks compress fixed-point data during compression. Therefore, the encoder needs to quantize floating-point feature data into fixed-point feature data. Characteristic data of point type is encoded.

Optionally, the feature data of the fixed-point type includes the feature data of the integer type, that is, the encoder quantizes the feature data of the floating-point number type into the feature data of the integer type.

In some embodiments, the feature data of the floating-point number type of the current image includes the feature data of the floating-point number type of N channels, where N is a positive integer. Types of feature data are quantified.

In some embodiments, the methods for quantizing the floating-point type feature data of at least one of the N channels in S303 include, but are not limited to, the following:

Manner 1: Use the same quantization method to quantize the floating-point feature data of all channels in the N channels; in this case, transmit a set of quantization parameters in all the N channels in the code stream.

The second method is to use a quantization method to quantize the floating point type feature data of each channel in the N channels; at this time, each channel of the N channels in the code stream transmits a set of quantization parameters.

Manner 3: Group the N channels, and use a quantization method to quantize the floating point type feature data of each group of channels. At this time, a set of quantization parameters are transmitted in the same group of channels in the code stream.

In some embodiments, the quantization method for quantizing the floating point type feature data of at least one channel may include a linear uniform quantization method, a nonlinear uniform quantization method, or a look-up table quantization method. Among them, the nonlinear uniform quantization method further includes nonlinear exponential function quantization and nonlinear logarithmic function quantization. It should be noted that the quantization methods in the embodiments of the present application include but are not limited to the above several quantization methods, and other quantization methods may also be used to quantify the characteristic data of the floating point type into the characteristic data of the fixed point type. No restrictions.

In the image encoding method provided by the present application, the current image to be encoded is obtained, and the current image is input into a neural network to obtain floating-point feature data of the current image, wherein the floating-point feature data of the current image includes floating point data of N channels. point type feature data; quantize the floating point type feature data of at least one channel in the N channels; encode the quantized feature data of at least one channel to obtain a code stream. In this way, the progress of the feature data output from the middle layer of the neural network is fixed, so that the existing video and image coding and decoding standards can be reused to encode the feature data, and the feature data of at least one channel of the N channels can be encoded at the same time. Fixed-pointing is performed, so as to improve the encoding efficiency of the fixed-point feature data and realize efficient compression of the feature data. In addition, the present application considers the channel information of the feature data in the quantization process at the encoding end, and can process the feature data between different channels, thereby improving the quantization reliability of the feature data.

The following describes in detail the process of using the same quantization method to quantize the feature data of the floating point type of all channels in the N channels with reference to FIG. 4 .

FIG. 4 is a schematic flowchart of an image encoding method 400 provided by an embodiment of the present application, as shown in FIG. 4 , including:

S401. Acquire a current image to be encoded.

S402, input the current image into the neural network, and obtain the feature data of the floating point type of N channels of the current image;

S403, use the same quantization method to quantize the floating-point type feature data of all channels in the N channels;

S404: Encode the feature data of the fixed-point type of the current image to obtain a code stream.

Optionally, the quantization method includes linear uniform quantization, nonlinear function quantization, and look-up table quantization.

The above code stream includes fixed-point feature data under all channels.

In some embodiments, if the present embodiment adopts the linear uniform quantization method to quantize the floating point type feature data of all channels in the N channels, the above S403 includes the following S403-A1 and S403-A2:

S403-A1, obtain the preset first quantization bit width, and the first eigenvalue and the second eigenvalue in the feature data of the floating point type of all channels in the N channels;

S403-A2. According to the first eigenvalue, the second eigenvalue, and the first quantization bit width, use a linear uniform quantization method to quantize the floating point type feature data of each channel in the N channels.

In the embodiment of the present application, the feature data of the floating point type of all channels in the N channels are taken as a whole, and the first feature value and the second feature value are obtained from the feature data of the floating point number type of all channels in the N channels.

Optionally, the above-mentioned preset first quantization bit width may be preset and set in the configuration file of the encoder.

Optionally, the first eigenvalue is the smallest eigenvalue in the feature data of the floating point type of all channels in the N channels of the current image, and the second eigenvalue is the floating point number of all channels in the N channels of the current image The largest eigenvalue in the eigendata of type.

In some embodiments, the encoder quantizes the floating-point type feature data of all channels in the N channels according to the following formula (1):

Among them, x _cij is the eigenvalue of the floating-point type of the i-th row and the j-th column of the c-th channel; x _cmax1 and x _cmin1 are the second eigenvalues and The first eigenvalue; bitdepth1 is the first quantization bit width, int[ ] represents the integerization function; y _cij is the eigenvalue of the fixed-point type of the i-th row and the j-th column of the c-th channel after quantization; Δ is a polar It is a small value, and can be set to 0, which is used to map the floating-point feature data into a value range of left closed and right open.

It should be noted that the above formula (1) is just an example, and the linear uniform quantization method of the present application also includes transforming the above formula (1), for example, transforming it into formula (2):

Alternatively, add, multiply or divide one or more coefficients in the above formula (1).

According to the above formula (1), after quantizing the floating-point feature data of all channels of the current image into fixed-point feature data, encode the fixed-point feature data to form a code stream.

According to different nonlinear functions, nonlinear uniform quantization methods include nonlinear logarithmic uniform quantization methods and nonlinear exponential uniform quantization methods.

In some embodiments, if the non-linear logarithmic uniform quantization method is used in this embodiment to quantize the floating point type feature data of all channels in the N channels, the above S403 includes the following S403-B1 and S403-B2:

S403-B1, obtain the preset second quantization bit width and the first base of the logarithmic function, and the first eigenvalue and the second eigenvalue in the characteristic data of the floating point type of all channels in the N channels;

S403-B2. According to the first eigenvalue and the second eigenvalue, as well as the second quantization bit width and the first base of the logarithmic function, use a non-linear logarithmic uniform quantization method to quantify the floating point number of each channel in the N channels Types of feature data are quantified.

Optionally, the preset second quantization bit width and the first base of the logarithmic function may be preset by the user and set in the configuration file of the encoder. The second quantization bit width may be determined according to the size of the first characteristic value, and the first base of the logarithmic function is determined according to the characteristic of the characteristic data.

In some embodiments, the encoder quantizes the floating point type feature data of each of the N channels according to the following formula (3):

Wherein, bitdepth2 is the second quantization bit width, and log_base1 is the first base of the logarithmic function used in logarithmic quantization.

It should be noted that the above formula (3) is just an example, and the nonlinear logarithmic uniform quantization method of the present application also includes transforming the above formula (3), for example, transforming it into formula (4):

Or add, multiply or divide one or more coefficients in the above formula (3).

Optionally, the second quantization bit width is equal to the first quantization bit width.

In some embodiments, if the non-linear exponential uniform quantization method is used to quantize the characteristic data of the floating point type of all channels in the N channels, the above S403 includes the following S403-C1 and S403-C2:

S403-C1, obtain the preset third quantization bit width and the first base of the exponential function, and the first eigenvalue and the second eigenvalue in the characteristic data of the floating point type of all channels in the N channels;

S403-C2. According to the first eigenvalue and the second eigenvalue, as well as the third quantization bit width and the first base of the exponential function, use a non-linear exponential uniform quantization method to quantify the floating point number type of each channel in the N channels. Feature data is quantified.

Optionally, the above-mentioned preset third quantization bit width and the first base of the exponential function may be preset by the user and set in the configuration file of the encoder. The third quantization bit width may be determined according to the size of the first characteristic value, and the first base of the exponential function is determined according to the characteristic of the characteristic data.

In some embodiments, the encoder quantizes the floating point type feature data of each of the N channels according to the following formula (5):

Wherein, bitdepth3 is the third quantization bit width, and e_base is the first base of the exponential function used in exponential quantization.

It should be noted that the above formula (5) is only an example, and the nonlinear logarithmic uniform quantization method of the present application also includes transforming the above formula (5), for example, transforming it into formula (6):

Or add, multiply or divide one or more coefficients in the above formula (5).

Optionally, the above-mentioned third quantization bit width is equal to the above-mentioned first quantization bit width.

In some embodiments, if the present embodiment adopts the table lookup quantization method to quantize the floating point type feature data of all channels in the N channels, the above S403 includes the following S403-D1 to S403-D3:

S403-D1. Sort the floating-point feature data of all channels in the N channels according to the value size, and obtain the sorted first feature data;

S403-D2, dividing the sorted first characteristic data into a plurality of first quantization intervals, wherein each first quantization interval includes the same amount of characteristic data;

S403-D3. For each first quantization interval, quantize the value of the feature data in the first quantization interval into an index value of the first quantization interval.

In this embodiment, the feature data of the floating-point number type of all channels in the N channels are taken as a whole, and each feature value in the feature data of the floating-point number type of all channels in the N channels is sorted in descending order according to the value size. Or sort from small to large to obtain the feature data of the floating point type of all channels after sorting. For the convenience of description, the sorted feature data is called the sorted first feature data. The sorted first feature data is divided into a plurality of first quantization intervals, and each first quantization interval includes the same quantity of feature data. Each first quantization interval is represented by an index that can be represented by a corresponding quantization bit width, so that each first quantization interval has an index. In this way, during quantization, the value of the feature data in each first quantization interval can be quantized into an index value of each first quantization interval.

For the table lookup quantification method, since the 0 value in the feature data accounts for a large proportion, the feature data other than the 0 value after sorting can be divided into intervals containing the same amount of feature data, that is, all the 0 values of the sorted feature data are recorded as Index 0, the corresponding reconstruction value is also set to 0 value.

The following describes the process of using a quantization method to quantize the feature data of the floating point type of each channel of the N channels of the current image in detail with reference to FIG. 5 .

FIG. 5 is a schematic flowchart of an image encoding method 500 provided by an embodiment of the present application, as shown in FIG. 5 , including:

S501. Acquire a current image to be encoded.

S502: Input the current image into a neural network to obtain feature data of the floating point type of N channels of the current image.

S503. Quantize the feature data of the floating point type of each channel in the N channels by using a quantization method respectively.

S505: Encode the feature data of the fixed-point type of the current image to obtain a code stream.

In some embodiments, if the present embodiment adopts a linear uniform quantization method to quantize the floating point type feature data of each of the N channels, the above S503 includes the following S503-A1 and S503-A2:

S503-A1, for each channel in the N channels, obtain a preset fourth quantization bit width, and the third eigenvalue and the fourth eigenvalue in the feature data of the floating point type of the channel;

S503-A2. According to the third eigenvalue, the fourth eigenvalue, and the fourth quantization bit width, use a linear uniform quantization method to quantize the floating point type feature data of the channel.

In the embodiment of the present application, the feature data of the floating point number type of each channel in the N channels is taken as a whole, and a quantization method is used to quantize the feature data of the floating point number type of each channel.

It should be noted that the quantization process for the characteristic data of each channel in the N channels is the same, and for convenience of description, one channel in the N channels is taken as an example. Obtain the largest eigenvalue and the smallest eigenvalue from the feature data of the floating point type of the channel, record the largest eigenvalue as the third eigenvalue, and record the smallest eigenvalue as the fourth eigenvalue.

Optionally, the above-mentioned preset fourth quantization bit width may be preset by the user and set in the configuration file of the encoder. The fourth quantization bit width may be determined according to the size of the third eigenvalue.

Optionally, the third eigenvalue is the largest eigenvalue in the feature data of the floating point type of the channel, and the fourth eigenvalue is the smallest eigenvalue in the feature data of the floating point type of the channel.

In some embodiments, the encoder quantizes the floating point type feature data of the channel according to the following formula (7):

Among them, the current channel is the c-th channel, x _cij is the eigenvalue of the floating-point type of the i-th row and the j-th column of the channel, and x _cmax2 and x _cmin2 are the second largest among the floating-point type characteristic data of the channel. value and the second minimum value, bitdepth4 is the fourth quantization bit width, int[ ] represents the integerization function, y _cij is the eigenvalue of the fixed-point type of the i-th row and the j-th column of the channel after quantization, and Δ is a polar Small value, which can be set to 0, is used to map the floating-point feature data into a value range of left closed and right open.

It should be noted that the above formula (7) is only an example, and the linear uniform quantization method of the present application also includes transforming the above formula (7), for example, the following formula (8):

Or add, multiply or divide one or more coefficients in the above formula (7).

According to the above formula (7), after the characteristic data of the floating point type of the channel is quantized into the characteristic data of the fixed point type, the characteristic data of the fixed point type is encoded to form a code stream.

In the present application, according to the difference of nonlinear functions, nonlinear uniform quantization methods include nonlinear logarithmic uniform quantization methods and nonlinear exponential uniform quantization methods.

In some embodiments, if the non-linear logarithmic uniform quantization method is used to quantize the characteristic data of the floating point type of the channel, the above S503 includes the following S503-B1 and S503-B2:

S503-B1, for each channel of the N channels, obtain a preset fifth quantization bit width and the second base of the logarithmic function, and the third eigenvalue and the third eigenvalue in the feature data of the floating point type of the channel Four eigenvalues;

S503-B2. According to the third eigenvalue and the fourth eigenvalue, and the fifth quantization bit width and the second base of the logarithmic function, use the nonlinear logarithmic uniform quantization method to perform the floating point number type feature data of the channel. quantify.

Optionally, the preset fifth quantization bit width and the second base of the logarithmic function may be preset by the user and set in the configuration file of the encoder. Wherein, the fifth quantization bit width can be determined according to the size of the third characteristic value, and the second base of the logarithmic function is determined according to the characteristics of the characteristic data in the channel.

In some embodiments, the encoder quantizes the floating point type feature data of the channel according to the following formula (9):

Wherein, bitdepth5 is the fifth quantization bit width, and log_base2 is the second base of the logarithmic function used in logarithmic quantization, for example, 10.

It should be noted that the above formula (9) is only an example, and the nonlinear logarithmic uniform quantization method of the present application also includes transforming the above formula (9), for example, the following formula (10):

Or add, multiply or divide one or more coefficients in the above formula (9).

Optionally, the fifth quantization bit width is equal to the fourth quantization bit width.

In some embodiments, if the non-linear exponential uniform quantization method is used to quantize the characteristic data of the floating point type of the channel, the above S503 includes the following S503-C1 and S503-C2:

S503-C1, for each channel of the N channels, obtain the preset sixth quantization bit width and the second base of the exponential function, as well as the third eigenvalue and the fourth feature in the feature data of the channel's floating point type value;

S503-C2. According to the third eigenvalue and the fourth eigenvalue, the sixth quantization bit width and the second base of the exponential function, use a nonlinear exponential uniform quantization method to quantize the floating point type feature data of the channel.

Optionally, the above-mentioned preset sixth quantization bit width and the second base of the exponential function may be preset by the user and set in the configuration file of the encoder. Wherein, the sixth quantization bit width can be determined according to the size of the third characteristic value, and the second base of the exponential function is determined according to the characteristics of the characteristic data under the channel.

In some embodiments, the encoder quantizes the floating point type feature data of the channel according to the following formula (11):

Wherein, bitdepth6 is the sixth quantization bit width, and e_base2 is the second base of the exponential function used in exponential quantization.

It should be noted that the above formula (11) is only an example, and the nonlinear logarithmic uniform quantization method of the present application also includes transforming the above formula (11), for example, the following formula (12):

Or add, multiply or divide one or more coefficients in the above formula (11).

Optionally, the above-mentioned sixth quantization bit width is equal to the above-mentioned fourth quantization bit width.

In some embodiments, if this embodiment adopts the table look-up quantization method to quantize the characteristic data of the floating point type of the channel, the above S503 includes the following S503-D1 to S503-D3:

S503-D1, for each channel in the N channels, sort the feature data of the floating point type of the channel according to the value size, and obtain the sorted second feature data under the channel;

S503-D2, the second characteristic data sorted under the channel is divided into a plurality of second quantization intervals, wherein each second quantization interval includes the same amount of characteristic data;

S503-D3: For each second quantization interval, quantize the value of the feature data in the second quantization interval into an index value of the second quantization interval.

In this embodiment, the feature data of the floating point type of the channel is sorted from large to small or from small to large according to the value size. For the convenience of description, the sorted feature data of the channel is called sorted Second characteristic data. The second feature data sorted in the channel is divided into a plurality of second quantization intervals, and each second quantization interval includes the same quantity of feature data. Each second quantization interval is represented by an index that can be represented by a corresponding quantization bit width, so that each second quantization interval has an index. In this way, during quantization, the value of the feature data in each second quantization interval can be quantized into an index value of each second quantization interval.

The following describes in detail the process of using a quantization method to quantize the characteristic data of the floating point type of each channel of the M groups of channels with reference to FIG. 6 .

FIG. 6 is a schematic flowchart of an image encoding method 600 provided by an embodiment of the present application, as shown in FIG. 6 , including:

S601. Acquire a current image to be encoded.

S602: Input the current image into a neural network to obtain floating point type feature data of N channels of the current image.

S603: Quantize the feature data of the floating point type of each group of channels using a quantization method respectively.

S604: Encode the feature data of the fixed-point type of the current image to obtain a code stream.

In some embodiments, if the present embodiment adopts the linear uniform quantization method to quantize the floating point type characteristic data of the group of channels, the above S603 includes the following S603-A1 and S603-A2:

S603-A1, for each group of channels, obtain the preset seventh quantization bit width, and the fifth eigenvalue and the sixth eigenvalue in the feature data of the floating point type of the group of channels;

S603-A2. According to the fifth eigenvalue, the sixth eigenvalue, and the seventh quantization bit width, use a linear uniform quantization method to quantize the floating point type feature data of each channel in the group of channels.

In the embodiment of the present application, the N channels are divided into multiple groups of channels, the feature data of the floating-point number type of each group of channels is taken as a whole, and a quantization method is used for the feature data of the floating-point number type of each group of channels. quantify.

It should be noted that the quantization process for the characteristic data of each group of channels is the same. For the convenience of description, a group of channels is taken as an example. Obtain the largest eigenvalue and the smallest eigenvalue from the feature data of the floating point type of the set of channels, record the largest eigenvalue as the fifth eigenvalue, and record the smallest eigenvalue as the sixth eigenvalue.

Optionally, the above-mentioned preset seventh quantization bit width may be preset by the user and set in the configuration file of the encoder. The seventh quantization bit width may be determined according to the size of the fifth eigenvalue.

Optionally, the fifth characteristic value is the largest characteristic value in the floating point type characteristic data of the group of channels, and the sixth characteristic value is the smallest characteristic value in the floating point type characteristic data of the group of channels.

