KR102421719B1 - Apparatus and method for performing artificial intelligence encoding and artificial intelligence decoding of image by using low-complexity neural network - Google Patents

Abstract

The processor executing one or more instructions stored in the AI encoding device inputs a first reduced image downscaled from the original image and a reduced feature map having a resolution smaller than the resolution of the original image into the downscale DNN, and AI downloads from the original image. An AI encoding apparatus according to an embodiment is disclosed for acquiring a scaled first image in downscale DNN, generating image data by first encoding the first image, and outputting the image data.

Description

Apparatus and method for AI encoding of images using low-complexity neural network, apparatus and method for AI decoding

The present disclosure relates to the field of image processing. More specifically, the present disclosure relates to a method and apparatus for encoding and decoding an image based on AI.

The image is encoded by a codec conforming to a predetermined data compression standard, for example, a Moving Picture Expert Group (MPEG) standard, and then stored in a recording medium in the form of a bitstream or transmitted through a communication channel.

With the development and dissemination of hardware capable of reproducing and storing high-resolution/high-definition images, the need for a codec capable of effectively encoding and decoding high-resolution/high-definition images is increasing.

An apparatus and method for AI encoding an image and an apparatus and method for decoding an AI according to an embodiment have a technical task of encoding and decoding an image based on AI to achieve a low bit rate.

An AI encoding apparatus according to an embodiment includes a processor executing one or more instructions stored in the AI encoding apparatus, wherein the processor provides a first reduced image downscaled from the original image and a resolution smaller than the resolution of the original image. inputting a reduced feature map having a reduced feature map into the downscale DNN, acquiring a first image downscaled by AI from the original image in the downscale DNN, first encoding the first image to generate image data, and the image data can be printed out.

In an exemplary embodiment, the processor may obtain a residual image between the second reduced image downscaled from the original image and the first reduced image as the reduced feature map.

In an exemplary embodiment, the processor acquires a plurality of first reduced images including pixels positioned at different points within the pixel groups of the original image, and the plurality of first reduced images and the second reduced image A plurality of residual images between the reduced images may be obtained as the reduced feature map.

In an exemplary embodiment, the sum of the number of the plurality of first reduced images and the number of the plurality of residual images may be equal to the number of input channels of the first layer of the downscale DNN.

In an exemplary embodiment, the first image may be obtained as the second reduced image and the output image of the last layer of the downscale DNN are summed.

In an exemplary embodiment, the processor may obtain an edge map corresponding to the original image as the reduced feature map.

In an exemplary embodiment, the first image may be obtained as the third reduced image downscaled from the original image and the output image of the last layer of the downscale DNN are summed.

In an exemplary embodiment, the processor obtains a downscaled and upscaled transformed image from the original image, and reduces the residual image between the downscaled fourth reduced image and the first reduced image from the transformed image. It can be obtained as a feature map.

In an exemplary embodiment, output data of any one of the layers of the downscale DNN may be combined with output data of layers before the one of the layers and input to the next layer.

An AI encoding method according to an embodiment comprises: inputting a first reduced image downscaled from an original image and a reduced feature map having a resolution smaller than a resolution of the original image into a downscale DNN; acquiring an AI downscaled first image from the original image in the downscale DNN; generating image data by first encoding the first image; and outputting the image data.

An AI decoding apparatus according to an embodiment includes a processor executing one or more instructions stored in the AI decoding apparatus, wherein the processor acquires image data generated as a result of a first encoding of a first image, and the image data first decodes to obtain a second image, and inputs a first enlarged image upscaled from the second image and an enlarged feature map having a resolution greater than the resolution of the second image as an upscale DNN, and the second image The AI upscaled third image may be obtained from the image in the upscale DNN, and the third image may be provided as a display.

The apparatus and method for AI encoding of an image and the apparatus and method for decoding an AI according to an embodiment may process an image at a low bit rate through AI-based image encoding and decoding.

However, effects that can be achieved by the apparatus and method for AI encoding of an image and the apparatus and method for decoding an AI according to an embodiment are not limited to those mentioned above, and other effects not mentioned above are disclosed from the description below. It will be clearly understood by those of ordinary skill in the art to which it belongs.

In order to more fully understand the drawings cited herein, a brief description of each drawing is provided.
1 is a diagram for explaining an AI encoding process and an AI decoding process according to an embodiment.
2 is a block diagram illustrating a configuration of an AI decoding apparatus according to an embodiment.
3 is an exemplary diagram illustrating a second DNN for AI upscaling of a second image.
4 is a diagram for explaining a convolution operation using a convolution layer.
5 is an exemplary diagram illustrating a mapping relationship between various pieces of image-related information and various pieces of DNN configuration information.
6 is a diagram illustrating a second image composed of a plurality of frames.
7 is a block diagram illustrating a configuration of an AI encoding apparatus according to an embodiment.
8 is an exemplary diagram illustrating a first DNN for AI downscaling of an original video.
9 is a diagram illustrating a structure of AI encoded data according to an embodiment.
10 is a diagram illustrating a structure of AI encoded data according to another embodiment.
11 is a diagram for explaining a method of training a first DNN and a second DNN.
12 is a diagram for explaining a training process of a first DNN and a second DNN by a training apparatus.
13 is a block diagram illustrating a configuration of an AI encoding apparatus according to another embodiment.
14 is a diagram for explaining an AI downscaling process using a first DNN according to an embodiment.
15 is a diagram for explaining an AI downscaling process using a first DNN according to another embodiment.
16 is a diagram for explaining an AI downscaling process using a first DNN according to another embodiment.
17 is an exemplary diagram illustrating a first DNN according to another embodiment.
18 is a block diagram illustrating a configuration of an AI decoding apparatus according to another embodiment.
19 is a diagram for explaining an AI upscaling process using a second DNN according to an embodiment.
20 is a diagram for explaining an AI upscaling process using a second DNN according to another embodiment.
21 is a diagram for explaining an AI upscaling process using a second DNN according to another embodiment.
22 is a diagram for describing a method of obtaining a residual image using a first reduced image (or a first enlarged image) and a second reduced image (a second enlarged image).
23 is a diagram for explaining a method of obtaining a residual image using a first reduced image (or a first enlarged image) and a second reduced image (a second enlarged image).
24 is a diagram for explaining a method of acquiring a residual image using a first reduced image (or a first enlarged image) and a second reduced image (a second enlarged image).
25 is a diagram for explaining another method of training a first DNN and a second DNN.