In some embodiments, the encoder quantizes the floating point type feature data of the set of channels according to the following formula (13):

Among them, the c-th channel is a channel in the group of channels, x _cij is the eigenvalue of the floating-point number type in the i-th row and the j-th column of the c-th channel, and x _cmax3 and x _cmin3 are the floating-point number type of the group of channels respectively. The third maximum value and the third minimum value in the feature data, bitdepth7 is the seventh quantization bit width, int[ ] represents the integerization function, y _cij is the fixed-point number of the i-th row and the j-th column of the c-th channel after quantization The eigenvalue of the type, Δ is a minimum value, which is used to map the eigendata of the floating-point type into the value range of left closed and right open.

It should be noted that the above formula (12) is only an example, and the linear uniform quantization method of the present application also includes transforming the above formula (13), for example, the following formula (14):

Or add, multiply or divide one or more coefficients in the above formula (14).

According to the above formula (14), after the feature data of the floating point type of the group of channels is quantized into the feature data of the fixed point type, the feature data of the fixed point type is encoded to form a code stream.

In some embodiments, if the non-linear logarithmic uniform quantization method is used to quantize the characteristic data of the floating point type of the group of channels, the above S603 includes the following S603-B1 and S603-B2:

S603-B1, for each group of channels, obtain the preset eighth quantization bit width and the third base of the logarithmic function, and the fifth eigenvalue and the sixth eigenvalue in the feature data of the floating point type of the group channel;

S603-B2. According to the fifth eigenvalue and the sixth eigenvalue, as well as the eighth quantization bit width and the third base of the logarithmic function, use a nonlinear logarithmic uniform quantization method to quantify the floating point number of each channel in the group of channels Types of feature data are quantified.

Optionally, the above-mentioned preset eighth quantization bit width and the third base of the logarithmic function may be preset by the user and set in the configuration file of the encoder. The eighth quantization bit width may be determined according to the size of the fifth characteristic value, and the third base of the logarithmic function is determined according to the characteristics of the characteristic data in the group of channels.

In some embodiments, the encoder quantizes the floating point type feature data of the set of channels according to the following formula (15):

Wherein, bitdepth8 is the eighth quantization bit width, and log_base3 is the third base of the logarithmic function used in logarithmic quantization, for example, 10.

It should be noted that the above formula (15) is just an example, and the nonlinear logarithmic uniform quantization method of the present application also includes transforming the above formula (15), for example, the following formula (16):

Or add, multiply or divide one or more coefficients in the above formula (15).

Optionally, the above-mentioned eighth quantization bit width is equal to the above-mentioned eighth quantization bit width.

In some embodiments, if the non-linear exponential uniform quantization method is used to quantize the characteristic data of the floating point type of the channel, the above S603 includes the following S603-C1 and S603-C2:

S603-C1, for each group of channels, obtain the preset ninth quantization bit width and the third base of the exponential function, and the fifth eigenvalue and the sixth eigenvalue in the feature data of the floating point type of the group of channels;

S603-C2. According to the fifth eigenvalue and the sixth eigenvalue, as well as the ninth quantization bit width and the third base of the exponential function, use a non-linear logarithmic uniform quantization method to obtain the floating point number of each channel in the group of channels Types of feature data are quantified.

Optionally, the above-mentioned preset ninth quantization bit width and the third base of the exponential function may be preset by the user and set in the configuration file of the encoder. Wherein, the ninth quantization bit width may be determined according to the size of the fifth characteristic value, and the third base of the exponential function is determined according to the characteristics of the characteristic data under the group of channels.

In some embodiments, the encoder quantizes the floating point type feature data of the set of channels according to the following formula (17):

Wherein, bitdepth9 is the ninth quantization bit width, and e_base3 is the third base of the exponential function used in exponential quantization.

It should be noted that the above formula (17) is just an example, and the nonlinear logarithmic uniform quantization method of the present application also includes transforming the above formula (17), for example, the following formula (18):

Or add, multiply or divide one or more coefficients in the above formula (18).

Optionally, the above-mentioned ninth quantization bit width is equal to the above-mentioned ninth quantization bit width.

In some embodiments, if this embodiment adopts the table look-up quantization method to quantize the floating point type characteristic data of the group of channels, the above S603 includes the following S603-D1 to S603-D3:

S603-D1. For each group of channels, sort the floating-point type feature data of the group channel according to the value size, and obtain the sorted third feature data under the group channel;

S603-D2, dividing the sorted third characteristic data under the group channel into a plurality of third quantization intervals, wherein each third quantization interval includes the same amount of characteristic data;

S603-D3: For each third quantization interval, quantize the value of the feature data in the third quantization interval into an index value of the third quantization interval.

In this embodiment, the feature data of the floating point type of the group of channels is sorted according to the value size from large to small or from small to large. For the convenience of description, the sorted feature data under the group of channels is called sorting After the third characteristic data. The sorted third characteristic data in the group of channels is divided into a plurality of third quantization intervals, and each third quantization interval includes the same quantity of characteristic data. Each third quantization interval is represented by an index that can be represented by a corresponding quantization bit width, so that each third quantization interval has an index. In this way, during quantization, the value of the feature data in each third quantization interval can be quantized into an index value of each third quantization interval.

The quantization process at the encoding end is described above, and the content indicated by the first information is described below.

In the present application, after the encoding end quantizes the floating point type feature data of at least one channel into a fixed point number type according to the above steps, the encoding end encodes the fixed point number type characteristic data in a code stream and sends it to the decoding end. At the same time, the encoding end carries first information in the code stream, where the first information indicates to perform inverse quantization on the feature data of the fixed-point type of at least one channel.

In some embodiments, the code stream further includes second information, where the second information is used to indicate an inverse quantization method used when performing inverse quantization on the characteristic data of the fixed-point type of at least one channel.

The inverse quantization method used when performing inverse quantization on the fixed-point type characteristic data of at least one channel includes any one of the following: linear uniform inverse quantization method, nonlinear exponential uniform inverse quantization method, nonlinear logarithmic uniform inverse quantization method, Look-up table inverse quantification method. It should be noted that the inverse quantization methods in the embodiments of the present application include but are not limited to the above several quantization methods, and other inverse quantization methods can also be used to inverse quantize the characteristic data of the fixed-point type into the characteristic data of the floating-point type. There is no restriction on the inverse quantization method.

In some embodiments, the first information includes at least one parameter required for inverse quantization of fixed-point type feature data of at least one channel.

At least one parameter included in the first information in this application includes the following situations:

In case 1, the first information indicates that inverse quantization is performed on the characteristic data of the fixed-point type of all channels in the N channels. At this time, according to the different inverse quantization methods, the first information includes the following example 1, example 2, example 3 or example Any of the four:

Example 1, if the inverse quantization method for performing inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels is a linear uniform inverse quantization method, the first information includes the first target feature value, the first target scaling value and the The first target quantization bit width.

Wherein, the first target feature value is one feature value in the feature data of all the channels in the N channels, for example, the first target feature value is the minimum value of the feature data of all the channels in the N channels.

The first target scaling value is the scaling value corresponding to the feature data of all channels in the N channels during quantization, and the first target quantization bit width is the quantization bit width corresponding to the feature data of all channels in the N channels during quantization.

The following describes the process of determining the first target scaling value in conjunction with the encoding mode of the encoding end.

In an example, if the quantization method of the encoding end for all channels is a linear uniform quantization method, the encoding end may use the first eigenvalue and the second eigenvalue in the characteristic data of all the channels in the N channels, and the A target quantization bit width determines the first target scaling value.

Optionally, the first target scaling value s _c1 may be determined according to the following formula (19):

Wherein, x _cmin1 and x _cmax1 are the first eigenvalue and the second eigenvalue in the feature data of all channels in the N channels, respectively. The first target quantization bit width 1bitdepth may be the first quantization bit width bitdepth1 in the above formula (1).

It should be noted that the above formula (19) is only an example, and the formula for determining the first target scaling value s _c1 in the present application also includes the modification of the above formula (19), or the addition and addition of the above formula (19). Multiply or divide one or more coefficients, etc.

In another example, if the encoding end quantizes all channels in a non-linear logarithmic uniform quantization mode, the encoding end may use the first eigenvalue and the second feature in the feature data of all channels in the N channels. value, together with the first target quantization bit width and the first base of the logarithmic function to determine the first target scaling value.

Optionally, the first target scaling value s _c1 may be determined according to the following formula (20):

Wherein, log _{log_base1} is the first base of the logarithmic function, and the first target quantization bit width may be the second quantization bit width in the above formula (3).

It should be noted that the above formula (20) is just an example, and the formula for determining the first target scaling value s _c1 in the present application also includes the modification of the above formula (20), or the addition and addition of the above formula (20). Multiply or divide one or more coefficients, etc.

In another example, if the encoding end quantizes all channels in a non-linear exponential uniform quantization mode, the encoding end may use the first eigenvalue and the second eigenvalue in the feature data of all channels in the N channels. , and the first target quantization bit width and the first base of the exponential function determine the first target scaling value.

Optionally, the first target scaling value s _c1 may be determined according to the following formula (21):

Wherein, e_base1 is the first base of the exponential function, and the first target quantization bit width may be the third quantization bit width bitdepth3 in the above formula (5).

It should be noted that the above formula (21) is only an example, and the formula for determining the first target scaling value s _c1 in the present application also includes the modification of the above formula (21), or the addition and addition of the above formula (21). Multiply or divide one or more coefficients, etc.

In this way, the decoding end can parse the first information from the code stream, and according to the first target feature value, the first target scaling value and the first target quantization bit width included in the first information, use the linear uniform inverse quantization method to Inverse quantization is performed on the fixed-point type feature data of all channels in the channel.

Example 2, if the inverse quantization method for performing inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels is a nonlinear logarithmic uniform inverse quantization method, at this time, the first information includes the first target eigenvalue, the first The target scaling value and the first target quantization bit width, or the first information includes the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base, or the first information includes the first target feature value, the first target scaling value, and the indication information of the first target quantization bit width and the first logarithmic base.

Specifically, if the first information includes the first target feature value, the first target scaling value, and the first target quantization bit width, the decoding end uses the first target feature value, the first target scaling value, and the first target quantization bit width and The default logarithmic base, which uses the nonlinear logarithmic uniform inverse quantization method to dequantize the fixed-point type feature data of all the N channels.

If the first information includes the first target eigenvalue, the first target scaling value, the first target quantization bit width and the first logarithmic base, the decoding end directly uses the first target eigenvalue, the first target scaling carried by the first information value, the first target quantization bit width and the first logarithmic base, and use the nonlinear logarithmic uniform inverse quantization method to perform inverse quantization on the fixed-point type feature data of all the N channels.

If the first information includes the first target eigenvalue, the first target scaling value, the first target quantization bit width and the first logarithmic base indication information, the first logarithmic base indication information is used to indicate the multiple The first logarithmic base is determined from the logarithmic bases. In this way, the decoding end parses the first information from the code stream, determines the first logarithmic base from the preset multiple logarithmic bases according to the indication information of the first logarithmic base, and then determines the first logarithmic base according to the first target characteristic value, The first target scaling value, the first target quantization bit width, and the first logarithmic base are used to perform inverse quantization on the fixed-point type feature data of all channels in the N channels by using a non-linear logarithmic uniform inverse quantization method.

Example 3, if the inverse quantization method for performing inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels is a nonlinear exponential uniform inverse quantization method, the first information includes the first target eigenvalue, the first target scaling value and the first target quantization bit width, or the first information includes the first target eigenvalue, the first target scaling value, the first target quantization bit width and the first exponent base, or the first information includes the first target eigenvalue, the first A target scaling value and indication information of a first target quantization bit width and a first exponent base.

Specifically, if the first information includes the first target feature value, the first target scaling value, and the first target quantization bit width, the decoding end uses the first target feature value, the first target scaling value, and the first target quantization bit width and The default exponential base, which uses the nonlinear exponential uniform inverse quantization method to inverse quantize the fixed-point feature data of all channels in the N channels.

If the first information includes the first target eigenvalue, the first target scaling value, the first target quantization bit width and the first exponent base, the decoding end directly uses the first target eigenvalue and the first target scaling value carried in the first information With the first target quantization bit width and the first exponent base, the non-linear exponent uniform inverse quantization method is used to inverse quantize the fixed-point type feature data of all the N channels.

If the first information includes the first target eigenvalue, the first target scaling value, the first target quantization bit width and the indication information of the first exponent base, the indication information of the first exponent base is used to indicate multiple exponents from preset In the base number, the base of the first exponent is determined. In this way, the decoding end parses the first information from the code stream, determines the first exponent base from the preset multiple exponent bases according to the indication information of the first exponent base, and then determines the first exponent base according to the first target characteristic value, the first target The scaling value, the first target quantization bit width and the first exponent base are used to inversely quantize the fixed-point feature data of all channels in the N channels by using a non-linear exponential uniform inverse quantization method.

Example 4, if the inverse quantization method for performing inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels is a table look-up inverse quantization method, then the first information includes the index value of the quantization interval and the inverse quantization value of the quantization interval. The first correspondence between the N channels is determined based on the pre-quantization value and the post-quantization value of the feature data of all channels in the N channels. The index of the quantization interval can be understood as a fixed-point eigenvalue, and the inverse quantization value of the quantization interval can be understood as the weighted average value of each eigenvalue in the quantization interval, or the eigenvalue corresponding to the center position of the quantization interval. The weighted average value of each eigenvalue in the quantization interval can also be called the eigenvalue corresponding to the probability distribution center of the quantization interval. The inverse quantization value may also be called a reconstruction value.

For the table lookup inverse quantization method, since the 0 value in the feature data accounts for a large proportion, the feature data other than the 0 value after sorting can be divided into intervals containing the same amount of feature data, that is, all 0 values of the sorted feature data are recorded. is index 0, and the corresponding reconstruction value is also set to 0 value.

In a specific embodiment, the quantization method of the encoding end corresponds to the inverse quantization method of the decoding end one-to-one. The decoding end uses linear inverse quantization to perform inverse quantization on the fixed-point feature data of all channels of the N channels. If the decoding end uses the nonlinear logarithmic uniform quantization method to quantize the floating-point type feature data of all channels in the N channels, the decoding end uses the nonlinear logarithmic uniform inverse quantization method to quantize all the N channels. The feature data of fixed-point type is inversely quantized. If the decoding end uses the nonlinear exponential uniform quantization method to quantize the floating-point type feature data of all channels in the N channels, the decoding end uses the nonlinear exponential uniform inverse quantization method to quantize the fixed-point data of all channels of the N channels. Type of feature data for inverse quantization. If the decoding end uses the table lookup quantization method to quantize the floating point type feature data of all channels in the N channels, the decoding end uses the table lookup inverse quantization method to quantize the fixed point type characteristic data of all the N channels. Do inverse quantization.

The embodiments of the present application may adopt the linear uniform inverse quantization method, the nonlinear logarithmic function inverse quantization method, the nonlinear exponential function inverse quantization method, and the look-up table inverse quantization method.

In some embodiments, the inverse quantization information related to the inverse quantization feature data of the present application may be recorded in the supplementary enhancement information, for example, recorded in the Supplemental Enhancement Information (Supplemental Enhancement Information) of the existing video coding standards H.265/HEVC and H.266/VVC SEI) or AVS standard extension data (Extension Data).

In an example, a new SEI category is added to sei_paylod() of sei_message() in sei_rbsp() of existing video coding standards AVC/HEVC/VVC/EVC, namely Feature data quantization SEI message, payloadType can be defined It is any number that has not been used by other SEI, such as 183. At this time, the syntax structure of sei_payload() is shown in Table 1.

Table 1

Among them, feature_data_quantization represents the inverse quantization of feature data.

Perform inverse quantization on the fixed-point feature data of all the N channels of the current image. When the inverse quantization methods are different, the syntax structures thereof are also different, and the syntax structures corresponding to the different inverse quantization methods are described below.

In some embodiments, if the method of performing inverse quantization on all channels is linear uniform inverse quantization, its syntax structure is shown in Table 2:

Table 2

Syntax elements can be encoded in different efficient entropy coding methods, where the syntax elements are:

flag_channel: used to describe the symbol bit indicating the processing object of the decoding end. When it is 0, it means that all channels are uniformly inverse quantized. When it is 1, it means that each channel is inversely quantized. When it is 2, it means that each group of channels is quantized separately; here the value of flag_channel is 0;

flag_iquantization: used to describe the sign bit indicating the inverse quantization method at the decoding end. When it is 0, it means linear inverse quantization, when it is 1, it means nonlinear logarithmic inverse quantization, when it is 2, it means nonlinear exponential inverse quantization, and when it is 3, it means lookup table Inverse quantization; here the value of flag_iquantization is 0;

channel_num: The number of channels used to describe the feature data is channel_num;

scale_num: The scaling value used to describe the feature data in all channels is scale_num, which can be understood as the above-mentioned first target scaling value;

min_num: The minimum value used to describe the feature data under all channels is min_num, which can be understood as the above-mentioned first target feature value.

In some embodiments, if the method of performing inverse quantization on all channels is nonlinear logarithmic function inverse quantization, its syntax structure is shown in Table 3:

table 3

flag_iquantization: used to describe the sign bit indicating the inverse quantization method at the decoding end. When it is 0, it means linear inverse quantization, when it is 1, it means nonlinear logarithmic inverse quantization, when it is 2, it means nonlinear exponential inverse quantization, and when it is 3, it means lookup table Inverse quantization; the value of flag_iquantization here is 1;

min_num: the minimum value used to describe the feature data under all channels is min_num, which can be understood as the above-mentioned first target feature value;

log_base: The base used to describe the logarithmic inverse quantization is log_base, which can be understood as the first logarithmic base above.

In some embodiments, if the method of performing inverse quantization on all channels is nonlinear logarithmic function inverse quantization, its syntax structure is shown in Table 4:

Table 4

flag_iquantization: used to describe the sign bit indicating the inverse quantization method at the decoding end. When it is 0, it means linear inverse quantization, when it is 1, it means nonlinear logarithmic inverse quantization, when it is 2, it means nonlinear exponential inverse quantization, and when it is 3, it means lookup table Inverse quantization; the value of flag_iquantization here is 2.

e_base: The base of the exponential function used to describe the exponential inverse quantization is e_base, which can be understood as the base of the first exponent.

In some embodiments, if the way of performing inverse quantization on all channels is table lookup inverse quantization, wherein table lookup inverse quantization includes histogram equalization inverse quantization. The grammatical structure of look-up table inverse quantization is shown in Table 5:

table 5

flag_iquantization: used to describe the sign bit indicating the inverse quantization method at the decoding end. When it is 0, it means linear inverse quantization, when it is 1, it means nonlinear logarithmic inverse quantization, when it is 2, it means nonlinear exponential inverse quantization, and when it is 3, it means lookup table Inverse quantization; here flag_iquantization is 3;

hist_codebook_num: the number of inverse quantization values hist_codebook_num included in the reconstructed codebook formed by the first correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval;

hist_codebook: used to describe the inverse quantization value corresponding to the i-th quantization interval index in the reconstructed codebook under table lookup inverse quantization.

In case 2, the first information indicates that inverse quantization is performed on the characteristic data of the fixed-point number type of each channel in the N channels. For each channel, according to the different inverse quantization methods, the content included in the first information is as follows: Example 1. Any of the examples shown in Example 2, Example 3 or Example 4:

Example 1, if the inverse quantization method for inverse quantization of the characteristic data of the fixed-point type of the channel is a linear uniform inverse quantization method, then the first information includes the second target eigenvalue, the second target scaling value and the second target quantization bit. width.

Wherein, the second target feature value is a feature value in the feature data of the channel, for example, the second target feature value is the minimum value of the feature data of the channel.

The second target scaling value is the scaling value corresponding to the feature data of the channel during quantization, and the second target quantization bit width is the quantization bit width corresponding to the feature data of the channel during quantization.

The following describes the process of determining the second target scaling value in combination with the encoding mode of the encoding end.

In an example, if the encoding end quantizes the channel in a linear uniform quantization mode, the encoding end can use the third eigenvalue and the fourth eigenvalue in the feature data of the channel and the second target quantization bit according to the The width determines the second target scaling value determines the second target scaling value.

Optionally, the second target scaling value s _c2 may be determined according to the following formula (22):

Wherein, x _cmax2 and x _cmin2 are the third eigenvalue and the second eigenvalue in the feature data of the channel, respectively. The second target quantization bit width 2bitdepth may be the fourth quantization bit width bitdepth4 in the above formula (7).

It should be noted that the above formula (21) is only an example, and the formula for determining the second target scaling value s _c2 in the present application also includes the modification of the above formula (21), or the addition and addition of the above formula (21). Multiply or divide one or more coefficients, etc.

In another example, if the way that the encoding end quantizes the channel is a non-linear logarithmic uniform quantization method, the encoding end can use the second eigenvalue and the second eigenvalue in the feature data of the channel, and the first eigenvalue. The second target scaling value is determined by the two target quantization bit widths and the second base of the logarithmic function.

Optionally, the second target scaling value s _c2 may be determined according to the following formula (23):

Wherein, log _{log_base2} is the second base of the logarithmic function, and the second target quantization bit width may be the fifth quantization bit width in the above formula (9).

It should be noted that the above formula (23) is only an example, and the formula for determining the second target scaling value s ₂ in the present application also includes the modification of the above formula (23), or the addition and addition of the above formula (23). Multiply or divide one or more coefficients, etc.

In another example, if the encoding end quantizes the channel in a non-linear exponential uniform quantization mode, the encoding end can use the third eigenvalue and the fourth eigenvalue in the feature data of the channel, and the second eigenvalue. The target quantization bit width and the second base of the exponential function determine a second target scaling value.

Optionally, the second target scaling value s _c2 may be determined according to the following formula (24):

Wherein, e_base2 is the second base of the exponential function, and the second target quantization bit width may be the sixth quantization bit width bitdepth6 in the above formula (11).

It should be noted that the above formula (24) is only an example, and the formula for determining the second target scaling value s _c2 in the present application also includes the modification of the above formula (24), or the addition and addition of the above formula (24). Multiply or divide one or more coefficients, etc.

In this way, the decoding end can parse the first information from the code stream, and use the linear uniform inverse quantization method for the channel according to the second target eigenvalue, the second target scaling value and the second target quantization bit width included in the first information. The feature data of fixed-point type is inverse quantized.

Example 2, if the inverse quantization method for inverse quantization of the characteristic data of the fixed-point type of the channel is a nonlinear logarithmic uniform inverse quantization method, at this time, the first information includes the second target eigenvalue, the second target scaling value and the first target eigenvalue. Two target quantization bit widths, or the first information includes the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base, or the first information includes the second target feature value, the second target scaling value and indication of the second target quantization bit width and the second logarithmic base.

Specifically, if the first information includes the second target eigenvalue, the second target scaling value, and the second target quantization bit width, the decoding end uses the sum of the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second target quantization bit width. The default logarithmic base, which uses the nonlinear logarithmic uniform inverse quantization method to dequantize the fixed-point type feature data of this channel.

If the first information includes the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second logarithmic base, the decoding end directly uses the second target eigenvalue, the second target scaling carried by the first information value, the second target quantization bit width and the second logarithmic base, and use the non-linear logarithmic uniform inverse quantization method to perform inverse quantization on the fixed-point type feature data of the channel.

If the first information includes the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second logarithmic base indication information, the second logarithmic base indication information is used to indicate the data from the preset multiple The second logarithmic base is determined from the logarithmic bases. In this way, the decoding end parses the first information from the code stream, determines the second logarithmic base from the preset multiple logarithmic bases according to the indication information of the second logarithmic base, and then determines the second logarithmic base according to the second target characteristic value, The second target scaling value, the second target quantization bit width, and the second logarithmic base are used to inversely quantize the fixed-point feature data of the channel by using a non-linear logarithmic uniform inverse quantization method.

Example 3, if the inverse quantization method for inverse quantization of the characteristic data of the fixed-point type of the channel is the nonlinear exponential uniform inverse quantization method, then the first information includes the second target eigenvalue, the second target scaling value and the second target. The quantization bit width, or the first information includes the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second exponent base, or the first information includes the second target eigenvalue, the second target scaling value and The indication information of the second target quantization bit width and the second exponent base.

Specifically, if the first information includes the second target eigenvalue, the second target scaling value, and the second target quantization bit width, the decoding end uses the sum of the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second target quantization bit width. The default exponential base, which uses the non-linear exponential uniform inverse quantization method to inverse quantize the fixed-point feature data of this channel.

If the first information includes the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second exponent base, the decoding end directly uses the second target eigenvalue and the second target scaling value carried by the first information With the second target quantization bit width and the second exponent base, use the non-linear exponential uniform inverse quantization method to inverse quantize the fixed-point type characteristic data of the channel.

If the first information includes the second target eigenvalue, the second target scaling value, the second target quantization bit width and the indication information of the second exponent base, the indication information of the second exponent base is used to indicate multiple exponents from preset In the base, the base of the second exponent is determined. In this way, the decoding end parses the first information from the code stream, determines the second exponent base from the preset multiple exponent bases according to the indication information of the second exponent base, and then determines the second exponent base according to the second target eigenvalue, the second target The scaling value, the second target quantization bit width and the second exponent base are inversely quantized using the non-linear exponential uniform inverse quantization method for the fixed-point type feature data of the channel.

Example 4, if the inverse quantization method for inverse quantization of the characteristic data of the fixed-point type of the channel is the table look-up inverse quantization method, then the first information includes the second index value between the index value of the quantization interval and the inverse quantization value of the quantization interval. A corresponding relationship, the second corresponding relationship is determined based on the value before quantization and the value after quantization of the characteristic data of the channel. The index of the quantization interval can be understood as a fixed-point eigenvalue, and the inverse quantization value of the quantization interval can be understood as the weighted average value of each eigenvalue in the quantization interval, or the eigenvalue corresponding to the center position of the quantization interval. The weighted average value of each eigenvalue in the quantization interval may also be referred to as the eigenvalue corresponding to the center of the probability distribution of the quantization interval.

In a specific embodiment, the quantization method of the encoding end corresponds to the inverse quantization method of the decoding end. For example, when the encoding end uses a linear quantization method to quantize the characteristic data of the floating point type of the channel, the decoding end uses a linear inverse quantization method. Inverse quantization of the fixed-point type feature data of this channel. If the decoding end uses the nonlinear logarithmic uniform quantization method to quantize the feature data of the floating point type of the channel, the decoding end uses the nonlinear logarithmic uniform inverse quantization method to inverse quantize the fixed point type feature data of the channel. . If the decoding end uses the nonlinear exponential uniform quantization method to quantize the characteristic data of the floating point type of the channel, the decoding end uses the nonlinear exponential uniform inverse quantization method to inverse quantize the fixed point type characteristic data of the channel. If the decoding end uses the table lookup quantization method to quantize the floating point type feature data of the channel, the decoding end uses the table lookup inverse quantization method to inverse quantize the fixed point type characteristic data of the channel.

In this embodiment of the present application, the above-mentioned linear uniform inverse quantization method, nonlinear function inverse quantization method, and table look-up inverse quantization method can be used to perform inverse quantization on the floating-point type feature data of each channel of the N channels of the current image. When the inverse quantization methods are different, the syntax structures thereof are also different, and the syntax structures corresponding to the different inverse quantization methods are described below.

In some embodiments, if the inverse quantization method is linear uniform inverse quantization 6, its syntax structure is shown in Table 6:

Table 6

flag_channel: used to describe the sign bit indicating the processing object of the decoding end. When it is 0, it means that all channels are uniformly inverse quantized. When it is 1, it means that each channel is dequantized separately. When it is 2, it means that each group of channels is quantized separately; here flag_channel is 1 ;

flag_iquantization: used to describe the sign bit indicating the inverse quantization method at the decoding end. When it is 0, it means linear inverse quantization, when it is 1, it means nonlinear logarithmic inverse quantization, when it is 2, it means nonlinear exponential inverse quantization, and when it is 3, it means lookup table Inverse quantization; here flag_iquantization is 0;

channel_num: The number of channels used to describe feature data is channel_num;

scale_num[i]: The scaling value used to describe the feature data under the i-th channel is scale_num[i], which can be understood as the above-mentioned second target scaling value;

min_num[i]: The minimum value used to describe the feature data under the i-th channel is min_num[i], which can be understood as the above-mentioned second target feature value.

In some embodiments, if the inverse quantization method is nonlinear logarithmic function inverse quantization, its syntax structure is shown in Table 7:

Table 7

flag_channel: used to describe the symbol bit indicating the processing object of the decoding end. When it is 0, it means that all channels are uniformly inverse quantized. When it is 1, it means that each channel is inversely quantized. When it is 2, it means that each group of channels is quantized separately; here the value of flag_channel is 1;

log_base: used to describe the base log_base of the logarithmic function used in logarithmic inverse quantization, which can be understood as the second logarithmic base above.

In some embodiments, if the inverse quantization method is nonlinear logarithmic function inverse quantization, its syntax structure is shown in Table 8:

Table 8

flag_iquantization: used to describe the sign bit indicating the inverse quantization method at the decoding end. When it is 0, it means linear inverse quantization, when it is 1, it means nonlinear logarithmic inverse quantization, when it is 2, it means nonlinear exponential inverse quantization, and when it is 3, it means lookup table Inverse quantization; the value of flag_iquantization here is 2;

e_base: The base used to describe the exponential function used in logarithmic inverse quantization is e_base, which can be understood as the above-mentioned second exponential base.

In some embodiments, if the inverse quantization method is table lookup inverse quantization, optionally, the table lookup inverse quantization includes histogram equalization inverse quantization. The grammatical structure of look-up table inverse quantization is shown in Table 9:

Table 9

flag_iquantization: used to describe the sign bit indicating the inverse quantization method at the decoding end. When it is 0, it means linear inverse quantization, when it is 1, it means nonlinear logarithmic inverse quantization, when it is 2, it means nonlinear exponential inverse quantization, and when it is 3, it means lookup table Inverse quantization; the value of flag_iquantization here is 3;

hist_codebook_num[i]: the size of the reconstructed codebook formed by the second correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval is hist_codebook_num[i];

hist_codebook[i][j]: The inverse quantization value of the index used to describe the jth quantization interval in the reconstructed codebook corresponding to the ith channel is hist_codebook[i][j].

Case 3, the first information indicates that inverse quantization is performed on the characteristic data of the fixed-point type of each group of channels in the M groups of channels. For each group of channels, according to different inverse quantization methods, the content included in the first information is as follows. 1. Any one of Example 2, Example 3 or Example 4:

Example 1, if the inverse quantization method for inverse quantization of the characteristic data of the fixed-point type of the set of channels is a linear uniform inverse quantization method, then the first information includes the third target feature value, the third target scaling value and the third target quantization. bit width.

Wherein, the third target feature value is a feature value in the feature data of the group of channels, for example, the third target feature value is the minimum value of the feature data of the group of channels.

The third target scaling value is the scaling value corresponding to the feature data of the group of channels during quantization, and the third target quantization bit width is the quantization bit width corresponding to the feature data of the group of channels during quantization.

The following describes the process of determining the third target scaling value in combination with the encoding mode of the encoding end.

In an example, if the encoding end quantizes the group of channels in a linear uniform quantization manner, the encoding end may use the fifth and sixth eigenvalues in the feature data of the group of channels, and the third target The quantization bit width determines the third target scaling value.

Optionally, the third target scaling value s _c3 can be determined according to the following formula (25):

Wherein, x _cmax3 and x _cmin3 are the fifth eigenvalue and the fifth eigenvalue in the feature data of the group of channels, respectively. The third target quantization bit width 3bitdepth may be the seventh quantization bit width bitdepth7 in the above formula (13).

It should be noted that the above formula (25) is only an example, and the formula for determining the third target scaling value s _c3 in the present application also includes the modification of the above formula (25), or the addition and addition of the above formula (25). Multiply or divide one or more coefficients, etc.

In another example, if the way that the encoding end quantizes the group of channels is a non-linear logarithmic uniform quantization method, then the encoding end can, according to the fifth eigenvalue and the fifth eigenvalue in the feature data of the group of channels, and a third target quantization bit width and a third base of the logarithmic function to determine a third target scaling value.

Optionally, the third target scaling value s _c3 may be determined according to the following formula (26):

Wherein, log _{log_base3} is the third base of the logarithmic function, and the third target quantization bit width may be the eighth quantization bit width in the above formula (15).

It should be noted that the above formula (26) is only an example, and the formula for determining the third target scaling value s _c3 in the present application also includes the modification of the above formula (26), or the addition and addition of the above formula (26). Multiply or divide one or more coefficients, etc.

In another example, if the way that the encoding end quantizes the group of channels is a nonlinear exponential uniform quantization method, the encoding end can use the fifth and sixth eigenvalues in the feature data of the group of channels, and The third target quantization bit width and the third base of the exponential function are determined.

Optionally, the third target scaling value s _c3 can be determined according to the following formula (27):

Wherein, e_base3 is the third base of the exponential function, and the third target quantization bit width may be the ninth quantization bit width bitdepth9 in the above formula (18).

It should be noted that the above formula (27) is only an example, and the formula for determining the third target scaling value s _c3 in the present application also includes the modification of the above formula (27), or the addition and addition of the above formula (27). Multiply or divide one or more coefficients, etc.

In this way, the decoding end can parse out the first information from the code stream, and use the linear uniform inverse quantization method to perform a linear uniform inverse quantization method according to the third target eigenvalue, the third target scaling value and the third target quantization bit width included in the first information. The channel's fixed-point feature data is inversely quantized.

Example 2, if the inverse quantization method for inverse quantization of the characteristic data of the fixed-point type of the set of channels is a nonlinear logarithmic uniform inverse quantization method, at this time, the first information includes the third target eigenvalue, the third target scaling value and the The third target quantization bit width, or the first information includes the third target eigenvalue, the third target scaling value, the third target quantization bit width and the third logarithmic base, or the first information includes the third target eigenvalue, the third Indication of the target scaling value and the third target quantization bit width and the third log base.

Specifically, if the first information includes the third target eigenvalue, the third target scaling value, and the third target quantization bit width, the decoding end uses the third target eigenvalue, the third target scaling value, the third target quantization bit width and the sum The default logarithmic base, which uses the nonlinear logarithmic uniform inverse quantization method to dequantize the fixed-point type feature data of this group of channels.

If the first information includes the third target eigenvalue, the third target scaling value, the third target quantization bit width and the third logarithmic base, the decoding end directly uses the third target eigenvalue, the third target scaling carried by the first information value, the third target quantization bit width and the third logarithmic base, and use the non-linear logarithmic uniform inverse quantization method to perform inverse quantization on the fixed-point type feature data of the group of channels.

If the first information includes the third target feature value, the third target scaling value, the third target quantization bit width and the third logarithmic base indication information, the third logarithmic base indication information is used to indicate that the The third logarithmic base is determined from the logarithmic bases. In this way, the decoding end parses the first information from the code stream, determines the third logarithmic base from the preset multiple logarithmic bases according to the indication information of the third logarithmic base, and then determines the third logarithmic base according to the third target eigenvalue, The third target scaling value, the third target quantization bit width, and the third logarithmic base are used to inversely quantize the fixed-point feature data of the group of channels by using a non-linear logarithmic uniform inverse quantization method.

Example 3, if the inverse quantization method for inverse quantization of the fixed-point type characteristic data of the group of channels is a nonlinear exponential uniform inverse quantization method, the first information includes the third target eigenvalue, the third target scaling value and the third target eigenvalue. The target quantization bit width, or the first information includes the third target feature value, the third target scaling value, the third target quantization bit width and the third exponent base, or the first information includes the third target feature value, the third target scaling value and the indication information of the third target quantization bit width and the third exponent base.

Specifically, if the first information includes the third target eigenvalue, the third target scaling value, and the third target quantization bit width, the decoding end uses the third target eigenvalue, the third target scaling value, the third target quantization bit width and the sum The default exponential base, which uses the nonlinear exponential uniform inverse quantization method to inverse quantize the fixed-point feature data of this group of channels.

If the first information includes the third target eigenvalue, the third target scaling value, the third target quantization bit width and the third exponent base, the decoding end directly uses the third target eigenvalue and the third target scaling value carried by the first information and the third target quantization bit width and the third exponent base, and use the non-linear exponential uniform inverse quantization method to inverse quantize the fixed-point type characteristic data of the group of channels.

If the first information includes the third target eigenvalue, the third target scaling value, the third target quantization bit width and the indication information of the third exponent base, the indication information of the third exponent base is used to indicate multiple exponents from preset The base of the third exponent is determined in the base. In this way, the decoding end parses the first information from the code stream, determines the third exponent base from the preset multiple exponent bases according to the indication information of the third exponent base, and then determines the third exponent base according to the third target eigenvalue, the third target The scaling value, the third target quantization bit width and the third exponent base are used to inversely quantize the fixed-point type feature data of the group of channels by using a non-linear exponential uniform inverse quantization method.