Since the present disclosure can make various changes and can have various embodiments, specific embodiments are illustrated in the drawings and will be described in detail through the detailed description. However, this is not intended to limit the embodiments of the present disclosure, and it should be understood that the present disclosure includes all modifications, equivalents and substitutes included in the spirit and scope of various embodiments.

In describing the embodiment, if it is determined that a detailed description of a related known technology may unnecessarily obscure the subject matter of the present disclosure, the detailed description thereof will be omitted. In addition, numbers (eg, first, second, etc.) used in the description process of the specification are only identifiers for distinguishing one component from other components.

In addition, in this specification, when a component is referred to as "connected" or "connected" with another component, the component may be directly connected or directly connected to the other component, but in particular It should be understood that, unless there is a description to the contrary, it may be connected or connected through another element in the middle.

In addition, in the present specification, components expressed as '~ part (unit)', 'module', etc. are two or more components combined into one component, or two or more components for each more subdivided function. may be differentiated into In addition, each of the components to be described below may additionally perform some or all of the functions of other components in addition to the main functions they are responsible for, and some of the main functions of each component may be different It goes without saying that it may be performed exclusively by the component.

Also, in this specification, an 'image' or 'picture' may indicate a still image, a moving picture composed of a plurality of continuous still images (or frames), or a video.

In addition, in the present specification, a 'deep neural network (DNN)' is a representative example of an artificial neural network model simulating a brain nerve, and is not limited to an artificial neural network model using a specific algorithm.

Also, in the present specification, a 'parameter' is a value used in a calculation process of each layer constituting a neural network, and may include, for example, a weight used when an input value is applied to a predetermined calculation expression. The parameter may be expressed in a matrix form. A parameter is a value set as a result of training, and may be updated through separate training data if necessary.

In addition, in this specification, 'first DNN' means a DNN used for AI downscaling of an image, and 'second DNN' means a DNN used for AI upscaling of an image.

In addition, in the present specification, 'DNN configuration information' includes the aforementioned parameters as information related to elements constituting the DNN. The first DNN or the second DNN may be configured using the DNN configuration information.

In addition, in this specification, the 'original image' refers to an image to be subjected to AI encoding, and the 'first image' refers to an image obtained as a result of AI downscaling of the original image in the AI encoding process. In addition, the 'second image' refers to an image obtained through the first decoding in the AI decoding process, and the 'third image' refers to an image obtained by AI upscaling the second image in the AI decoding process.

In addition, in this specification, 'AI downscaling' refers to a process of reducing the resolution of an image based on AI, and 'first encoding' refers to an encoding process by a frequency transform-based image compression method. In addition, 'first decoding' refers to a decoding process by a frequency transformation-based image restoration method, and 'AI upscaling' refers to a process of increasing the resolution of an image based on AI.

1 is a diagram for explaining an artificial intelligence (AI) encoding process and an AI decoding process according to an embodiment.

As described above, as the resolution of an image rapidly increases, the amount of information processing for encoding/decoding increases. Accordingly, a method for improving encoding and decoding efficiency of an image is required.

As shown in FIG. 1 , according to an embodiment of the present disclosure, a first image 115 is acquired by AI downscaling 110 of an original image 105 having a high resolution. In addition, since the first encoding 120 and the first decoding 130 are performed on the first image 115 of a relatively small resolution, the first encoding 120 and Compared to the case of performing the first decoding 130 , the bit rate may be greatly reduced.

1 , in an embodiment, in the AI encoding process, the original image 105 is AI downscaled 110 to obtain a first image 115 , and the first image 115 is first Encoding (120). In the AI decoding process, AI encoding data including AI data and image data obtained as a result of AI encoding is received, a second image 135 is obtained through the first decoding 130, and the second image 135 is A third image 145 is obtained by AI upscaling 140 .

Looking at the AI encoding process in more detail, when the original image 105 is input, the original image 105 is AI downscaled 110 to obtain the first image 115 of a predetermined resolution and/or a predetermined image quality. At this time, the AI downscaling 110 is performed based on AI, the AI for the AI downscaling 110 must be trained in connection with the AI for the AI upscaling 140 of the second image 135 (joint trained) do. Because, when the AI for the AI downscaling 110 and the AI for the AI upscaling 140 are separately trained, between the original image 105 that is the AI encoding target and the third image 145 restored through AI decoding because the difference between

In an embodiment of the present disclosure, AI data may be used in order to maintain this linkage in the AI encoding process and the AI decoding process. Therefore, the AI data obtained through the AI encoding process should include information indicating the upscale target, and in the AI decoding process, the second image 135 is AI upscaled ( 140) should be done.