Example 4, if the inverse quantization method for inverse quantization of the characteristic data of the fixed-point type of the group of channels is the table look-up inverse quantization method, then the first information includes the index value of the quantization interval and the inverse quantization value of the quantization interval. Three correspondences, where the third correspondence is determined based on the pre-quantization value and the post-quantization value of the characteristic data of the group of channels. The index of the quantization interval can be understood as a fixed-point eigenvalue, and the inverse quantization value of the quantization interval can be understood as the weighted average value of each eigenvalue in the quantization interval, or the eigenvalue corresponding to the center position of the quantization interval. The weighted average value of each eigenvalue in the quantization interval may also be referred to as the eigenvalue corresponding to the center of the probability distribution of the quantization interval.

In a specific embodiment, the quantization method of the encoding end corresponds to the inverse quantization method of the decoding end. For example, when the encoding end uses a linear quantization method to quantize the characteristic data of the floating point type of the group of channels, the decoding end uses linear inverse quantization. way to inverse quantize the fixed-point feature data of the group of channels. If the decoding end uses the nonlinear logarithmic uniform quantization method to quantize the floating-point type feature data of the group of channels, the decoding end uses the nonlinear logarithmic uniform inverse quantization method to quantize the fixed-point number type feature data of the group of channels. Inverse quantization. If the decoding end uses the nonlinear exponential uniform quantization method to quantize the floating-point type feature data of the group of channels, the decoding end uses the nonlinear exponential uniform inverse quantization method to inverse quantize the fixed-point type feature data of the group of channels. . If the decoding end uses the table lookup quantization method to quantize the floating point type feature data of the group of channels, the decoding end uses the table lookup inverse quantization method to inverse quantize the fixed point type characteristic data of the group channel.

In some embodiments, if the inverse quantization method is linear uniform inverse quantization, its syntax structure is shown in Table 10:

Table 10

flag_channel: used to describe the symbol bit indicating the processing object of the decoding end. When it is 0, it means that all channels are uniformly inverse quantized. When it is 1, it means that each channel is inversely quantized. When it is 2, it means that each group of channels is quantized separately; here the value of flag_channel is 2;

group_num: The number of groups used to describe the feature data is group_num;

group_channel: The number of channels under each group used to describe the feature data is group_channel;

scale_num[i]: The scaling value used to describe the feature data under the i-th channel is scale_num[i], which can be understood as the above-mentioned third target scaling value;

min_num[i]: The minimum value used to describe the feature data of all channels under the ith group is min_num[i], which can be understood as the third target feature value above.

In some embodiments, if the inverse quantization method is nonlinear logarithmic function inverse quantization, its syntax structure is shown in Table 11:

Table 11

flag_channel: used to describe the sign bit indicating the processing object of the decoding end. When it is 0, it means that all channels are uniformly inverse quantized. When it is 1, it means that each channel is dequantized separately. When it is 2, it means that each group of channels is quantized separately;

flag_iquantization: used to describe the sign bit indicating the inverse quantization method at the decoding end. When it is 0, it means linear inverse quantization, when it is 1, it means nonlinear logarithmic inverse quantization, when it is 2, it means nonlinear exponential inverse quantization, and when it is 3, it means lookup table inverse quantification;

group_num: The number of groups used to describe the feature data is group_num;

min_num[i]: The minimum value used to describe the feature data of all channels under the ith group is min_num[i], which can be understood as the third target feature value above;

log_base: The base log_base used to describe the logarithmic function used in logarithmic inverse quantization, which can be understood as the third logarithmic base above.

In some embodiments, if the inverse quantization method is nonlinear exponential function inverse quantization, its syntax structure is shown in Table 12:

Table 12

group_num: The number of groups used to describe the feature data is group_num;

e_base: The base used to describe the exponential function used in logarithmic quantization is e_base, which can be understood as the above-mentioned third exponential base.

In some embodiments, if the inverse quantization method is look-up table inverse quantization, its syntax structure is as shown in Table 14:

Table 14

group_num: The number of groups used to describe the feature data is group_num;

hist_codebook[i]: The reconstructed codebook formed by describing the third correspondence between the index value of the quantization interval under the ith group and the inverse quantization value of the quantization interval is hist_codebook[i].

hist_codebook_num[i]: The size of the reconstructed codebook formed by the third correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval under the ith grouping is hist_codebook_num[i];

hist_codebook[i][j]: used to describe the inverse quantization value of the jth quantization interval index in the reconstructed codebook corresponding to the ith group hist_codebook[i][j].

In some embodiments, the decoder uses a default inverse quantization method to perform inverse quantization on the fixed-point type feature data of at least one channel.

The image encoding process is described above with reference to FIG. 3 to FIG. 7 , and the image decoding process at the decoding end is described below based on the foregoing embodiment.

The decoder that performs the image decoding process at the decoding end may be the decoder shown in FIG. 2 .

FIG. 7 is a schematic flowchart of an image decoding method 700 provided by an embodiment of the present application, as shown in FIG. 7 , including:

S701, decoding the code stream to obtain feature data of the current image, wherein the feature data of the current image includes the feature data of N channels, and the N is a positive integer;

S702. Decode the code stream to obtain first information, where the first information is used to instruct inverse quantization of feature data of at least one channel in the N channels;

S703. Perform inverse quantization on the feature data of at least one channel according to the first information.

It can be seen from the above that when the encoding end quantizes the feature data of the current image, the quantization is performed with the channel of the feature data as a consideration. Therefore, when inverse quantization is performed on the feature data at the decoding end, inverse quantization is also performed in consideration of the channel. Specifically, the decoder parses the code stream to obtain characteristic data of N channels of the current image and first information, and inversely quantizes the characteristic data of at least one of the N channels according to the first information.

In some embodiments, performing inverse quantization on the feature data of at least one channel according to the first information in the above S703 includes: according to the first information, inverse quantizing the feature data of the fixed-point type of at least one channel into at least one channel Characteristic data of the channel's float type.

In some embodiments, the inverse quantization method used by the decoder to perform inverse quantization on the fixed-point type feature data of at least one channel includes any one of the following: a linear uniform inverse quantization method, a nonlinear uniform inverse quantization method, or a look-up table inverse quantization. The nonlinear uniform inverse quantization method further includes nonlinear exponential function inverse quantization and nonlinear logarithmic function inverse quantization. It should be noted that the inverse quantization methods in the embodiments of the present application include but are not limited to the above several inverse quantization methods, and other inverse quantization methods can also be used to inverse quantize the characteristic data of fixed-point type. make restrictions.

In some embodiments, the inverse quantization method is default, that is, the decoder uses the default inverse quantization method to perform inverse quantization on the fixed-point type feature data of at least one channel according to the first information.

In some embodiments, the code stream includes second information, where the second information is used to indicate an inverse quantization method used when performing inverse quantization on the fixed-point feature data of at least one channel. In this case, the decoder may The first information uses the inverse quantization mode indicated by the second information to perform inverse quantization on the feature data of the fixed-point type of at least one channel.

In some embodiments, the first information in the code stream includes at least one parameter required for inverse quantization of the fixed-point type feature data of at least one channel. For example, the first information includes parameters corresponding to the inverse quantization method.

In some embodiments, according to the first information in the above S703, the manners of performing inverse quantization on the fixed-point type feature data of at least one channel of the N channels include but are not limited to the following:

Mode 1, if the first information indicates to perform inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels, use the same inverse quantization method to perform inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels ;

Mode 2, if the first information indicates that inverse quantization is performed on the characteristic data of the fixed-point type of each channel in the N channels, then for each channel, use the inverse quantization method corresponding to the channel to perform the inverse quantization of the fixed-point type of the channel. Inverse quantification of feature data;

Mode 3: If the first information indicates to perform inverse quantization on the fixed-point type feature data of M groups of channels, the N channels are divided into M groups of channels, and for each group of channels, the inverse quantization method corresponding to the group of channels is used. , perform inverse quantization on the fixed-point feature data of this group of channels.

The image decoding method provided by the present application obtains fixed-point type feature data of N channels of the current image by decoding the code stream; decodes the code stream to obtain first information, and the first information indicates the data of at least one of the N channels. The fixed-point feature data is inversely quantized, so that the decoder performs inverse quantization on the fixed-point feature data of at least one channel of the N channels according to the first information to obtain floating-point feature data of the current image. In the present application, the feature data output from the intermediate layer of the neural network is fixed-point, so that the technology in the existing video and image coding and decoding standards can be reused to decode the feature data, and at least one inverse quantization method is used at the same time. The fixed-point feature data of each channel is inversely quantized, thereby improving the decoding efficiency of the fixed-point feature data. In addition, the present application considers the channel information of the feature data in the inverse quantization process at the decoding end, and can process the feature data between different channels, thereby improving the reliability of the inverse quantization of the feature data.

The following describes in detail the process of inverse quantization using a quantization method for the characteristic data of the fixed-point type of all channels in the N channels with reference to FIG. 8 .

FIG. 8 is a schematic flowchart of an image decoding method 800 provided by an embodiment of the present application, including:

S801. Decode the code stream to obtain first information;

S802. According to the first information, use an inverse quantization method to perform inverse quantization on the fixed-point type feature data of all channels in the N channels.

For different inverse quantization methods, the parameters included in the first information may be different. The following describes the process of inverse quantization of the fixed-point feature data of all channels in the N channels using different inverse quantization methods for the decoder. .

In some embodiments, if the inverse quantization method is a linear uniform inverse quantization method, the above S802 includes the following S802-A1 and S802-A2:

S802-A1, analyzing the first information to obtain the first target feature value, the first target scaling value and the first target quantization bit width;

S802-A2. According to the first target feature value, the first target scaling value and the first target quantization bit width, use a linear uniform inverse quantization method to perform inverse quantization on the fixed-point feature data of all channels in the N channels.

Wherein, the above-mentioned first target characteristic value is one characteristic value in the characteristic data of all the channels in the N channels, and the above-mentioned first target scaling value is the corresponding scaling value when the characteristic data of all the channels in the N channels are quantized, The above-mentioned first target quantization bit width is the quantization bit width corresponding to the characteristic data of all channels in the N channels during quantization.

Optionally, the above-mentioned first objective feature value is the smallest feature value among the feature data of all channels in the N channels of the current image.

In this embodiment, the first information includes the first target eigenvalue, the first target scaling value and the first target quantization bit width required by the linear uniform inverse quantization method. In this way, the decoder can A target eigenvalue, a first target scaling value and a first target quantization bit width are used to inverse quantize the fixed-point feature data of all channels in the N channels by using a linear uniform inverse quantization method. For example, the decoder determines several bits as an inverse quantization value according to the first target quantization bit width, and then, according to the first target eigenvalue and the first target scaling value, uses a linear uniform inverse quantization method to quantify all the N channels. The feature data of the channel is inverse quantized.

For example, the decoder performs inverse quantization on the fixed-point type feature data of all channels according to the following formula (28):

Among them, y _cij is the quantized value of the i-th row and the j-th column of the c-th channel, s _c1 is the first target scaling value of the feature data under all channels, x1 _cmin is the first target feature value of the feature data under all channels, x _cij is the reconstruction value or inverse quantization value of the i-th row and the j-th column of the c-th channel.

According to different nonlinear functions, the above-mentioned nonlinear uniform quantization methods include nonlinear logarithmic uniform inverse quantization methods and nonlinear exponential uniform inverse quantization methods.

In some embodiments, if the inverse quantization method is a nonlinear logarithmic uniform inverse quantization method, the above S802 includes the following S802-B1 and S802-B2:

S802-B1, according to the first information, determine the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base;

S802-B2. According to the first target eigenvalue, the first target scaling value, the first target quantization bit width, and the first logarithmic base, use a nonlinear logarithmic uniform inverse quantization method to quantify the fixed-point type of all channels in the N channels inverse quantization of the feature data.

Wherein, according to the different parameters included in the first information, the above S802-B1 according to the first information determines the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base. Not limited to the following:

Mode 1, if the first information includes the first target eigenvalue, the first target scaling value, the first target quantization bit width, and the first logarithmic base, the decoder can directly analyze the first information to obtain the first target eigenvalue. , a first target scaling value and a first target quantization bit width and a first logarithmic base.

Mode 2, if the first information includes the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base indication information, in this way, the decoder parses the first information to obtain the first indication information of the target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base; according to the indication information of the first logarithmic base, from the preset multiple logarithmic bases, determine the first Logarithmic base.

Mode 3, if the first information includes the first target feature value, the first target scaling value and the first target quantization bit width, but does not include the first logarithmic base, in this way, the decoder obtains the first target feature by parsing the first information value, a first target scaling value, and a first target quantization bit width, and determine the default log base as the first log base.

After the decoder determines the first target eigenvalue, the first target scaling value, the first target quantization bit width and the first logarithmic base according to the above method, the first target eigenvalue, the first target scaling value and the first target quantization Bit width and the first logarithmic base, use the nonlinear logarithmic uniform inverse quantization method to dequantize the fixed-point type feature data of all channels in the N channels.

For example, the decoder performs inverse quantization on the fixed-point type feature data of all channels according to the following formula (29):

where log_base ₁ is the first logarithmic base.

In some embodiments, if the inverse quantization method is a nonlinear exponential uniform inverse quantization method, the above S802 includes the following S802-C1 and S802-C2:

S802-C1, according to the first information, determine the first target feature value, the first target scaling value, the first target quantization bit width and the first exponent base;

S802-C2. According to the first target eigenvalue, the first target scaling value, the first target quantization bit width and the first exponential base, use a non-linear exponential uniform inverse quantization method to perform a uniform inverse quantization method on the fixed-point number type of all channels in the N channels. The feature data is inverse quantized.

In some embodiments, according to different parameters included in the first information, in S802-C1, the first target feature value, the first target scaling value, the first target quantization bit width and the first exponent base are determined according to the first information. The methods include but are not limited to the following:

Mode 1: If the first information includes the first target feature value, the first target scaling value, the first target quantization bit width, and the first exponent base, the decoder directly parses the first information to obtain the first target feature value, the first target feature value, and the first index base. A target scaling value, a first target quantization bit width, and a first exponent base.

Mode 2, if the first information includes the first target eigenvalue, the first target scaling value, the first target quantization bit width and the indication information of the first exponent base, then the decoder parses the first information to obtain the first target eigenvalue, The indication information of the first target scaling value, the first target quantization bit width and the first exponent base; and according to the indication information of the first exponent base, the first exponent base is determined from a plurality of preset exponent bases.

Mode 3: If the first information includes the first target feature value, the first target scaling value, and the first target quantization bit width, the decoder parses the first information to obtain the first target feature value, the first target scaling value, the first target The target quantization bit width, and the default exponent base is determined as the first exponent base.

After the decoder determines the first target eigenvalue, the first target scaling value, the first target quantization bit width and the first exponent base according to the above method, the first target eigenvalue, the first target scaling value, the first target quantization bit The width and the first exponent base are used to inversely quantize the fixed-point feature data of all channels in the N channels by using the nonlinear exponential uniform inverse quantization method.

For example, the decoder performs inverse quantization on the fixed-point type feature data of all channels according to the following formula (30):

Among them, e_base ₁ is the base of the first exponent.

In some embodiments, if the inverse quantization method is a look-up table inverse quantization method, the above S802 includes the following S802-D1 to S802-D3:

S802-D1. Determine the first correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval, where the first correspondence is based on the value before quantization and the value after quantization of the characteristic data of all channels in the N channels value is determined;

S802-D2. For the characteristic data of each fixed-point number type of all channels in the N channels, use the value of the characteristic data of the fixed-point number type as the index of the quantization interval, and in the first correspondence, query the fixed-point number type of the characteristic data. The target inverse quantization value corresponding to the value of the feature data;

S802-D3: Determine the target inverse quantization value as a floating-point value of the feature data of the fixed-point type.

The corresponding relationship between the index value of the quantization interval and the inverse quantization value of the quantization interval is default; or, the first information includes the corresponding relationship between the index value of the quantization interval and the inverse quantization value of the quantization interval.

Optionally, the inverse quantization value of the quantization interval is the eigenvalue corresponding to the center position in the quantization interval, or the weighted average value of each eigenvalue in the quantization interval. The weighted average value of each eigenvalue in the quantization interval may also be referred to as the eigenvalue corresponding to the center of the probability distribution of the quantization interval.

The following describes in detail the process of inverse quantization using an inverse quantization method for the feature data of the fixed-point type of each channel of the N channels of the current image in conjunction with FIG. 9 .

FIG. 9 is a schematic flowchart of an image decoding method 900 provided by an embodiment of the present application, including:

S901, for each channel in the N channels, decode the code stream to obtain characteristic data of the fixed-point type of the channel;

S902. According to the first information, use the inverse quantization mode corresponding to the channel to perform inverse quantization on the feature data of the fixed-point type of the channel.

Optionally, the inverse quantization method includes linear uniform inverse quantization, nonlinear function inverse quantization, and look-up table inverse quantization.

In some embodiments, if the inverse quantization method is a linear uniform inverse quantization method, the above S902 includes the following S902-A1 and S902-A2:

S902-A1, parse the first information to obtain the second target feature value, the second target scaling value and the second target quantization bit width;

S902-A2. According to the second target feature value, the second target scaling value, and the second target quantization bit width, use a linear uniform inverse quantization method to perform inverse quantization on the fixed-point type feature data of the channel.

Wherein, the second target feature value is a feature value in the feature data of the group of channels, the second target scaling value is the scaling value corresponding to the feature data of the channel during quantization, and the second target quantization bit width is the feature of the channel The corresponding quantization bit width when the data is quantized.

Optionally, the second target feature value is the smallest feature value in the feature data of the channel.

In this embodiment, the first information includes the second target eigenvalue, the second target scaling value, and the second target quantization bit width required by the linear uniform inverse quantization method. In this way, the decoder can The second target eigenvalue, the second target scaling value and the second target quantization bit width are inversely quantized using a linear uniform inverse quantization method for the fixed-point type feature data of the channel. For example, the decoder determines several bits as an inverse quantization value according to the second target quantization bit width, and then, according to the second target feature value and the second target scaling value, uses a linear uniform inverse quantization method for the feature data of the channel Do inverse quantization.

For example, according to the following formula (31), inverse quantization is performed on the fixed-point type feature data of the channel:

Among them, it is assumed that the current channel is the c-th channel, y _cij is the quantized value of the i-th row and the j-th column of the c-th channel, s _c2 is the second target scaling value of the feature data under this channel, and x2 _cmin is the feature under this channel. The second target eigenvalue of the data, x _cij is the reconstructed value of the i-th row and the j-th column of the c-th channel.