AI for AI downscaling 110 and AI for AI upscaling 140 may be implemented as a deep neural network (DNN). As will be described later with reference to FIG. 11 , since the first DNN and the second DNN are jointly trained through sharing of loss information under a predetermined target, the AI encoding apparatus uses target information used when the first DNN and the second DNN are jointly trained. is provided to the AI decoding apparatus, and the AI decoding apparatus may upscale the AI to the image quality and/or resolution targeting the second image 135 based on the received target information.

When the first encoding 120 and the first decoding 130 shown in FIG. 1 are described in detail, the first image 115 AI downscaled 110 from the original image 105 is the first encoding 120 . can reduce the amount of information. The first encoding 120 includes a process of predicting the first image 115 to generate prediction data, a process of generating residual data corresponding to a difference between the first image 115 and the prediction data, and a spatial domain component. It may include a process of transforming the residual data into a frequency domain component, a process of quantizing the residual data transformed into a frequency domain component, and a process of entropy encoding the quantized residual data. Such a first encoding process 120 is MPEG-2, H.264 Advanced Video Coding (AVC), MPEG-4, High Efficiency Video Coding (HEVC), VC-1, VP8, VP9 and AV1 (AOMedia Video 1). It can be implemented through one of the image compression methods using frequency transformation, etc.

The second image 135 corresponding to the first image 115 may be restored through the first decoding 130 of image data. The first decoding 130 includes a process of generating quantized residual data by entropy-decoding image data, a process of inverse quantizing the quantized residual data, a process of converting the residual data of a frequency domain component into a spatial domain component, a process of predicting data It may include a process of generating , and a process of reconstructing the second image 135 using prediction data and residual data. The first decoding 130 process as described above is an image compression using frequency conversion such as MPEG-2, H.264, MPEG-4, HEVC, VC-1, VP8, VP9 and AV1 used in the first encoding 120 process. It may be implemented through an image restoration method corresponding to one of the methods.

The AI encoded data obtained through the AI encoding process may include image data obtained as a result of the first encoding 120 of the first image 115 and AI data related to the AI downscale 110 of the original image 105. can The image data may be used in the first decoding 130 process, and the AI data may be used in the AI upscaling 140 process.

The image data may be transmitted in the form of a bitstream. The image data may include data obtained based on pixel values in the first image 115 , for example, residual data that is a difference between the first image 115 and prediction data of the first image 115 . . Also, the image data includes information used in the process of the first encoding 120 of the first image 115 . For example, the image data includes prediction mode information used for the first encoding 120 of the first image 115 , motion information, and quantization parameter related information used in the first encoding 120 , etc. can do. The image data is an image compression method used in the first encoding 120 process among image compression methods using frequency conversion, such as MPEG-2, H.264 AVC, MPEG-4, HEVC, VC-1, VP8, VP9, and AV1. It may be generated according to a rule of, for example, a syntax.

AI data is used for AI upscaling 140 based on the second DNN. As described above, since the first DNN and the second DNN are jointly trained, the AI data includes information enabling the accurate AI upscaling 140 of the second image 135 through the second DNN to be performed. . In the AI decoding process, the AI upscaling 140 may be performed to a resolution and/or image quality targeting the second image 135 based on the AI data.

AI data may be transmitted together with image data in the form of a bitstream. According to an embodiment, AI data may be transmitted separately from image data in the form of frames or packets. Alternatively, according to an embodiment, AI data may be transmitted while being included in image data. The image data and AI data may be transmitted through the same network or different networks.

2 is a block diagram illustrating a configuration of an AI decoding apparatus 200 according to an embodiment.

Referring to FIG. 2 , the AI decoding apparatus 200 according to an embodiment includes a receiving unit 210 and an AI decoding unit 230 . The AI decoding unit 230 may include a parsing unit 232 , a first decoding unit 234 , an AI upscaling unit 236 , and an AI setting unit 238 .

Although the receiver 210 and the AI decoder 230 are illustrated as separate devices in FIG. 2 , the receiver 210 and the AI decoder 230 may be implemented through one processor. In this case, the receiving unit 210 and the AI decoding unit 230 may be implemented as a dedicated processor, and a general-purpose processor such as an application processor (AP), a central processing unit (CPU), or a graphic processing unit (GPU) and S/W It may be implemented through a combination of In addition, the dedicated processor may include a memory for implementing an embodiment of the present disclosure or a memory processing unit for using an external memory.

The receiver 210 and the AI decoder 230 may include a plurality of processors. In this case, it may be implemented as a combination of dedicated processors, or may be implemented through a combination of S/W with a plurality of general-purpose processors such as an AP, CPU, or GPU. In one embodiment, the receiving unit 210 is implemented as a first processor, the first decoding unit 234 is implemented as a second processor different from the first processor, the parsing unit 232, the AI upscaling unit 236 and the AI setting unit 238 may be implemented as a third processor different from the first processor and the second processor.

The receiving unit 210 receives AI encoded data obtained as a result of AI encoding. As an example, the AI-encoded data may be a video file having a file format such as mp4 or mov.

The receiver 210 may receive AI-encoded data transmitted through a network. The receiving unit 210 outputs the AI encoded data to the AI decoding unit 230 .

In one embodiment, the AI-encoded data is a hard disk, magnetic media such as floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floppy disks. It may be obtained from a data storage medium including an optical medium).

The parsing unit 232 parses the AI encoded data and transmits image data generated as a result of the first encoding of the first image 115 to the first decoder 234 , and transmits the AI data to the AI setting unit 238 . transmit

In an embodiment, the parsing unit 232 may parse the image data and the AI data separately included in the AI encoded data. The parsing unit 232 may read the header in the AI-encoded data to distinguish the AI data and the image data included in the AI-encoded data. In one example, AI data may be included in a Vendor Specific InfoFrame (VSIF) in the HDMI stream. The structure of AI encoded data including AI data and image data separated from each other will be described later with reference to FIG. 9 .

In another embodiment, the parsing unit 232 parses the image data from the AI encoded data, extracts the AI data from the image data, transmits the AI data to the AI setting unit 238, and first decodes the remaining image data. may be transmitted to the unit 234 . That is, AI data may be included in image data. For example, AI data may be included in supplemental enhancement information (SEI), which is an additional information area of a bitstream corresponding to image data. The structure of AI-encoded data including image data including AI data will be described later with reference to FIG. 10 .

In another embodiment, the parsing unit 232 divides the bitstream corresponding to the image data into a bitstream to be processed by the first decoding unit 234 and a bitstream corresponding to AI data, and divides each divided bitstream It may output to the first decoding unit 234 and the AI setting unit 238 .

The parsing unit 232 obtains image data included in the AI encoded data through a predetermined codec (eg, MPEG-2, H.264, MPEG-4, HEVC, VC-1, VP8, VP9, or AV1). It can also be confirmed that it is the image data that has been processed. In this case, the corresponding information may be transmitted to the first decoder 234 so that the image data can be processed by the identified codec.

The first decoder 234 reconstructs the second image 135 corresponding to the first image 115 based on the image data received from the parser 232 . The second image 135 obtained by the first decoder 234 is provided to the AI upscaler 236 .

According to an embodiment, first decoding-related information such as prediction mode information, motion information, and quantization parameter information may be provided from the first decoding unit 234 to the AI setting unit 238 . The first decoding related information may be used to obtain DNN configuration information.

The AI data provided to the AI setting unit 238 includes information enabling AI upscaling of the second image 135 . In this case, the upscale target of the second image 135 should correspond to the downscale target of the first DNN. Therefore, the AI data should include information that can confirm the downscale target of the first DNN.

Specifically exemplifying the information included in the AI data, there are information about a difference between the resolution of the original image 105 and the resolution of the first image 115 , and information related to the first image 115 .

The difference information may be expressed as information about a resolution conversion degree of the first image 115 compared to the original image 105 (eg, resolution conversion rate information). And, since the resolution of the first image 115 is known through the resolution of the restored second image 135 and the degree of resolution conversion can be confirmed through this, the difference information can be expressed only with the resolution information of the original image 105 . may be Here, the resolution information may be expressed as a horizontal/vertical screen size, or it may be expressed as a ratio (16:9, 4:3, etc.) and a size of one axis. In addition, when there is preset resolution information, it may be expressed in the form of an index or a flag.

The information related to the first image 115 is used for the resolution of the first image 115 , the bit rate of image data obtained as a result of the first encoding of the first image 115 , and the first encoding of the first image 115 . information on at least one of the specified codec types may be included.

The AI setting unit 238 may determine an upscale target of the second image 135 based on at least one of difference information included in the AI data and information related to the first image 115 . The upscaling target may indicate, for example, to which resolution the second image 135 should be upscaled.

When the upscaling target is determined, the AI upscaling unit 236 AI upscales the second image 135 through the second DNN to obtain a third image 145 corresponding to the upscaling target.

Prior to explaining a method for the AI setting unit 238 to determine an upscaling target based on AI data, an AI upscaling process through the second DNN will be described with reference to FIGS. 3 and 4 .

3 is an exemplary diagram illustrating a second DNN 300 for AI upscaling of a second image 135, and FIG. 4 is a convolution operation in the first convolution layer 310 shown in FIG. is showing

3 , the second image 135 is input to the first convolutional layer 310 . 3X3X4 displayed in the first convolution layer 310 shown in FIG. 3 exemplifies convolution processing on one input image using four filter kernels having a size of 3×3. As a result of the convolution process, four feature maps are generated by four filter kernels. Each feature map represents unique characteristics of the second image 135 . For example, each feature map may represent a vertical direction characteristic, a horizontal direction characteristic, or an edge characteristic of the second image 135 .

A convolution operation in the first convolution layer 310 will be described in detail with reference to FIG. 4 .

Through the multiplication operation and addition operation between the parameters of the filter kernel 430 having a size of 3 X 3 used in the first convolution layer 310 and the pixel values in the second image 135 corresponding thereto, one A feature map 450 may be generated. Since four filter kernels are used in the first convolution layer 310 , four feature maps may be generated through a convolution operation process using the four filter kernels.

In FIG. 4 , I1 to I49 displayed on the second image 135 indicate pixels of the second image 135 , and F1 to F9 displayed on the filter kernel 430 indicate parameters of the filter kernel 430 . Also, M1 to M9 displayed in the feature map 450 indicate samples of the feature map 450 .

4 illustrates that the second image 135 includes 49 pixels, but this is only an example, and when the second image 135 has a resolution of 4K, for example, 3840 X 2160 pixels It may contain pixels.

In the convolution operation process, pixel values of I1, I2, I3, I8, I9, I10, I15, I16, and I17 of the second image 135 and F1, F2, F3, F4, and F5 of the filter kernel 430 are respectively , F6, F7, F8, and F9 may each be multiplied, and a value obtained by combining (eg, addition operation) result values of the multiplication operation may be assigned as the value of M1 of the feature map 450 . If the stride of the convolution operation is 2, pixel values of I3, I4, I5, I10, I11, I12, I17, I18, and I19 of the second image 135 and F1 and F2 of the filter kernel 430, respectively , F3, F4, F5, F6, F7, F8, and F9 may each be multiplied, and a value obtained by combining result values of the multiplication operation may be assigned as the value of M2 of the feature map 450 .