In some embodiments, if the inverse quantization method is a nonlinear logarithmic uniform inverse quantization method, the above S902 includes the following S902-B1 and S902-B1:

S902-B1, according to the first information, determine the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base;

S902-B2. According to the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second logarithmic base, use the non-linear logarithmic uniform inverse quantization method to obtain the fixed-point type feature data of the channel Do inverse quantization.

Wherein, according to different parameters included in the first information, the above-mentioned S902-B1, according to the first information, determines the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base. But not limited to the following:

Mode 1, if the first information includes the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second logarithmic base, the decoder directly parses the first information to obtain the second target eigenvalue, the second target eigenvalue, and the second logarithmic base. Two target scaling values, a second target quantization bit width, and a second log base.

Method 2: If the first information includes the indication information of the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base, the decoder parses the first information to obtain the second target feature value , the second target scaling value, the second target quantization bit width and the indication information of the second logarithmic base; and according to the indication information of the second logarithmic base, from the preset multiple logarithmic bases, determine the second logarithm base.

Mode 3: If the first information includes the second target feature value, the second target scaling value and the second target quantization bit width, the decoder parses the first information to obtain the second target feature value, the second target scaling value and the second target scaling value. The target quantization bit width, and determines the default log base as the second log base.

After the decoder determines the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second logarithmic base according to the above method, the second target eigenvalue, the second target scaling value, the second target quantization Bit width and second logarithmic base, use the non-linear logarithmic uniform inverse quantization method to dequantize the fixed-point type feature data of this channel.

For example, the decoder performs inverse quantization on the fixed-point type feature data of all channels according to the following formula (32):

where log_base ₂ is the second logarithmic base.

In some embodiments, if the inverse quantization method corresponding to the channel is a nonlinear exponential uniform inverse quantization method, the above S902 includes the following S902-C1 and S902-C2:

S902-C1, according to the first information, determine the second target feature value, the second target scaling value, the second target quantization bit width and the second exponent base;

S902-C2. According to the second target feature value, the second target scaling value, the second target quantization bit width and the second exponential base, use the nonlinear exponential uniform inverse quantization method to inverse the feature data of the fixed-point type of the channel quantify.

In some embodiments, according to different parameters included in the first information, the above S902-B1 determines the second target feature value, the second target scaling value, the second target quantization bit width, and the second exponent base according to the first information. Methods include but are not limited to the following:

Mode 1, if the first information includes the second target eigenvalue, the second target scaling value, the second target quantization bit width, and the second exponent base, the decoder directly parses the first information to obtain the second target eigenvalue, the second target eigenvalue, and the second index base. The target scaling value, the second target quantization bit width, and the second exponent base.

Mode 2: If the first information includes the second target feature value, the second target scaling value, the second target quantization bit width, and the second logarithmic base indication information, the decoder parses the first information, and obtains that the first information includes the first information. The indication information of the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base; and according to the indication information of the second logarithmic base, from the preset multiple exponential bases, determine the first Two exponential bases.

Mode 3: If the first information includes the second target feature value, the second target scaling value, and the second target quantization bit width, the decoder parses the first information to obtain the second target feature value, the second target scaling value, and the second target scaling value. Target quantization bit width, and establishes the default exponent base as the second exponent base.

After the decoder determines the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second exponent base according to the above method, according to the second target eigenvalue, the second target scaling value, the second target quantization bit The width and the second exponential base are used to inversely quantize the fixed-point feature data of the channel using the nonlinear exponential uniform inverse quantization method.

For example, the decoder performs inverse quantization on the fixed-point feature data of the channel according to the following formula (33):

Among them, e_base ₂ is the second exponent base.

In some embodiments, if the inverse quantization method corresponding to the channel is a look-up table inverse quantization method, the above S902 includes S902-D1 to S902-D3:

S902-D1, determine the second correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval, and the second correspondence is determined based on the value before quantization and the value after quantization of the characteristic data of the channel;

S902-D2. For each characteristic data of the fixed-point type in the channel, use the value of the characteristic data of the fixed-point type as the index of the quantization interval, and in the second correspondence, query the characteristic data of the fixed-point type. The target inverse quantization value corresponding to the value;

S902-D3: Determine the target inverse quantization value as the value of the floating point type of the feature data of the fixed point type.

In the following, in conjunction with FIG. 10 , the process of inverse quantization using an inverse quantization method is described in detail for the characteristic data of the floating point type of each group of channels.

FIG. 10 is a schematic flowchart of an image decoding method 1000 provided by an embodiment of the present application, including:

S101, for each group of channels, decode the code stream to obtain characteristic data of the fixed-point type of the group of channels;

S102. According to the first information, use an inverse quantization method corresponding to the group of channels to perform an inverse quantization method on the characteristic data of the fixed-point number type of the group of channels.

In some embodiments, if the inverse quantization method corresponding to the group of channels is a linear uniform inverse quantization method, the above S102 includes the following S102-A1 and S102-A2:

S102-A1, analyzing the first information to obtain the third target feature value, the third target scaling value and the third target quantization bit width;

S102-A2. According to the third target feature value, the third target scaling value, and the third target quantization bit width, use a linear uniform inverse quantization method to perform inverse quantization on the fixed-point type feature data of the group of channels.

Among them, the third target feature value is a feature value in the feature data of the group of channels, the third target scaling value is the scaling value corresponding to the feature data of the group of channels during quantization, and the third target quantization bit width is the group of channels. The corresponding quantization bit width of the feature data during quantization.

Optionally, the third target feature value is the smallest feature value in the feature data of the group of channels.

In this embodiment, the first information includes the third target eigenvalue, the third target scaling value, and the third target quantization bit width required by the linear uniform inverse quantization method. In this way, the decoder can The three target eigenvalues, the third target scaling value, and the third target quantization bit width are inversely quantized using a linear uniform inverse quantization method to perform inverse quantization on the fixed-point type feature data of the group of channels. For example, the decoder determines several bits as an inverse quantization value according to the third target quantization bit width, and then, according to the third target feature value and the third target scaling value, uses a linear uniform inverse quantization method for the characteristics of the group of channels Data is dequantified.

For example, according to the following formula (34), inverse quantization is performed on the fixed-point type feature data of this group of channels:

Among them, the c-th channel is a channel in the current group of channels, y _cij is the quantized value of the c-th channel in the i-th row and the j-th column, s _c3 is the third target scaling value of the feature data under this group of channels, and x3 _cmin is The third target eigenvalue of the feature data under this group of channels, x _cij is the reconstructed value of the i-th row and the j-th column of the c-th channel.

The above-mentioned nonlinear uniform quantization methods include nonlinear logarithmic uniform inverse quantization methods and nonlinear exponential uniform inverse quantization methods.

In some embodiments, if the inverse quantization method corresponding to the group of channels is a nonlinear logarithmic uniform inverse quantization method, the above S102 includes the following S102-B1 and S102-B2:

S102-B1, according to the first information, determine the third target feature value, the third target scaling value, the third target quantization bit width and the third logarithmic base;

S102-B2. According to the third target eigenvalue, the third target scaling value, the third target quantization bit width and the third logarithmic base, use the non-linear logarithmic uniform inverse quantization method to obtain the fixed-point type feature of the set of channels Data is dequantified.

In some embodiments, the manners of determining the third target feature value, the third target scaling value, the third target quantization bit width and the third logarithmic base in the above S102-B1 include but are not limited to the following:

Mode 1, if the first information includes the third target eigenvalue, the third target scaling value, the third target quantization bit width, and the third logarithmic base, the decoder directly parses the first information to obtain the third target eigenvalue, the third target eigenvalue, and the third logarithmic base. Three target scaling values, a third target quantization bit width, and a third log base.

Method 2: If the first information includes the third target eigenvalue, the third target scaling value, the third target quantization bit width and the third logarithmic base indication information, the decoder parses the first information to obtain the third target eigenvalue , the third target scaling value, the third target quantization bit width and the third logarithmic base; and according to the third logarithmic base instruction information, from the preset multiple logarithmic bases, determine the third logarithm base;

Mode 3: If the first information includes the third target eigenvalue, the third target scaling value and the third target quantization bit width, the decoder parses the first information to obtain the third target eigenvalue, the third target scaling value and the third target eigenvalue. The target quantization bit width, and determines the default log base as the third log base.

After the decoder determines the third target eigenvalue, the third target scaling value, the third target quantization bit width and the third logarithmic base according to the above method, the decoder determines the third target eigenvalue, the third target scaling value, the third target quantization The bit width and the third logarithmic base are used to inversely quantize the fixed-point feature data of this group of channels using a non-linear logarithmic uniform inverse quantization method.

For example, according to the following formula (35), inverse quantization is performed on the fixed-point type feature data of this group of channels:

where log_base ₃ is the third logarithmic base.

In some embodiments, if the inverse quantization method corresponding to the group of channels is a nonlinear exponential uniform inverse quantization method, the above S102 includes the following S102-C1 and S102-C2:

S102-C1, according to the first information, determine the third target feature value, the third target scaling value, the third target quantization bit width and the third exponent base;

S102-C2. According to the third target eigenvalue, the third target scaling value, the third target quantization bit width and the third exponent base, use the non-linear exponential uniform inverse quantization method to perform the fixed-point number type feature data on the set of channels. Inverse quantization.

In some embodiments, the manners of determining the third target feature value, the third target scaling value, the third target quantization bit width, and the third exponent base in S102-C1 include but are not limited to the following manners:

Mode 1, if the first information includes the third target eigenvalue, the third target scaling value, the third target quantization bit width, and the third exponent base, the decoder directly parses the first information to obtain the third target eigenvalue, the third target eigenvalue, and the third index base. The target scaling value, the third target quantization bit width, and the third exponent base.

Mode 2: If the first information includes the third target feature value, the third target scaling value, the third target quantization bit width and the third logarithmic base indication information, the decoder parses the first information, and obtains that the first information includes the third target quantization bit width and the third logarithmic base. The indication information of the three target feature values, the third target scaling value, the third target quantization bit width and the third logarithmic base; and according to the indication information of the third logarithmic base, from the preset multiple index bases, determine the three-exponential base;

Mode 3: If the first information includes the third target feature value, the third target scaling value, and the third target quantization bit width, the decoder parses the first information to obtain the third target feature value, the third target scaling value, and the third target scaling value. The target quantization bit width, and determines the default exponent base as the third exponent base.

After the decoder determines the third target eigenvalue, the third target scaling value, the third target quantization bit width and the third exponent base according to the above method, according to the third target eigenvalue, the third target scaling value, the third target quantization bit The width and the third exponent base are inversely quantized using the non-linear exponential uniform inverse quantization method to perform inverse quantization on the fixed-point type feature data of this group of channels.

For example, according to the following formula (36), inverse quantization is performed on the fixed-point type feature data of this group of channels:

Among them, e_base ₃ is the third exponent base.

In some embodiments, if the inverse quantization method corresponding to the group of channels is a table look-up inverse quantization method, the above S102 includes the following S102-D1 to S102-D3:

S102-D1, determine the third correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval, and the third correspondence is determined based on the value before quantization and the value after quantization of the characteristic data of this group of channels;

S102-D2. For each feature data of the fixed-point number type in the set of channels, use the value of the feature data of the fixed-point number type as the index of the quantization interval, and in the third correspondence, query the feature data of the fixed-point number type The value of the corresponding target inverse quantization value;

S102-D3: Determine the target inverse quantization value as a floating-point value of the feature data of the fixed-point type.

It should be understood that the above-mentioned FIG. 3 to FIG. 10 are only examples of the present application, and should not be construed as a limitation on the present application.

The preferred embodiments of the present application have been described in detail above with reference to the accompanying drawings. However, the present application is not limited to the specific details of the above-mentioned embodiments. Within the scope of the technical concept of the present application, various simple modifications can be made to the technical solutions of the present application. These simple modifications all belong to the protection scope of the present application. For example, the specific technical features described in the above-mentioned specific embodiments can be combined in any suitable manner unless they are inconsistent. In order to avoid unnecessary repetition, this application does not describe any possible combination. State otherwise. For another example, the various embodiments of the present application can also be combined arbitrarily, as long as they do not violate the idea of the present application, they should also be regarded as the content disclosed in the present application.

It should also be understood that, in the various method embodiments of the present application, the size of the sequence numbers of the above-mentioned processes does not mean the sequence of execution, and the execution sequence of each process should be determined by its functions and internal logic, and should not be dealt with in the present application. The implementation of the embodiments constitutes no limitation. In addition, in this embodiment of the present application, the term "and/or" is only an association relationship for describing associated objects, indicating that there may be three kinds of relationships. Specifically, A and/or B can represent three situations: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character "/" in this document generally indicates that the related objects are an "or" relationship.

The method embodiments of the present application are described in detail above with reference to FIGS. 3 to 10 , and the apparatus embodiments of the present application are described in detail below with reference to FIG. 14 .

FIG. 11 is a schematic block diagram of a video encoder 10 provided by an embodiment of the present application.

As shown in Figure 11, the video encoder 10 includes:

an acquisition unit 110 for acquiring the current image to be encoded;

A feature extraction unit 120, configured to input the current image into a neural network to obtain feature data of the current image, where the feature data of the current image includes feature data of N channels, and N is a positive integer;

a quantization unit 130, configured to quantify the characteristic data of at least one channel in the N channels;

The encoding unit 140 is configured to encode the quantized feature data of the at least one channel to obtain a code stream, where the code stream includes first information, and the first information is used to indicate whether the N channels are to be The feature data of at least one channel is inverse quantized.

In some embodiments, the quantization unit 130 is specifically configured to quantize the feature data of the floating point type of at least one channel of the N channels into the feature data of the fixed point type.

In some embodiments, the quantization method for quantizing the floating-point type feature data of at least one channel of the N channels includes any one of the following: a linear uniform quantization method, a nonlinear exponential uniform quantization method, and a nonlinear logarithmic method. Uniform quantization method, look-up table quantization method.

In some embodiments, the quantization unit 130 is specifically configured to use the same quantization method to quantize the floating-point type feature data of all channels in the N channels; The feature data of the floating-point number type is quantized by using different quantization methods respectively; or, the N channels are grouped, and the feature data of the floating-point number type of each group of channels is quantized using a quantization method respectively.

In some embodiments, if the quantization method is a linear uniform quantization method, the quantization unit 130 is specifically configured to obtain a preset first quantization bit width, and the floating point type feature data of all channels in the N channels. The first eigenvalue and the second eigenvalue of the The feature data of the channel's floating point type is quantized.

In some embodiments, if the quantization method is a nonlinear logarithmic uniform quantization method, the quantization unit 130 is specifically configured to obtain a preset second quantization bit width and the first base of the logarithmic function, and the N channels The first eigenvalue and the second eigenvalue in the feature data of the floating point type of all channels in the The first base is to use the nonlinear logarithmic uniform quantization method to quantize the floating point type feature data of each of the N channels.

In some embodiments, if the quantization method is a nonlinear exponential uniform quantization method, the quantization unit 130 is specifically configured to obtain a preset third quantization bit width and the first base of the exponential function, as well as all of the N channels. the first eigenvalue and the second eigenvalue in the feature data of the floating point type of the channel; according to the first eigenvalue and the second eigenvalue, the third quantization bit width and the first base of the exponential function , using the nonlinear exponential uniform quantization method to quantize the floating point type feature data of each channel in the N channels.

In some embodiments, if the quantization method is a look-up table quantization method, the quantization unit 130 specifically sorts the floating-point type feature data of all channels in the N channels according to the value size, and obtains the sorted first characteristic data; dividing the sorted first characteristic data into a plurality of first quantization intervals, wherein each first quantization interval includes the same amount of characteristic data; for each of the first quantization intervals, all The value of the feature data in the first quantization interval is quantized as an index value of the first quantization interval.

Optionally, the first eigenvalue is the smallest eigenvalue in the floating-point type feature data of all channels in the N channels, and the second eigenvalue is the floating-point number of all channels in the N channels The largest eigenvalue in the eigendata of type.

In some embodiments, if the quantization method is a linear uniform quantization method, the quantization unit 130 specifically obtains a preset fourth quantization bit width for each of the N channels, and the floating value of the channel. The third eigenvalue and the fourth eigenvalue in the feature data of the point type; according to the third eigenvalue and the fourth eigenvalue, and the fourth quantization bit width, using the linear uniform quantization method, the The feature data of the channel's floating point type is quantized.

In some embodiments, if the quantization method is a nonlinear logarithmic uniform quantization method, the quantization unit 130 specifically obtains a preset fifth quantization bit width and a logarithmic function for each channel of the N channels The second base of , and the third eigenvalue and the fourth eigenvalue in the characteristic data of the floating point type of the channel; according to the third and fourth eigenvalues, and the fifth quantization bit width and The second base of the logarithmic function uses the nonlinear logarithmic uniform quantization method to quantize the floating point type feature data of the channel.

In some embodiments, if the quantization method is a non-linear exponential uniform quantization method, the quantization unit 130 specifically obtains the preset sixth quantization bit width and the sixth index of the exponential function for each channel of the N channels. A base-two number, and the third eigenvalue and the fourth eigenvalue in the feature data of the floating point type of the channel; according to the third and fourth eigenvalues, the sixth quantization bit width and the The second base of the exponential function uses the nonlinear exponential uniform quantization method to quantize the floating point type feature data of the channel.

In some embodiments, if the quantization method is a look-up table quantization method, the quantization unit 130 specifically uses, for each channel in the N channels, the feature data of the floating point type of the channel according to the value size. Perform sorting to obtain the sorted second feature data under the channel; divide the sorted second feature data under the channel into a plurality of second quantization intervals, wherein each second quantization interval includes the same number of features data; for each of the second quantization intervals, the value of the feature data in the second quantization interval is quantized as an index value of the second quantization interval.

In some embodiments, if the quantization method is a linear uniform quantization method, the quantization unit 130 obtains a preset seventh quantization bit width for each group of channels and the characteristic data of the floating point type of the group of channels. The fifth eigenvalue and the sixth eigenvalue in the Quantize the feature data of the floating point type of each channel.

In some embodiments, if the quantization method is a non-linear logarithmic uniform quantization method, the quantization unit 130 specifically obtains a preset eighth quantization bit width and the third base of the logarithmic function for each group of channels, and the fifth eigenvalue and the sixth eigenvalue in the feature data of the floating point type of the set of channels; according to the fifth eigenvalue and the sixth eigenvalue, and the eighth quantization bit width and the logarithmic function The third base of , using the nonlinear logarithmic uniform quantization method to quantize the floating point type feature data of each channel in the group of channels.