A convolution operation between the pixel values in the second image 135 and the parameters of the filter kernel 430 is performed while the filter kernel 430 moves along the stride until the last pixel of the second image 135 is reached. By being performed, a feature map 450 having a predetermined size may be obtained.

According to the present disclosure, parameters of the second DNN through joint training of the first DNN and the second DNN, for example, parameters of the filter kernel used in convolutional layers of the second DNN (eg, filter The values of F1, F2, F3, F4, F5, F6, F7, F8, and F9 of the kernel 430 may be optimized. The AI setting unit 238 determines an upscale target corresponding to the downscale target of the first DNN based on the AI data, and uses parameters corresponding to the determined upscale target in the convolution layers of the second DNN. It can be determined by parameters of the filter kernel.

The convolution layers included in the first DNN and the second DNN may be processed according to the convolution operation process described with reference to FIG. 4 , but the convolution operation process described in FIG. 4 is only an example, and the it is not

Referring again to FIG. 3 , the feature maps output from the first convolutional layer 310 are input to the first activation layer 320 .

The first activation layer 320 may provide a non-linear characteristic to each feature map. The first activation layer 320 may include a sigmoid function, a Tanh function, a Rectified Linear Unit (ReLU) function, and the like, but is not limited thereto.

Giving the nonlinear characteristic in the first activation layer 320 means changing and outputting some sample values of the feature map, which is an output of the first convolution layer 310 . At this time, the change is performed by applying a non-linear characteristic.

The first activation layer 320 determines whether to transfer sample values of the feature maps output from the first convolution layer 310 to the second convolution layer 330 . For example, some sample values among the sample values of the feature maps are activated by the first activation layer 320 and transmitted to the second convolution layer 330 , and some sample values are activated by the first activation layer 320 . It is deactivated and is not transmitted to the second convolutional layer 330 . The intrinsic characteristic of the second image 135 indicated by the feature maps is emphasized by the first activation layer 320 .

The feature maps 325 output from the first activation layer 320 are input to the second convolution layer 330 . Any one of the feature maps 325 shown in FIG. 3 is a result of processing the feature map 450 described with reference to FIG. 4 in the first activation layer 320 .

3X3X4 displayed in the second convolution layer 330 exemplifies convolution processing on the input feature maps 325 using four filter kernels having a size of 3×3. The output of the second convolutional layer 330 is input to the second activation layer 340 . The second activation layer 340 may provide a non-linear characteristic to the input data.

The feature maps 345 output from the second activation layer 340 are input to the third convolution layer 350 . 3X3X1 displayed in the third convolution layer 350 shown in FIG. 3 exemplifies convolution processing to generate one output image using one filter kernel having a size of 3×3. The third convolution layer 350 is a layer for outputting a final image and generates one output using one filter kernel. According to the example of the present disclosure, the third convolution layer 350 may output the third image 145 through a convolution operation.

DNN setting information indicating the number of filter kernels of the first convolutional layer 310, the second convolutional layer 330, and the third convolutional layer 350 of the second DNN 300, parameters of the filter kernel, etc. As will be described later, there may be a plurality of pieces, and a plurality of pieces of DNN configuration information should be associated with a plurality of pieces of DNN configuration information of the first DNN. The association between the plurality of DNN configuration information of the second DNN and the plurality of DNN configuration information of the first DNN may be implemented through association learning of the first DNN and the second DNN.

3 shows that the second DNN 300 includes three convolutional layers 310 , 330 , 350 and two activation layers 320 , 340 , but this is only an example, and the implementation Accordingly, the number of convolutional layers and activation layers may be variously changed. Also, according to an embodiment, the second DNN 300 may be implemented through a recurrent neural network (RNN). In this case, it means changing the CNN structure of the second DNN 300 according to the example of the present disclosure to the RNN structure.

In an embodiment, the AI upscaling unit 236 may include at least one arithmetic logic unit (ALU) for the above-described convolution operation and operation of the activation layer. The ALU may be implemented as a processor. For the convolution operation, the ALU includes a multiplier that performs a multiplication operation between the sample values of the feature map output from the second image 135 or the previous layer and the sample values of the filter kernel, and an adder that adds the result values of the multiplication. can In addition, for the calculation of the activation layer, the ALU is a multiplier that multiplies an input sample value by a weight used in a predetermined sigmoid function, a Tanh function, or a ReLU function, and compares the multiplication result with a predetermined value to calculate the input sample value It may include a comparator that determines whether to transfer to the next layer.

Hereinafter, a method in which the AI setting unit 238 determines the upscaling target and the AI upscaling unit 236 AI upscales the second image 135 according to the upscaling target will be described.

In an embodiment, the AI setting unit 238 may store a plurality of DNN setting information settable in the second DNN.

Here, the DNN configuration information may include information on at least one of the number of convolutional layers included in the second DNN, the number of filter kernels for each convolutional layer, and parameters of each filter kernel.

The plurality of DNN configuration information may each correspond to various upscale targets, and the second DNN may operate based on the DNN configuration information corresponding to a specific upscale target. The second DNN may have a different structure according to the DNN configuration information. For example, the second DNN may include three convolutional layers according to certain DNN configuration information, and the second DNN may include four convolutional layers according to other DNN configuration information.

In an embodiment, the DNN configuration information may include only parameters of a filter kernel used in the second DNN. In this case, the structure of the second DNN is not changed, but only the parameters of the internal filter kernel may be changed according to the DNN configuration information.

The AI setting unit 238 may acquire DNN setting information for AI upscaling of the second image 135 among a plurality of DNN setting information. Each of the plurality of DNN configuration information used herein is information for acquiring the third image 145 of a predetermined resolution and/or predetermined image quality, and is trained in association with the first DNN.

For example, the DNN setting information of any one of the plurality of DNN setting information is a third image 145 having a resolution twice that of the second image 135, for example, a second image of 2K (2048 * 1080). It may include information for acquiring the third image 145 of 4K (4096*2160) that is twice as large as the 2nd image 135 , and the other DNN setting information is 4 more than the resolution of the second image 135 . Information for obtaining a third image 145 of twice the resolution, for example, a third image 145 of 8K (8192 * 4320) that is 4 times larger than the second image 135 of 2K (2048 * 1080) may include

Each of the plurality of DNN setting information is created in association with the DNN setting information of the first DNN of the AI encoding device 700, and the AI setting unit 238 is set at an enlargement ratio corresponding to the reduction rate of the DNN setting information of the first DNN. Accordingly, one piece of DNN configuration information among a plurality of DNN configuration information is acquired. To this end, the AI setting unit 238 should check the information of the first DNN. In order for the AI setting unit 238 to check the information of the first DNN, the AI decoding apparatus 200 according to an embodiment receives AI data including the information of the first DNN from the AI encoding apparatus 700 .

In other words, the AI setting unit 238 uses the information received from the AI encoding device 700 to identify information targeted by the DNN setting information of the first DNN used to obtain the first image 115 , and , it is possible to obtain the DNN configuration information of the second DNN trained in association with it.

When DNN setting information for AI upscaling of the second image 135 is obtained among a plurality of DNN setting information, the DNN setting information is transmitted to the AI upscaling unit 236, and a second DNN operating according to the DNN setting information. Based on the input data may be processed.

For example, when any one DNN configuration information is obtained, the AI upscaling unit 236 includes the first convolutional layer 310 and the second convolutional layer 330 of the second DNN 300 shown in FIG. 3 . ) and the third convolutional layer 350, the number of filter kernels included in each layer and parameters of the filter kernels are set to values included in the obtained DNN configuration information.

Specifically, the AI upscaling unit 236 determines that the parameters of the filter kernel of 3 X 3 used in any one convolution layer of the second DNN shown in FIG. 4 are {1, 1, 1, 1, 1, 1 , 1, 1, 1}, and when there is a change in DNN configuration information, the parameters of the filter kernel are changed to {2, 2, 2, 2, 2, 2, 2, 2, 2, parameters included in the changed DNN configuration information. } can be replaced.

The AI setting unit 238 may acquire DNN setting information for upscaling the second image 135 among a plurality of DNN setting information based on the information included in the AI data, which is used to obtain the DNN setting information. AI data will be explained in detail.

In an embodiment, the AI setting unit 238 may obtain DNN setting information for upscaling the second image 135 among a plurality of DNN setting information, based on the difference information included in the AI data. For example, based on the difference information, the resolution of the original image 105 (eg, 4K (4096 * 2160)) is higher than the resolution of the first image 115 (eg, 2K (2048 * 1080)). When it is confirmed that the size is 2 times larger, the AI setting unit 238 may obtain DNN setting information capable of increasing the resolution of the second image 135 by 2 times.

In another embodiment, the AI setting unit 238 is based on the information related to the first image 115 included in the AI data, DNN setting information for AI upscaling the second image 135 among a plurality of DNN setting information. can be obtained. The AI setting unit 238 may determine a mapping relationship between the image related information and the DNN setting information in advance, and obtain DNN setting information mapped to the first image 115 related information.

5 is an exemplary diagram illustrating a mapping relationship between various pieces of image-related information and various pieces of DNN configuration information.

5, it can be seen that the AI encoding/AI decoding process according to an embodiment of the present disclosure does not only consider a change in resolution. As shown in FIG. 5, DNN setting information considering resolutions such as SD, HD, Full HD, bitrates such as 10Mbps, 15Mbps, and 20Mbps, and codec information such as AV1, H.264, and HEVC individually or all selection can be made. For this consideration, training in consideration of each element in the AI training process should be performed in connection with the encoding and decoding processes (see FIG. 11 ).

Accordingly, as shown in FIG. 5 according to the training content, when a plurality of DNN setting information is provided based on image-related information including a codec type and image resolution, the first image ( 115) DNN setting information for AI upscaling of the second image 135 may be acquired based on the related information.

That is, the AI setting unit 238 matches the image-related information shown on the left side of the table shown in FIG. 5 with the DNN setting information on the right side of the table, so that the DNN setting information according to the image-related information can be used.

5 , from the information related to the first image 115 , the resolution of the first image 115 is SD, and the bit rate of image data obtained as a result of the first encoding of the first image 115 is 10Mbps, , when it is confirmed that the first image 115 is first coded with the AV1 codec, the AI setting unit 238 may acquire A DNN setting information among a plurality of DNN setting information.

In addition, from the information related to the first image 115 , the resolution of the first image 115 is HD, the bit rate of image data obtained as a result of the first encoding is 15 Mbps, and the first image 115 is converted to the H.264 codec. When it is confirmed that the first encoding is performed, the AI setting unit 238 may acquire B DNN setting information among a plurality of DNN setting information.