In some embodiments, if the quantization method is a nonlinear exponential uniform quantization method, the quantization unit 130 specifically obtains a preset ninth quantization bit width and the third base of the exponential function for each group of channels, and the The fifth eigenvalue and the sixth eigenvalue in the feature data of the floating point type of the group channel; according to the fifth eigenvalue and the sixth eigenvalue, and the ninth quantization bit width and the third eigenvalue of the exponential function Base, using the nonlinear logarithmic uniform quantization method to quantize the floating point type feature data of each channel in the group of channels.

In some embodiments, if the quantization method is a look-up table quantization method, the quantization unit 130 specifically uses, for each group of channels, to sort the characteristic data of the floating point type of the group of channels according to the value size, to obtain the The third feature data sorted under the set of channels; the third feature data sorted under the set of channels is divided into a plurality of third quantization intervals, wherein each third quantization interval includes the same amount of feature data; for each each of the third quantization intervals, and quantizing the value of the feature data in the third quantization interval into an index value of the third quantization interval.

Optionally, the fifth characteristic value is the largest characteristic value in the floating point type characteristic data of the channel group, and the sixth characteristic value is the smallest characteristic value in the floating point type characteristic data of the group channel. value.

In some embodiments, the first information indicates inverse quantization of fixed-point type feature data of all of the N channels; alternatively, the first information indicates that each of the N channels is inverse-quantized The feature data of the fixed-point number type of the channel is respectively inverse-quantized; or, the first information indicates that the feature data of the fixed-point number type of each group of channels in the M groups of channels is respectively inverse-quantized, wherein the M groups of channels are Obtained by grouping the N channels, each group of channels includes at least one channel in the N channels.

In some embodiments, the inverse quantization method used when performing inverse quantization on the feature data of the fixed-point type of the at least one channel includes any one of the following: a linear uniform inverse quantization method, a nonlinear exponential uniform inverse quantization method, a non-linear uniform inverse quantization method, and a non-linear uniform inverse quantization method. Linear logarithmic uniform inverse quantization method, look-up table inverse quantization method.

In some embodiments, the first information includes at least one parameter required for inverse quantization of the fixed-point type feature data of the at least one channel.

In some embodiments, the first information indicates that inverse quantization is performed on the fixed-point feature data of all channels in the N channels, and the first information includes any one of the following:

If the inverse quantization method for performing inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels is a linear uniform inverse quantization method, the first information includes a first target feature value, a first target scaling value and the first target quantization bit width;

If the inverse quantization method for inverse quantization of the fixed-point type feature data of all channels in the N channels is a nonlinear logarithmic uniform inverse quantization method, the first information includes a first target eigenvalue, a first A target scaling value and a first target quantization bit width, or the first information includes a first target feature value, a first target scaling value, a first target quantization bit width, and a first logarithmic base, or the first information includes Indication information of the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base;

If the inverse quantization method for inverse quantization of the fixed-point type characteristic data of all channels in the N channels is a nonlinear exponential uniform inverse quantization method, the first information includes a first target eigenvalue, a first target The scaling value and the first target quantization bit width, or the first information includes the first target feature value, the first target scaling value, the first target quantization bit width and the first exponent base, or the first information includes the first The indication information of the target feature value, the first target scaling value, the first target quantization bit width and the first exponent base;

If the inverse quantization method for performing inverse quantization on the fixed-point type characteristic data of all channels in the N channels is a table look-up inverse quantization method, the first information includes the index value of the quantization interval and the inverse quantization of the quantization interval the first correspondence between the values, the first correspondence is determined based on the pre-quantized value and the quantized value of the characteristic data of all channels in the N channels;

The first target feature value is one feature value in the feature data of all channels in the N channels, and the first target scaling value corresponds to the feature data of all channels in the N channels during quantization The first target quantization bit width is the quantization bit width corresponding to the characteristic data of all channels in the N channels during quantization.

Optionally, the first target feature value is the minimum value of feature data of all channels in the N channels.

In some embodiments, the first information indicates that inverse quantization is performed on the fixed-point type feature data of each channel in the N channels, and for each channel, the first information includes any one of the following :

If the inverse quantization method for inverse quantization of the fixed-point type feature data of the channel is a linear uniform inverse quantization method, the first information includes the second target feature value, the second target scaling value, and the second target quantization bit width ;

If the feature data of the fixed-point type of the channel is inversely quantized into a non-linear logarithmic uniform inverse quantization method in an inverse quantization manner, the first information includes a second target feature value, a second target scaling value, and a second target The quantization bit width, or the first information includes the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second logarithmic base, or the first information includes the second target eigenvalue, the first Two target scaling values and the indication information of the second target quantization bit width and the second logarithmic base;

If the inverse quantization method for inverse quantization of the fixed-point type feature data of the channel is a nonlinear exponential uniform inverse quantization method, the first information includes the second target feature value, the second target scaling value, and the second target quantization bit width, or the first information includes the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second exponent base, or the first information includes the second target eigenvalue, the second target scaling value and indication information of the second target quantization bit width and the second exponent base;

If the inverse quantization method for inverse quantization of the fixed-point type characteristic data of the channel is the table look-up inverse quantization method, the first information includes the second correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval relationship, the second corresponding relationship is determined based on the value before quantization and the value after quantization of the characteristic data of the channel;

Wherein, the second target feature value is a feature value in the feature data of the channel, the second target scaling value is the scaling value corresponding to the feature data of the channel during quantization, and the second target quantization bit width It is the corresponding quantization bit width of the feature data of this channel during quantization.

Optionally, the second target feature value is the minimum value of feature data of the channel.

In some embodiments, the first information indicates that inverse quantization is performed on the fixed-point feature data of M groups of channels, respectively, and for each group of channels, the first information includes any one of the following:

If the inverse quantization is performed on the group of channels so that the inverse quantization method is a linear uniform inverse quantization method, the first information includes the third target eigenvalue, the third target scaling value and the third target quantization bit width;

If the inverse quantization method for inverse quantization of the fixed-point type feature data of the group of channels is a nonlinear logarithmic uniform inverse quantization method, the first information includes the third target eigenvalue, the third target scaling value, and the third target scaling value. The target quantization bit width, or the first information includes the third target eigenvalue, the third target scaling value, the third target quantization bit width, and the third logarithmic base, or the first information includes the third target eigenvalue, Indication information of the third target scaling value, the third target quantization bit width and the third logarithmic base;

If the inverse quantization method for inverse quantization of the fixed-point type feature data of the group of channels is a nonlinear exponential uniform inverse quantization method, the first information includes the third target feature value, the third target scaling value, and the third target Quantization bit width, or the first information includes a third target eigenvalue, a third target scaling value, a third target quantization bit width, and a third exponent base, or the first information includes a third target eigenvalue, a third Indication information of the target scaling value, the third target quantization bit width and the third exponent base;

If the inverse quantization method for performing inverse quantization on the fixed-point type feature data of the group of channels is the table look-up inverse quantization method, the first information includes the third index value between the index value of the quantization interval and the inverse quantization value of the quantization interval. Correspondence, the third correspondence is determined based on the value before quantization and the value after quantization of the characteristic data of this group of channels;

The M groups of channels are obtained by grouping the N channels, each group of channels includes at least one channel in the N channels, and the third target feature value is in the feature data of the group of channels. A characteristic value of , the third target scaling value is the corresponding scaling value of the characteristic data of this group of channels during quantization, and the third target quantization bit width is the corresponding quantization bit width of the characteristic data of this group of channels during quantization .

Optionally, the third target feature value is the minimum value of feature data of the group of channels.

In some embodiments, the code stream further includes second information, where the second information is used to indicate an inverse quantization method used when performing inverse quantization on the characteristic data of the fixed-point type of the at least one channel.

It should be understood that the apparatus embodiments and the method embodiments may correspond to each other, and similar descriptions may refer to the method embodiments. To avoid repetition, details are not repeated here. Specifically, the video encoder 10 shown in FIG. 11 may execute the methods of the embodiments of the present application, and the aforementioned and other operations and/or functions of the various units in the video encoder 10 are for implementing the methods in the methods 300 to 600, respectively. The corresponding process, for the sake of brevity, will not be repeated here.

FIG. 12 is a schematic block diagram of a video decoder 20 provided by an embodiment of the present application.

As shown in Figure 12, the video decoder 20 may include:

a decoding unit 210, configured to decode the code stream to obtain characteristic data of the current image, where the characteristic data of the current image includes characteristic data of N channels, and N is a positive integer;

The decoding unit 210 is further configured to decode the code stream to obtain first information, where the first information is used to instruct to perform inverse quantization on the feature data of at least one channel in the N channels;

An inverse quantization unit 220, configured to perform inverse quantization on the feature data of the at least one channel according to the first information.

In some embodiments, the inverse quantization unit 220 is specifically configured to, according to the first information, inverse quantize the feature data of the fixed-point type of the at least one channel into the feature of the floating-point type of the at least one channel data.

In some embodiments, the inverse quantization unit 220 is specifically configured to perform inverse quantization on the fixed-point type feature data of the at least one channel by using a default inverse quantization manner according to the first information.

In some embodiments, the code stream further includes second information, where the second information is used to indicate an inverse quantization method used when performing inverse quantization on the characteristic data of the fixed-point type of the at least one channel, corresponding to The inverse quantization unit 220 is specifically configured to perform inverse quantization on the fixed-point type feature data of the at least one channel by using the inverse quantization manner indicated by the second information according to the first information.

In some embodiments, the inverse quantization unit 220 is specifically configured to use the same inverse quantization method to perform inverse quantization on the fixed-point type feature data of all channels in the N channels if the first information indicates to perform inverse quantization. Perform inverse quantization on the feature data of the fixed-point number type of all channels in the N channels; or, if the first information indicates to perform inverse quantization on the feature data of the fixed-point number type of each channel in the N channels, then For each channel, use the inverse quantization method corresponding to the channel to perform inverse quantization on the characteristic data of the fixed-point type of the channel; or, if the first information indicates that the characteristic data of the fixed-point type of the M groups of channels are respectively inversely quantized For quantization, the N channels are divided into M groups of channels, and for each group of channels, the inverse quantization method corresponding to the group of channels is used to inversely quantize the fixed-point type feature data of the group of channels.

In some embodiments, if the inverse quantization method for performing inverse quantization on the fixed-point type feature data of all channels in the N channels is a linear uniform inverse quantization method, the inverse quantization unit 220 is specifically configured to parse the first A piece of information to obtain the first target feature value, the first target scaling value and the first target quantization bit width; according to the first target feature value, the first target scaling value and the first target quantization bit width, use linear uniform inverse quantization manner, inverse quantization is performed on the feature data of the fixed-point number type of all channels in the N channels.

In some embodiments, if the inverse quantization method for performing inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels is a nonlinear logarithmic uniform inverse quantization method, the inverse quantization unit 220 is specifically configured to: According to the first information, determine the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base; according to the first target feature value, the first target scaling value and the first The target quantization bit width and the first logarithmic base are used to perform inverse quantization on the fixed-point feature data of all channels in the N channels by using the nonlinear logarithmic uniform inverse quantization method.

Exemplarily, the inverse quantization unit 220 is specifically configured to parse the first information to obtain the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base; or, parsing For the first information, the indication information of the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base is obtained; according to the indication information of the first logarithmic base, from Among the preset multiple logarithmic bases, the first logarithmic base is determined; or, the first information is analyzed to obtain the first target characteristic value, the first target scaling value and the first target quantization bit width, And the default logarithmic base is determined as the first logarithmic base.

In some embodiments, if the inverse quantization method for performing inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels is a nonlinear exponential uniform inverse quantization method, the inverse quantization unit 220 is specifically configured to perform inverse quantization according to the the first information, determine the first target feature value, the first target scaling value, the first target quantization bit width and the first exponent base; according to the first target feature value, the first target scaling value, the first target quantization bit The width and the first exponent base are used to dequantize the fixed-point type feature data of all channels in the N channels by using the nonlinear exponential uniform inverse quantization method.

Exemplarily, the inverse quantization unit 220 is specifically configured to parse the first information to obtain the first target feature value, the first target scaling value, the first target quantization bit width, and the first exponent base; The first information is obtained, and the indication information of the first target characteristic value, the first target scaling value, the first target quantization bit width and the first exponent base is obtained; according to the indication information of the first exponent base, from the preset Among multiple exponent bases, determine the first exponent base; or, parse the first information to obtain the first target feature value, the first target scaling value, and the first target quantization bit width, and use the default exponent The base is determined as the first exponent base.

In some embodiments, the first target feature value is one feature value in the feature data of all channels in the N channels, and the first target scaling value is the feature data of all channels in the N channels The corresponding scaling value during quantization, and the first target quantization bit width is the quantization bit width corresponding to the feature data of all channels in the N channels during quantization.

Optionally, the first target feature value is the smallest feature value in feature data of all channels in the N channels.

In some embodiments, if the inverse quantization method for performing inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels is a table look-up inverse quantization method, the inverse quantization unit 220 is specifically configured to determine the size of the quantization interval. The first correspondence between the index value and the inverse quantization value of the quantization interval, the first correspondence is determined based on the value before quantization and the value after quantization of the characteristic data of all channels in the N channels; The characteristic data of each fixed-point number type of all channels in the N channels, the value of the characteristic data of the fixed-point number type is used as the index of the quantization interval, and in the first correspondence, the feature of the fixed-point number type is queried. The target inverse quantization value corresponding to the value of the data; the target inverse quantization value is determined as the value of the floating point type of the characteristic data of the fixed point type.

In some embodiments, if the inverse quantization method corresponding to the channel is a linear uniform inverse quantization method, the inverse quantization unit 220 is specifically configured to parse the first information to obtain the second target feature value and the second target scaling value and the second target quantization bit width; according to the second target feature value, the second target scaling value and the second target quantization bit width, using the linear uniform inverse quantization method, the characteristic data of the fixed-point number type of this channel is carried out. Inverse quantization.

In some embodiments, if the inverse quantization method corresponding to the channel is a nonlinear logarithmic uniform inverse quantization method, the inverse quantization unit 220 is specifically configured to determine the second target eigenvalue, the second target eigenvalue, the second a target scaling value, a second target quantization bit width, and a second logarithmic base; using the nonlinear The logarithmic uniform inverse quantization method performs inverse quantization on the fixed-point type feature data of the channel.

Exemplarily, the inverse quantization unit 220 is specifically configured to parse the first information to obtain the second target feature value, the second target scaling value, the second target quantization bit width, and the second logarithmic base; or, parsing The first information obtains the indication information of the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second logarithmic base; according to the indication information of the second logarithmic base, from Among the preset multiple logarithmic bases, the second logarithmic base is determined; or, the first information is analyzed to obtain the second target characteristic value, the second target scaling value and the second target quantization bit width, And the default logarithmic base is determined as the second logarithmic base.

In some embodiments, if the inverse quantization method corresponding to the channel is a nonlinear exponential uniform inverse quantization method, the inverse quantization unit 220 is specifically configured to determine the second target eigenvalue and the second target scaling according to the first information value, the second target quantization bit width and the second exponent base; according to the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second exponent base, the nonlinear exponent is used for uniform inverse quantization method, inverse quantization is performed on the feature data of the fixed-point type of the channel.

Exemplarily, the inverse quantization unit 220 is specifically configured to parse the first information to obtain the second target feature value, the second target scaling value, the second target quantization bit width and the second exponent base; The first information is obtained, and the first information includes the indication information of the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base; according to the second logarithmic base The indication information of , determine the second exponent base from a plurality of preset exponent bases; or, parse the first information to obtain the second target characteristic value, the second target scaling value, the second target quantization bit width, and the default exponent base is determined as the second exponent base.

In some embodiments, the second target feature value is a feature value in the feature data of the group of channels, the second target scaling value is a scaling value corresponding to the feature data of the channel during quantization, and the first target scaling value is The second target quantization bit width is the quantization bit width corresponding to the characteristic data of the channel during quantization.

In some embodiments, if the inverse quantization method corresponding to the channel is a look-up table inverse quantization method, the inverse quantization unit 220 is specifically configured to determine the second correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval , the second correspondence is determined based on the value before quantization and the value after quantization of the characteristic data of the channel; for each characteristic data of the fixed-point number type in the channel, the value of the characteristic data of the fixed-point number type is determined. The value is used as the index of the quantization interval, and in the second correspondence, the target inverse quantization value corresponding to the value of the characteristic data of the fixed-point number type is queried; the target inverse quantization value is determined as the characteristic data of the fixed-point number type. value of type floating point.

In some embodiments, if the inverse quantization method corresponding to the group of channels is a linear uniform inverse quantization method, the inverse quantization unit 220 is specifically configured to parse the first information to obtain the third target eigenvalue, the third target scaling value and the third target quantization bit width; according to the third target feature value, the third target scaling value and the third target quantization bit width, using the linear uniform inverse quantization method, the characteristics of the fixed-point number type of the group of channels Data is dequantified.

In some embodiments, if the inverse quantization method corresponding to the group of channels is a non-linear logarithmic uniform inverse quantization method, the inverse quantization unit 220 is specifically configured to determine the third target eigenvalue, the first Three target scaling values, the third target quantization bit width and the third logarithmic base; according to the third target feature value, the third target scaling value, the third target quantization bitwidth and the third logarithmic base, use the non- The linear logarithmic uniform inverse quantization method performs inverse quantization on the fixed-point feature data of this group of channels.

Exemplarily, the inverse quantization unit 220 is specifically configured to parse the first information to obtain the third target feature value, the third target scaling value, the third target quantization bit width, and the third logarithmic base; or, parsing For the first information, the indication information of the third target feature value, the third target scaling value, the third target quantization bit width and the third logarithmic base is obtained; according to the indication information of the third logarithmic base, from Among the preset multiple logarithmic bases, the third logarithmic base is determined; or, the first information is analyzed to obtain the third target characteristic value, the third target scaling value and the third target quantization bit width, And the default logarithmic base is determined as the third logarithmic base.

In some embodiments, if the inverse quantization method corresponding to the set of channels is a nonlinear exponential uniform inverse quantization method, the inverse quantization unit 220 is specifically configured to determine the third target eigenvalue, the third target eigenvalue and the third target according to the first information. The scaling value, the third target quantization bit width and the third exponent base; according to the third target eigenvalue, the third target scaling value, the third target quantization bit width and the third exponent base, the non-linear exponent is used to uniformly invert. In the quantization mode, inverse quantization is performed on the fixed-point feature data of this group of channels.