In addition, from the information related to the first image 115 , the resolution of the first image 115 is Full HD, the bit rate of image data obtained as a result of the first encoding of the first image 115 is 20 Mbps, and the first image ( 115) is confirmed to be first encoded by the HEVC codec, the AI setting unit 238 obtains C DNN setting information among a plurality of DNN setting information, the resolution of the first image 115 is Full HD, and the first When the bit rate of the image data obtained as a result of the first encoding of the image 115 is 15 Mbps and it is confirmed that the first image 115 is first encoded with the HEVC codec, the AI setting unit 238 sets a plurality of DNN setting information Among them, DNN configuration information can be obtained. Either one of the C DNN setting information and the D DNN setting information is selected according to whether the bit rate of the image data obtained as a result of the first encoding of the first image 115 is 20 Mbps or 15 Mbps. When the first image 115 of the same resolution is first encoded with the same codec, the fact that the bit rates of the image data are different from each other means that the image quality of the reconstructed images is different from each other. Therefore, the first DNN and the second DNN may be jointly trained based on a predetermined image quality, and accordingly, the AI setting unit 238 sets the DNN according to the bit rate of the image data indicating the image quality of the second image 135 . information can be obtained.

In another embodiment, the AI setting unit 238 relates to the information (prediction mode information, motion information, quantization parameter information, etc.) provided from the first decoder 234 and the first image 115 included in the AI data. DNN configuration information for AI upscaling of the second image 135 among a plurality of pieces of DNN configuration information may be acquired in consideration of all information. For example, the AI setting unit 238 receives quantization parameter information used in the first encoding process of the first image 115 from the first decoder 234 , and receives the quantization parameter information from the AI data of the first image 115 . A bit rate of image data obtained as a result of encoding may be checked, and DNN setting information corresponding to a quantization parameter and a bit rate may be obtained. Even at the same bit rate, there may be a difference in the quality of the reconstructed image depending on the complexity of the image, and the bit rate is a value representative of the entire first image 115 to be first encoded, and the bit rate of each frame in the first image 115 is The picture quality may be different. Accordingly, when the prediction mode information, motion information, and/or quantization parameters obtainable from the first decoder 234 for each frame are considered together, the DNN setting more suitable for the second image 135 compared to using only AI data information can be obtained.

Also, depending on the implementation, the AI data may include an identifier of mutually agreed DNN configuration information. The identifier of the DNN setting information is an upscale target corresponding to the downscale target of the first DNN, and the DNN setting information trained in association between the first DNN and the second DNN so that the second image 135 can be AI upscaled. It is information for distinguishing pairs of . The AI setting unit 238 obtains the identifier of the DNN setting information included in the AI data, and then obtains the DNN setting information corresponding to the identifier of the DNN setting information, and the AI upscaling unit 236 receives the corresponding DNN setting information. The second image 135 may be upscaled using AI. For example, an identifier indicating each of a plurality of pieces of DNN configuration information configurable in the first DNN and an identifier indicating each of a plurality of pieces of DNN configuration information configurable in the second DNN may be predefined. In this case, the same identifier may be designated for a pair of DNN configuration information configurable to each of the first DNN and the second DNN. The AI data may include an identifier of DNN setting information set in the first DNN for AI downscaling of the original image 105 . The AI setting unit 238 that has received the AI data acquires the DNN setting information indicated by the identifier included in the AI data among a plurality of DNN setting information, and the AI upscaling unit 236 uses the corresponding DNN setting information to obtain the second The image 135 may be upscaled by AI.

Also, according to implementation, AI data may include DNN configuration information. The AI setting unit 238 may obtain DNN setting information included in AI data, and the AI upscaling unit 236 may AI upscale the second image 135 using the corresponding DNN setting information.

According to the embodiment, when information constituting the DNN setting information (eg, the number of convolution layers, the number of filter kernels for each convolutional layer, parameters of each filter kernel, etc.) is stored in the form of a lookup table, The AI setting unit 238 obtains DNN setting information by combining some selected among the lookup table values based on information included in the AI data, and the AI upscaling unit 236 uses the corresponding DNN setting information to obtain the second image (135) may be AI upscaled.

According to an embodiment, when the structure of the DNN corresponding to the upscale target is determined, the AI setting unit 238 may obtain DNN setting information corresponding to the determined structure of the DNN, for example, parameters of the filter kernel.

As described above, the AI setting unit 238 obtains DNN setting information of the second DNN through AI data including information related to the first DNN, and the AI upscaling unit 236 sets the DNN setting information to the corresponding DNN setting information. AI upscales the second image 135 through the second DNN, which may reduce memory usage and computational amount compared to upscaling by directly analyzing the characteristics of the second image 135 .

In an embodiment, when the second image 135 consists of a plurality of frames, the AI setting unit 238 may independently obtain DNN setting information for each predetermined number of frames, or a common DNN for all frames. It is also possible to obtain setting information.

6 is a diagram illustrating a second image 135 composed of a plurality of frames.

As shown in FIG. 6 , the second image 135 may include frames corresponding to t0 to tn.

In one example, the AI setting unit 238 obtains DNN setting information of the second DNN through AI data, and the AI upscaling unit 236 AI frames corresponding to t0 to tn based on the corresponding DNN setting information. can be upscaled. That is, frames corresponding to t0 to tn may be AI upscaled based on common DNN configuration information.

In another example, the AI setting unit 238 obtains 'A' DNN setting information from AI data for some of the frames corresponding to t0 to tn, for example, frames corresponding to t0 to ta. and 'B' DNN configuration information may be obtained from AI data for frames corresponding to ta+1 to tb. Also, the AI setting unit 238 may obtain 'C' DNN setting information from AI data for frames corresponding to tb+1 to tn. In other words, the AI setting unit 238 independently obtains DNN setting information for each group including a predetermined number of frames among the plurality of frames, and the AI upscaling unit 236 independently selects the frames included in each group. AI upscaling is possible with the acquired DNN configuration information.