Exemplarily, the inverse quantization unit 220 is specifically configured to parse the first information to obtain the third target feature value, the third target scaling value, the third target quantization bit width, and the third exponent base; The first information is obtained, and the indication information that the first information includes the third target feature value, the third target scaling value, the third target quantization bit width and the third logarithmic base is obtained; according to the third logarithmic base the indication information, determine the third exponent base from a plurality of preset exponent bases; or, parse the first information to obtain the third target characteristic value, the third target scaling value, and the third target quantization bit width, and the default exponent base is determined as the third exponent base.

In some embodiments, the third target feature value is a feature value in the feature data of the set of channels, the third target scaling value is a scaling value corresponding to the feature data of the set of channels during quantization, and the The third target quantization bit width is the quantization bit width corresponding to the characteristic data of the group of channels during quantization.

In some embodiments, if the inverse quantization method corresponding to the set of channels is the table look-up inverse quantization method, the inverse quantization unit 220 is specifically configured to determine the third correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval relationship, the third corresponding relationship is determined based on the value before quantization and the value after quantization of the characteristic data of this group of channels;

For each characteristic data of the fixed-point type in the set of channels, the value of the characteristic data of the fixed-point type is used as the index of the quantization interval, and in the third corresponding relationship, the value of the characteristic data of the fixed-point type is queried. The corresponding target inverse quantization value;

The target inverse quantization value is determined as a value of the floating point type of the feature data of the fixed point type.

Optionally, the corresponding relationship between the index value of the quantization interval and the inverse quantization value of the quantization interval is default; or, the first information includes the index value of the quantization interval and the inverse quantization value of the quantization interval. Correspondence between.

It should be understood that the apparatus embodiments and the method embodiments may correspond to each other, and similar descriptions may refer to the method embodiments. To avoid repetition, details are not repeated here. Specifically, the video decoder 20 shown in FIG. 12 may correspond to the corresponding subject in performing the methods 700 to 1000 of the embodiments of the present application, and the aforementioned and other operations and/or functions of the respective units in the video decoder 20 are for the purpose of For the sake of brevity, the corresponding processes in each of the implementation methods 700 to 1000 will not be repeated here.

The apparatus and system of the embodiments of the present application are described above from the perspective of functional units with reference to the accompanying drawings. It should be understood that the functional unit may be implemented in the form of hardware, may also be implemented by an instruction in the form of software, or may be implemented by a combination of hardware and software units. Specifically, the steps of the method embodiments in the embodiments of the present application may be completed by hardware integrated logic circuits in the processor and/or instructions in the form of software, and the steps of the methods disclosed in conjunction with the embodiments of the present application may be directly embodied as hardware The execution of the decoding processor is completed, or the execution is completed by a combination of hardware and software units in the decoding processor. Optionally, the software unit may be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, and other storage media mature in the art. The storage medium is located in the memory, and the processor reads the information in the memory, and completes the steps in the above method embodiments in combination with its hardware.

FIG. 13 is a schematic block diagram of an electronic device 30 provided by an embodiment of the present application.

As shown in FIG. 13 , the electronic device 30 may be the video encoder or the video decoder described in this embodiment of the application, and the electronic device 30 may include:

A memory 33 and a processor 32 for storing a computer program 34 and transmitting the program code 34 to the processor 32 . In other words, the processor 32 can call and run the computer program 34 from the memory 33 to implement the methods in the embodiments of the present application.

For example, the processor 32 may be adapted to perform the steps of the above-described methods according to instructions in the computer program 34 .

In some embodiments of the present application, the processor 32 may include, but is not limited to:

General-purpose processor, Digital Signal Processor (DSP), Application Specific Integrated Circuit (ASIC), Field Programmable Gate Array (FPGA) or other programmable logic devices, discrete gates Or transistor logic devices, discrete hardware components, and so on.

In some embodiments of the present application, the memory 33 includes but is not limited to:

Volatile memory and/or non-volatile memory. Wherein, the non-volatile memory may be a read-only memory (Read-Only Memory, ROM), a programmable read-only memory (Programmable ROM, PROM), an erasable programmable read-only memory (Erasable PROM, EPROM), an electrically programmable read-only memory (Erasable PROM, EPROM). Erase programmable read-only memory (Electrically EPROM, EEPROM) or flash memory. Volatile memory may be Random Access Memory (RAM), which acts as an external cache. By way of illustration and not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (Double Data Rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (Enhanced SDRAM, ESDRAM), synchronous link dynamic random access memory (synch link DRAM, SLDRAM) and direct memory bus random access memory (Direct Rambus RAM, DR RAM).

In some embodiments of the present application, the computer program 34 may be divided into one or more units, and the one or more units are stored in the memory 33 and executed by the processor 32 to complete the procedures provided by the present application. Methods. The one or more units may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program 34 in the electronic device 30 .

As shown in FIG. 13 , the electronic device 30 may further include:

A transceiver 33 which can be connected to the processor 32 or the memory 33 .

The processor 32 can control the transceiver 33 to communicate with other devices, specifically, can send information or data to other devices, or receive information or data sent by other devices. The transceiver 33 may include a transmitter and a receiver. The transceiver 33 may further include antennas, and the number of the antennas may be one or more.

It should be understood that each component in the electronic device 30 is connected through a bus system, wherein the bus system includes a power bus, a control bus and a status signal bus in addition to a data bus.

As shown in FIG. 14 , the video encoding and decoding system 40 may include: a video encoder 41 and a video decoder 42 , wherein the video encoder 41 is used for executing the video encoding method involved in the embodiments of the present application, and the video decoder 42 is used for executing The video decoding method involved in the embodiments of the present application.

The present application also provides a computer storage medium on which a computer program is stored, and when the computer program is executed by a computer, enables the computer to execute the methods of the above method embodiments. In other words, the embodiments of the present application further provide a computer program product including instructions, when the instructions are executed by a computer, the instructions cause the computer to execute the methods of the above method embodiments.

When implemented in software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. The computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable device. The computer instructions may be stored on or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions may be transmitted over a wire from a website site, computer, server or data center (eg coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (eg infrared, wireless, microwave, etc.) means to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that includes one or more available media integrated. The available media may be magnetic media (eg, floppy disk, hard disk, magnetic tape), optical media (eg, digital video disc (DVD)), or semiconductor media (eg, solid state disk (SSD)), and the like.

Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of this application.

In the several embodiments provided in this application, it should be understood that the disclosed system, apparatus and method may be implemented in other manners. For example, the apparatus embodiments described above are only illustrative. For example, the division of the unit is only a logical function division. In actual implementation, there may be other division methods, for example, multiple units or components may be combined or Integration into another system, or some features can be ignored, or not implemented. On the other hand, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be in electrical, mechanical or other forms.

Units described as separate components may or may not be physically separated, and components shown as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment. For example, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit.

The above are only specific embodiments of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art who is familiar with the technical scope disclosed in the present application can easily think of changes or substitutions. Covered within the scope of protection of this application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

An image coding method, comprising:

Get the current image to be encoded;

Inputting the current image into a neural network to obtain feature data of the current image, where the feature data of the current image includes feature data of N channels, and N is a positive integer;

quantifying the feature data of at least one channel in the N channels;

Encoding the quantized feature data of the at least one channel to obtain a code stream, where the code stream includes first information, and the first information is used to indicate the characteristics of at least one channel in the N channels Data is dequantified.
The method according to claim 1, wherein the quantizing the feature data of at least one channel in the N channels comprises:

Quantizing the feature data of the floating point type of at least one channel of the N channels to obtain the feature data of the fixed point type of the at least one channel.
The method according to claim 2, wherein the quantization method for quantizing the floating point type characteristic data of at least one channel of the N channels includes any one of the following: a linear uniform quantization method, a nonlinear exponential uniform method Quantization method, nonlinear logarithmic uniform quantization method, look-up table quantization method.
The method according to claim 2 or 3, wherein the quantizing the floating point type feature data of at least one channel of the N channels comprises:

Use the same quantization method to quantize the floating-point type feature data of all channels in the N channels; or,

Quantize the characteristic data of the floating point type of each channel in the N channels by using different quantization methods respectively;

Alternatively, the N channels are grouped, and a quantization method is used to quantize the floating point type feature data of each group of channels.
The method according to claim 4, wherein, if the quantization method is a linear uniform quantization method, the feature data of the floating point type of all channels in the N channels is quantized using the same quantization method ,include:

Obtain a preset first quantization bit width, and the first eigenvalue and the second eigenvalue in the feature data of the floating point type of all channels in the N channels;

According to the first eigenvalue and the second eigenvalue, and the first quantization bit width, the linear uniform quantization method is used to quantize the floating point type feature data of each of the N channels.
The method according to claim 4, wherein if the quantization method is a non-linear logarithmic uniform quantization method, the same quantization method is used for the floating-point type feature data of all channels in the N channels. ways to quantify, including:

Obtain the preset second quantization bit width and the first base of the logarithmic function, and the first eigenvalue and the second eigenvalue in the feature data of the floating point type of all channels in the N channels;

According to the first eigenvalue and the second eigenvalue, as well as the second quantization bit width and the first base of the logarithmic function, using the nonlinear logarithmic uniform quantization method, for the N channels The feature data of float type for each channel is quantized.
The method according to claim 4, wherein if the quantization method is a nonlinear exponential uniform quantization method, the same quantization method is used for the floating-point type feature data of all channels in the N channels. Quantify, including:

Obtain the preset third quantization bit width and the first base of the exponential function, and the first eigenvalue and the second eigenvalue in the feature data of the floating point type of all channels in the N channels;

According to the first eigenvalue and the second eigenvalue, as well as the third quantization bit width and the first base of the exponential function, using the nonlinear exponential uniform quantization method, for each of the N channels The feature data of the channel's floating point type is quantized.
The method according to claim 4, wherein if the quantization method is a look-up table quantization method, the quantization is performed using the same quantization method for the floating-point type feature data of all channels in the N channels. ,include:

Sort the feature data of the floating point type of all channels in the N channels according to the value size to obtain the sorted first feature data;

dividing the sorted first characteristic data into a plurality of first quantization intervals, wherein each first quantization interval includes the same amount of characteristic data;

For each of the first quantization intervals, the value of the feature data in the first quantization interval is quantized as an index value of the first quantization interval.
The method according to any one of claims 5-7, wherein the first eigenvalue is the smallest eigenvalue in the floating-point type feature data of all channels in the N channels, and the second eigenvalue is The eigenvalue is the largest eigenvalue in the feature data of the floating point type of all channels in the N channels.
The method according to claim 4, wherein if the quantization method is a linear uniform quantization method, a quantization method is respectively used for the characteristic data of the floating point type of each channel in the N channels. Quantify, including:

For each channel in the N channels, obtain a preset fourth quantization bit width, and the third eigenvalue and the fourth eigenvalue in the feature data of the floating point type of the channel;

According to the third eigenvalue, the fourth eigenvalue, and the fourth quantization bit width, using the linear uniform quantization method, the feature data of the floating point type of the channel is quantized.
The method according to claim 4, wherein if the quantization method is a non-linear logarithmic uniform quantization method, the floating point type feature data of each channel in the N channels is respectively used as a quantified in various ways, including:

For each of the N channels, obtain a preset fifth quantization bit width and the second base of the logarithmic function, and the third eigenvalue and the fourth feature in the feature data of the floating point type of the channel value;

According to the third eigenvalue and the fourth eigenvalue, the fifth quantization bit width and the second base of the logarithmic function, using the nonlinear logarithmic uniform quantization method, the floating point type of the channel quantified feature data.
The method according to claim 4, wherein if the quantization method is a nonlinear exponential uniform quantization method, the floating-point type feature data of each channel in the N channels is divided into two types. Quantitative methods are quantified, including:

For each of the N channels, obtain a preset sixth quantization bit width and the second base of the exponential function, and the third and fourth eigenvalues in the floating-point type feature data of the channel ;

According to the third eigenvalue and the fourth eigenvalue, as well as the sixth quantization bit width and the second base of the exponential function, using the nonlinear exponential uniform quantization method, the floating point type feature of the channel is data is quantified.
The method according to claim 5, wherein if the quantization method is a look-up table quantization method, a quantization method is respectively used for the characteristic data of the floating point type of each channel in the N channels. Quantify, including:

For each channel in the N channels, sort the feature data of the floating point type of the channel according to the value size, and obtain the sorted second feature data under the channel;

dividing the sorted second characteristic data under the channel into a plurality of second quantization intervals, wherein each second quantization interval includes the same amount of characteristic data;

For each second quantization interval, the value of the feature data in the second quantization interval is quantized as an index value of the second quantization interval.
The method according to any one of claims 10-12, wherein the third eigenvalue is the largest eigenvalue in the feature data of the floating point type of the channel, and the fourth eigenvalue is the channel The smallest eigenvalue in eigendata of float type.
The method according to claim 4, wherein if the quantization method is a linear uniform quantization method, the feature data of the floating point type of each group of channels is quantized using a quantization method, comprising:

For each group of channels, obtain the preset seventh quantization bit width, and the fifth and sixth eigenvalues in the feature data of the floating point type of the group of channels;

According to the fifth eigenvalue and the sixth eigenvalue, and the seventh quantization bit width, using the linear uniform quantization method, the floating point type feature data of each channel in the group of channels is quantized.
The method according to claim 4, wherein if the quantization method is a non-linear logarithmic uniform quantization method, the characteristic data of the floating point type of each group of channels is quantized using a quantization method respectively. ,include:

For each group of channels, obtain the preset eighth quantization bit width and the third base of the logarithmic function, and the fifth and sixth eigenvalues in the feature data of the floating point type of the group of channels;

According to the fifth eigenvalue and the sixth eigenvalue, the eighth quantization bit width and the third base of the logarithmic function, using the non-linear logarithmic uniform quantization method, for each channel in the group of channels Quantize the feature data of the floating point type of each channel.
The method according to claim 4, wherein, if the quantization method is a nonlinear exponential uniform quantization method, the characteristic data of the floating point type of each group of channels is quantized using a quantization method respectively, include:

For each group of channels, obtain the preset ninth quantization bit width and the third base of the exponential function, and the fifth and sixth eigenvalues in the feature data of the floating point type of the group of channels;

According to the fifth eigenvalue and the sixth eigenvalue, and the ninth quantization bit width and the third base of the exponential function, using the nonlinear logarithmic uniform quantization method, each of the channels in the group of The feature data of the channel's floating point type is quantized.
The method according to claim 4, wherein, if the quantization method is a look-up table quantization method, the feature data of the floating point type of each group of channels is quantized by using a quantization method, comprising:

For each group of channels, sort the characteristic data of the floating point type of the group of channels according to the value size, and obtain the sorted third characteristic data under the group of channels;

dividing the sorted third characteristic data under the group of channels into a plurality of third quantization intervals, wherein each third quantization interval includes the same amount of characteristic data;

For each third quantization interval, the value of the feature data in the third quantization interval is quantized as an index value of the third quantization interval.
The method according to any one of claims 15-17, wherein the fifth eigenvalue is the largest eigenvalue in the floating-point type feature data of the group of channels, and the sixth eigenvalue is the group The smallest eigenvalue in the eigendata of the channel's float type.
The method according to any one of claims 2-19, wherein the first information indicates that inverse quantization is performed on fixed-point type feature data of all channels in the N channels; or,

The first information indicates that inverse quantization is performed on the fixed-point type feature data of each of the N channels; or,

The first information indicates that inverse quantization is performed on the fixed-point type feature data of each channel in the M groups of channels, wherein the M groups of channels are obtained by grouping the N channels, and each group of channels is obtained by grouping the N channels. At least one of the N channels is included.
The method according to any one of claims 2-20, wherein the inverse quantization method used when performing inverse quantization on the characteristic data of the fixed-point type of the at least one channel comprises any one of the following: linear uniform inverse Quantization method, nonlinear exponential uniform inverse quantization method, nonlinear logarithmic uniform inverse quantization method, and look-up table inverse quantization method.
The method according to any one of claims 2-21, wherein the first information includes at least one parameter required for inverse quantization of the characteristic data of the fixed-point type of the at least one channel.
The method according to claim 22, wherein the first information indicates that inverse quantization is performed on the fixed-point type feature data of all channels in the N channels, and the first information includes any one of the following :

If the inverse quantization method for performing inverse quantization on the characteristic data of the fixed-point type of all channels in the N channels is a linear uniform inverse quantization method, the first information includes a first target feature value, a first target scaling value and the first target quantization bit width;

If the inverse quantization method for inverse quantization of the fixed-point type feature data of all channels in the N channels is a nonlinear logarithmic uniform inverse quantization method, the first information includes a first target eigenvalue, a first A target scaling value and a first target quantization bit width, or the first information includes a first target feature value, a first target scaling value, a first target quantization bit width, and a first logarithmic base, or the first information includes The indication information of the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base;

If the inverse quantization method for inverse quantization of the fixed-point type characteristic data of all channels in the N channels is a nonlinear exponential uniform inverse quantization method, the first information includes a first target eigenvalue, a first target A scaling value and a first target quantization bit width, or the first information includes a first target feature value, a first target scaling value, a first target quantization bit width, and a first exponent base, or the first information includes a first The indication information of the target feature value, the first target scaling value, the first target quantization bit width and the first exponent base;

If the inverse quantization method for performing inverse quantization on the fixed-point type characteristic data of all channels in the N channels is a table look-up inverse quantization method, the first information includes the index value of the quantization interval and the inverse quantization of the quantization interval the first correspondence between the values, the first correspondence is determined based on the pre-quantized value and the quantized value of the characteristic data of all channels in the N channels;

The first target feature value is one feature value in the feature data of all channels in the N channels, and the first target scaling value corresponds to the feature data of all channels in the N channels during quantization The first target quantization bit width is the quantization bit width corresponding to the characteristic data of all channels in the N channels during quantization.
The method according to claim 23, wherein the first target feature value is the minimum value of feature data of all channels in the N channels.
The method according to claim 22, wherein the first information indicates that inverse quantization is performed on the fixed-point type feature data of each channel in the N channels, and for each channel, the first A message includes any of the following:

If the inverse quantization method for inverse quantization of the fixed-point type feature data of the channel is a linear uniform inverse quantization method, the first information includes the second target feature value, the second target scaling value, and the second target quantization bit width ;

If the feature data of the fixed-point type of the channel is inversely quantized into a non-linear logarithmic uniform inverse quantization method in an inverse quantization manner, the first information includes a second target feature value, a second target scaling value, and a second target The quantization bit width, or the first information includes the second target eigenvalue, the second target scaling value, the second target quantization bit width, and the second logarithmic base, or the first information includes the second target eigenvalue, the first Two target scaling values, the second target quantization bit width and the indication information of the second logarithmic base;

If the inverse quantization method for inverse quantization of the fixed-point type feature data of the channel is a nonlinear exponential uniform inverse quantization method, the first information includes the second target feature value, the second target scaling value, and the second target quantization bit width, or the first information includes a second target eigenvalue, a second target scaling value, a second target quantization bit width, and a second exponent base, or the first information includes a second target eigenvalue, a second target indication information of the scaling value, the second target quantization bit width and the second exponent base;

If the inverse quantization method for inverse quantization of the fixed-point type characteristic data of the channel is the table look-up inverse quantization method, the first information includes the second correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval relationship, the second corresponding relationship is determined based on the value before quantization and the value after quantization of the characteristic data of the channel;

Wherein, the second target feature value is a feature value in the feature data of the channel, the second target scaling value is the scaling value corresponding to the feature data of the channel during quantization, and the second target quantization bit width It is the corresponding quantization bit width of the feature data of this channel during quantization.
The method according to claim 25, wherein the second target characteristic value is the minimum value of characteristic data of the channel.
The method according to claim 22, wherein the first information indicates that inverse quantization is performed on the fixed-point type feature data of M groups of channels respectively, and for each group of channels, the first information includes any of the following A sort of:

If the inverse quantization is performed on the group of channels so that the inverse quantization method is a linear uniform inverse quantization method, the first information includes the third target eigenvalue, the third target scaling value and the third target quantization bit width;

If the inverse quantization method for inverse quantization of the fixed-point type feature data of the group of channels is a nonlinear logarithmic uniform inverse quantization method, the first information includes the third target eigenvalue, the third target scaling value, and the third target scaling value. The target quantization bit width, or the first information includes the third target eigenvalue, the third target scaling value, the third target quantization bit width, and the third logarithmic base, or the first information includes the third target eigenvalue, Indication information of the third target scaling value, the third target quantization bit width and the third logarithmic base;

If the inverse quantization method for inverse quantization of the fixed-point type feature data of the group of channels is a nonlinear exponential uniform inverse quantization method, the first information includes the third target feature value, the third target scaling value, and the third target Quantization bit width, or the first information includes a third target eigenvalue, a third target scaling value, a third target quantization bit width, and a third exponent base, or the first information includes a third target eigenvalue, a third Indication information of the target scaling value, the third target quantization bit width and the third exponent base;

If the inverse quantization method for performing inverse quantization on the fixed-point type feature data of the group of channels is the table look-up inverse quantization method, the first information includes the third index value between the index value of the quantization interval and the inverse quantization value of the quantization interval. Correspondence, the third correspondence is determined based on the value before quantization and the value after quantization of the characteristic data of this group of channels;

The M groups of channels are obtained by grouping the N channels, each group of channels includes at least one channel in the N channels, and the third target feature value is in the feature data of the group of channels. A characteristic value of , the third target scaling value is the corresponding scaling value of the characteristic data of this group of channels during quantization, and the third target quantization bit width is the corresponding quantization bit width of the characteristic data of this group of channels during quantization .
The method according to claim 27, wherein the third target feature value is the minimum value of feature data of the group of channels.
The method according to any one of claims 2-28, wherein the code stream further includes second information, and the second information is used to indicate that the characteristic data of the fixed-point type of the at least one channel is to be performed The inverse quantization method used during inverse quantization.
An image decoding method, comprising:

Decoding the code stream to obtain fixed-point feature data of the current image, where the feature data of the current image includes feature data of N channels, and N is a positive integer;

Decoding the code stream to obtain first information, where the first information is used to instruct to perform inverse quantization on the feature data of at least one channel in the N channels;

According to the first information, inverse quantization is performed on the feature data of the at least one channel.
The method according to claim 30, wherein the performing inverse quantization on the characteristic data of the at least one channel according to the first information comprises:

The inverse quantization is performed on the feature data of the fixed-point type of the at least one channel according to the first information, to obtain the feature data of the floating-point number type of the at least one channel.
The method according to claim 31, wherein the inverse quantization method used when performing inverse quantization on the characteristic data of the fixed-point type of the at least one channel comprises any one of the following: a linear uniform inverse quantization method, a nonlinear Exponential uniform inverse quantization method, nonlinear logarithmic uniform inverse quantization method, and look-up table inverse quantization method.
The method according to claim 31 or 32, wherein the performing inverse quantization on the fixed-point feature data of the at least one channel according to the first information comprises:

According to the first information, inverse quantization is performed on the fixed-point type feature data of the at least one channel by using a default inverse quantization manner.
The method according to claim 31 or 32, wherein the code stream further comprises second information, and the second information is used to indicate when performing inverse quantization on the characteristic data of the fixed-point type of the at least one channel The inverse quantization method used, in which the inverse quantization is performed on the characteristic data of the fixed-point type of the at least one channel according to the first information, including:

According to the first information, inverse quantization is performed on the fixed-point type feature data of the at least one channel by using the inverse quantization manner indicated by the second information.
The method according to claim 31 or 34, wherein the first information includes at least one parameter required for inverse quantization of the fixed-point type feature data of the at least one channel.
The method according to any one of claims 31-35, wherein the performing inverse quantization on the fixed-point feature data of the at least one channel according to the first information, comprises:

If the first information indicates to perform inverse quantization on the fixed-point feature data of all channels in the N channels, use the same inverse quantization method to perform inverse quantization on the fixed-point feature data of all channels in the N channels perform inverse quantification; or,

If the first information indicates that inverse quantization is performed on the feature data of the fixed-point type of each channel in the N channels, for each channel, use the inverse quantization method corresponding to the channel to perform the inverse quantization of the fixed-point type of the channel. inverse quantization of the feature data; or,

If the first information indicates that inverse quantization is performed on the fixed-point type feature data of M groups of channels, the N channels are divided into M groups of channels, and for each group of channels, the inverse quantization corresponding to the group of channels is used. method, inverse quantization is performed on the fixed-point type feature data of the group of channels.
The method according to claim 36, wherein if the inverse quantization method for inverse quantization of the fixed-point type feature data of all channels in the N channels is a linear uniform inverse quantization method, the same The inverse quantization method performs inverse quantization on the fixed-point type feature data of all channels in the N channels, including:

Parsing the first information to obtain the first target feature value, the first target scaling value and the first target quantization bit width;

According to the first target feature value, the first target scaling value and the first target quantization bit width, a linear uniform inverse quantization method is used to perform inverse quantization on the fixed-point type feature data of all channels in the N channels.
The method according to claim 36, wherein if the inverse quantization method for inverse quantization of the fixed-point type feature data of all channels in the N channels is a nonlinear logarithmic uniform inverse quantization method, then the The description uses the same inverse quantization method to perform inverse quantization on the fixed-point type feature data of all channels in the N channels, including:

According to the first information, determine the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base;

According to the first target eigenvalue, the first target scaling value, the first target quantization bit width and the first logarithmic base, the nonlinear logarithmic uniform inverse quantization method is used to determine the quantization of all channels in the N channels. The feature data of point type is inversely quantized.
The method according to claim 38, wherein determining the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base according to the first information comprises:

Parse the first information to obtain the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base; or,

Parse the first information to obtain the indication information of the first target feature value, the first target scaling value, the first target quantization bit width and the first logarithmic base; according to the indication information of the first logarithmic base, From a plurality of preset logarithmic bases, determine the first logarithmic base; or,

The first information is parsed to obtain the first target feature value, the first target scaling value and the first target quantization bit width, and the default logarithmic base is determined as the first logarithmic base.
The method according to claim 36, wherein if the inverse quantization method for inverse quantization of the fixed-point type characteristic data of all channels in the N channels is a nonlinear exponential uniform inverse quantization method, the using The same inverse quantization method performs inverse quantization on the fixed-point type feature data of all channels in the N channels, including:

According to the first information, determine the first target feature value, the first target scaling value, the first target quantization bit width and the first exponent base;

According to the first target eigenvalue, the first target scaling value, the first target quantization bit width and the first exponential base, using the nonlinear exponential uniform inverse quantization method, the fixed-point number of all channels in the N channels Type of feature data for inverse quantization.
The method according to claim 40, wherein determining the first target characteristic value, the first target scaling value, the first target quantization bit width and the first exponent base according to the first information comprises:

Parse the first information to obtain the first target feature value, the first target scaling value, the first target quantization bit width and the first exponent base; or,

Parse the first information to obtain the indication information of the first target feature value, the first target scaling value, the first target quantization bit width and the first exponent base; Among the multiple set exponent bases, determine the first exponent base; or,

The first information is parsed to obtain the first target characteristic value, the first target scaling value, and the first target quantization bit width, and the default exponent base is determined as the first exponent base.
The method according to any one of claims 37-41, wherein the first target feature value is one feature value in the feature data of all channels in the N channels, and the first target scaling value is is the scaling value corresponding to the feature data of all channels in the N channels during quantization, and the first target quantization bit width is the quantization bit width corresponding to the feature data of all channels in the N channels during quantization.
The method according to claim 42, wherein the first target feature value is the smallest feature value in feature data of all channels in the N channels.
The method according to claim 36, wherein, if the inverse quantization method for performing inverse quantization on the fixed-point type characteristic data of all channels in the N channels is a look-up table inverse quantization method, the same The inverse quantization method performs inverse quantization on the fixed-point type feature data of all channels in the N channels, including:

Determine the first correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval, and the first correspondence is based on the value before quantization and the value after quantization of the characteristic data of all channels in the N channels definite;

For the characteristic data of each fixed-point type of all channels in the N channels, the value of the characteristic data of the fixed-point type is used as the index of the quantization interval, and in the first correspondence, query the fixed-point type of the characteristic data. The target inverse quantization value corresponding to the value of the feature data;

The target inverse quantization value is determined as a value of the floating point type of the feature data of the fixed point type.
The method according to claim 36, wherein if the inverse quantization method corresponding to the channel is a linear uniform inverse quantization method, the use of the inverse quantization method corresponding to the channel is used for the fixed-point type feature data of the channel. Perform inverse quantization, including:

Parsing the first information to obtain a second target feature value, a second target scaling value and a second target quantization bit width;

According to the second target feature value, the second target scaling value and the second target quantization bit width, the linear uniform inverse quantization method is used to perform inverse quantization on the fixed-point type feature data of the channel.
The method according to claim 36, wherein if the inverse quantization method corresponding to the channel is a non-linear logarithmic uniform inverse quantization method, the inverse quantization method corresponding to the channel is used, and the fixed-point number of the channel is Types of feature data for inverse quantization, including:

According to the first information, determine the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base;

According to the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base, using the nonlinear logarithmic uniform inverse quantization method, the fixed-point type feature data of the channel Do inverse quantization.
The method according to claim 46, wherein determining the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base according to the first information comprises:

Parse the first information to obtain the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base; or,

Parse the first information to obtain the indication information of the second target feature value, the second target scaling value, the second target quantization bit width and the second logarithmic base; according to the indication information of the second logarithmic base, From a plurality of preset logarithmic bases, determine the second logarithmic base; or,

The first information is parsed to obtain the second target feature value, the second target scaling value and the second target quantization bit width, and the default logarithmic base is determined as the second logarithmic base.
The method according to claim 36, wherein if the inverse quantization method corresponding to the channel is a nonlinear exponential uniform inverse quantization method, the inverse quantization method corresponding to the channel is used, and the fixed-point number type of the channel is used. Inverse quantization of feature data, including:

According to the first information, determine the second target feature value, the second target scaling value, the second target quantization bit width and the second exponent base;

According to the second target eigenvalue, the second target scaling value, the second target quantization bit width and the second exponential base, use the nonlinear exponential uniform inverse quantization method to inverse the feature data of the fixed-point type of the channel quantify.
The method according to claim 48, wherein determining the second target characteristic value, the second target scaling value, the second target quantization bit width and the second exponent base according to the first information comprises:

Parse the first information to obtain the second target feature value, the second target scaling value, the second target quantization bit width and the second exponent base; or,

Parsing the first information to obtain indication information that the first information includes the second target feature value, the second target scaling value, the second target quantization bit width, and the second logarithmic base; The indication information of the number base, the second exponent base is determined from the preset multiple exponent bases; or,

The first information is parsed to obtain the second target characteristic value, the second target scaling value, and the second target quantization bit width, and the default exponent base is determined as the second exponent base.
The method according to any one of claims 45-49, wherein the second target feature value is a feature value in the feature data of the group of channels, and the second target scaling value is a feature of the channel The scaling value corresponding to the data during quantization, and the second target quantization bit width is the quantization bit width corresponding to the characteristic data of the channel during quantization.
The method according to claim 50, wherein the second target feature value is the smallest feature value in the feature data of the channel.
The method according to claim 36, wherein if the inverse quantization method corresponding to the channel is a look-up table inverse quantization method, the inverse quantization method corresponding to the channel is used to obtain the fixed-point type characteristic data of the channel. Perform inverse quantization, including:

Determine the second correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval, and the second correspondence is determined based on the value before quantization and the value after quantization of the characteristic data of the channel;

For each characteristic data of the fixed-point type in the channel, the value of the characteristic data of the fixed-point type is used as the index of the quantization interval, and in the second correspondence, query the value corresponding to the characteristic data of the fixed-point type. The target inverse quantization value of ;

The target inverse quantization value is determined as a value of the floating point type of the feature data of the fixed point type.
The method according to claim 36, wherein if the inverse quantization method corresponding to the group of channels is a linear uniform inverse quantization method, the use of the inverse quantization method corresponding to the group of channels is a fixed-point type of the group of channels. Inverse quantization of the feature data, including:

Parsing the first information to obtain a third target feature value, a third target scaling value and a third target quantization bit width;

According to the third target feature value, the third target scaling value and the third target quantization bit width, the linear uniform inverse quantization method is used to perform inverse quantization on the fixed-point type feature data of the group of channels.
The method according to claim 36, wherein if the inverse quantization method corresponding to the group of channels is a nonlinear logarithmic uniform inverse quantization method, the inverse quantization method corresponding to the group of channels is used to perform the inverse quantization method corresponding to the group of channels The feature data of fixed-point type is inverse quantized, including:

According to the first information, determine the third target feature value, the third target scaling value, the third target quantization bit width and the third logarithmic base;

According to the third target feature value, the third target scaling value, the third target quantization bit width, and the third logarithmic base, using the nonlinear logarithmic uniform inverse quantization method, the fixed-point type feature of the group of channels Data is dequantified.
The method according to claim 54, wherein determining the third target characteristic value, the third target scaling value, the third target quantization bit width and the third logarithmic base according to the first information comprises:

Parse the first information to obtain the third target feature value, the third target scaling value, the third target quantization bit width and the third logarithmic base; or,

Parse the first information to obtain the indication information of the third target feature value, the third target scaling value, the third target quantization bit width and the third logarithmic base; according to the indication information of the third logarithmic base, From a plurality of preset logarithmic bases, determine the third logarithmic base; or,

The first information is parsed to obtain the third target feature value, the third target scaling value and the third target quantization bit width, and the default logarithmic base is determined as the third logarithmic base.
The method according to claim 36, wherein if the inverse quantization method corresponding to the group of channels is a nonlinear exponential uniform inverse quantization method, the inverse quantization method corresponding to the group of channels is used to determine the set of channels. The feature data of point type is inversely quantized, including:

According to the first information, determine the third target feature value, the third target scaling value, the third target quantization bit width and the third exponent base;

According to the third target eigenvalue, the third target scaling value, the third target quantization bit width and the third exponential base, using the nonlinear exponential uniform inverse quantization method, the fixed-point type feature data of the group of channels is Inverse quantization.
The method according to claim 56, wherein determining the third target characteristic value, the third target scaling value, the third target quantization bit width and the third exponent base according to the first information comprises:

Parse the first information to obtain the third target feature value, the third target scaling value, the third target quantization bit width and the third exponent base; or,

Parse the first information to obtain indication information that the first information includes the third target feature value, the third target scaling value, the third target quantization bit width and the third logarithmic base; The indication information of the number base, and the third exponent base is determined from the preset multiple exponent bases; or,

The first information is parsed to obtain the third target feature value, the third target scaling value, and the third target quantization bit width, and the default exponent base is determined as the third exponent base.
The method according to any one of claims 53-57, wherein the third target feature value is a feature value in the feature data of the group of channels, and the third target scaling value is a feature value of the group of channels. The scaling value corresponding to the feature data during quantization, and the third target quantization bit width is the quantization bit width corresponding to the feature data of the group of channels during quantization.
The method according to claim 58, wherein the third target eigenvalue is the smallest eigenvalue in the eigendata of the group of channels.
The method according to claim 36, wherein if the inverse quantization method corresponding to the group of channels is a look-up table inverse quantization method, the inverse quantization method corresponding to the group of channels is used, and the fixed-point number type of the group of channels is used. Inverse quantization of the feature data, including:

Determine the third correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval, and the third correspondence is determined based on the value before quantization and the value after quantization of the characteristic data of the group of channels;

For each fixed-point type characteristic data in the set of channels, the value of the fixed-point type characteristic data is used as the index of the quantization interval, and in the third corresponding relationship, the value of the fixed-point type characteristic data is queried. The corresponding target inverse quantization value;

The target inverse quantization value is determined as a value of the floating point type of the feature data of the fixed point type.
The method according to claim 44, 52 or 60, wherein the target correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval is a default; or the first information includes the The target correspondence between the index value of the quantization interval and the inverse quantization value of the quantization interval, and the target correspondence includes a first pair, a second correspondence, or a third correspondence.
An image encoder, comprising:

an acquisition unit for acquiring the current image to be encoded;

A feature extraction unit, configured to input the current image into a neural network to obtain feature data of the current image, where the feature data of the current image includes feature data of N channels, and N is a positive integer;

a quantization unit, configured to quantify the feature data of at least one channel in the N channels;

The encoding unit is configured to encode the quantized feature data of the at least one channel to obtain a code stream, where the code stream includes first information, and the first information is used to indicate the The feature data of at least one channel is inverse quantized.
A video decoder, comprising:

a decoding unit, configured to decode the code stream to obtain feature data of the current image, where the feature data of the current image includes the feature data of N channels, where N is a positive integer; and, decode the code stream to obtain the first information, where the The first information is used to indicate that the feature data of at least one channel in the N channels is inversely quantized;

an inverse quantization unit, configured to perform inverse quantization on the feature data of the at least one channel according to the first information.
A video encoding and decoding system, comprising:

The video encoder of claim 62;

and the video decoder of claim 63.
A computer-readable storage medium, wherein computer-executable instructions are stored in the computer-readable storage medium, and when the computer-executable instructions are executed by a processor, are used to implement any one of claims 1 to 61. Methods.