In another example, the AI setting unit 238 may independently obtain DNN setting information for each frame constituting the second image 135 . For example, when the second image 135 consists of three frames, the AI setting unit 238 obtains DNN setting information in relation to the first frame, and DNN setting information in relation to the second frame. and may acquire DNN configuration information in relation to the third frame. That is, DNN configuration information may be independently obtained for each of the first frame, the second frame, and the third frame. DNN setting information is obtained based on the information (prediction mode information, motion information, quantization parameter information, etc.) provided from the above-described first decoder 234 and information related to the first image 115 included in the AI data According to the method, DNN setting information may be independently obtained for each frame constituting the second image 135 . This is because mode information, quantization parameter information, and the like can be independently determined for each frame constituting the second image 135 .

In another example, the AI data may include information indicating to which frame the DNN configuration information obtained based on the AI data is valid. For example, if the AI data includes information that the DNN setting information is valid until the ta frame, the AI setting unit 238 acquires the DNN setting information based on the AI data, and the AI upscaling unit 236 is AI upscales t0 to ta frames with corresponding DNN configuration information. And, when the information that the DNN setting information is valid until the tn frame is included in other AI data, the AI setting unit 238 acquires the DNN setting information based on the other AI data, and the AI upscaling unit 236 may AI upscale ta+1 to tn frames with the obtained DNN configuration information.

Hereinafter, an AI encoding apparatus 700 for AI encoding of the original image 105 will be described with reference to FIG. 7 .

7 is a block diagram illustrating a configuration of an AI encoding apparatus 700 according to an embodiment.

Referring to FIG. 7 , the AI encoding apparatus 700 may include an AI encoding unit 710 and a transmission unit 730 . The AI encoder 710 may include an AI downscaler 712 , a first encoder 714 , a data processor 716 , and an AI setting unit 718 .

7 illustrates the AI encoder 710 and the transmitter 730 as separate devices, the AI encoder 710 and the transmitter 730 may be implemented through one processor. In this case, it may be implemented as a dedicated processor, or it may be implemented through a combination of an AP or a general-purpose processor such as a CPU or GPU and S/W. In addition, the dedicated processor may include a memory for implementing an embodiment of the present disclosure or a memory processing unit for using an external memory.

Also, the AI encoder 710 and the transmitter 730 may include a plurality of processors. In this case, it may be implemented as a combination of dedicated processors, or may be implemented through a combination of S/W with a plurality of general-purpose processors such as an AP, CPU, or GPU. In an embodiment, the first encoding unit 714 is configured as a first processor, and the AI downscale unit 712 , the data processing unit 716 , and the AI setting unit 718 are configured as a second processor different from the first processor. In the implementation, the transmitter 730 may be implemented as a third processor different from the first processor and the second processor.

The AI encoder 710 performs AI downscaling of the original image 105 and the first encoding of the first image 115 , and transmits the AI encoded data to the transmitter 730 . The transmitter 730 transmits the AI encoded data to the AI decoding apparatus 200 .

The image data includes data obtained as a result of the first encoding of the first image 115 . The image data may include data obtained based on pixel values in the first image 115 , for example, residual data that is a difference between the first image 115 and prediction data of the first image 115 . . Also, the image data includes information used in the first encoding process of the first image 115 . For example, the image data may include prediction mode information used for first encoding the first image 115 , motion information, and quantization parameter related information used for first encoding the first image 115 , etc. .

The AI data includes information enabling the AI upscaling unit 236 to AI upscale the second image 135 to an upscale target corresponding to the downscale target of the first DNN.

In one example, the AI data may include difference information between the original image 105 and the first image 115 .

In an example, the AI data may include information related to the first image 115 . The information related to the first image 115 is used for the resolution of the first image 115 , the bit rate of image data obtained as a result of the first encoding of the first image 115 , and the first encoding of the first image 115 . information on at least one of the specified codec types may be included.

In an embodiment, the AI data may include an identifier of mutually agreed DNN configuration information so that the second image 135 can be AI upscaled to an upscale target corresponding to the downscale target of the first DNN. .

Also, in an embodiment, the AI data may include DNN configuration information settable in the second DNN.

The AI downscaler 712 may acquire the AI downscaled first image 115 from the original image 105 through the first DNN. The AI downscaling unit 712 may AI downscale the original image 105 using the DNN setting information provided from the AI setting unit 718 .

The AI setting unit 718 may determine a downscale target of the original image 105 based on a predetermined criterion.

In order to acquire the first image 115 matching the downscale target, the AI setting unit 718 may store a plurality of settable DNN setting information in the first DNN. The AI setting unit 718 obtains DNN setting information corresponding to the downscale target among a plurality of DNN setting information, and provides the obtained DNN setting information to the AI downscale unit 712 .

Each of the plurality of pieces of DNN configuration information may be trained to obtain the first image 115 having a predetermined resolution and/or a predetermined image quality. For example, the DNN setting information of any one of the plurality of DNN setting information is the first image 115 having a resolution that is 1/2 times smaller than the resolution of the original image 105, for example, 4K (4096*2160) may include information for obtaining the first image 115 of 2K (2048*1080) 1/2 times smaller than the original image 105 of A first image 115 of a resolution of 1/4 times smaller than that of a first image 115, for example, a first image 115 of 2K (2048*1080) which is 1/4 times smaller than an original image 105 of 8K (8192*4320). ) may include information for obtaining

According to the embodiment, when information constituting the DNN setting information (eg, the number of convolution layers, the number of filter kernels for each convolutional layer, parameters of each filter kernel, etc.) is stored in the form of a lookup table, The AI setting unit 718 may obtain DNN setting information by combining some selected among the lookup table values according to the downscale target, and may provide the obtained DNN setting information to the AI downscaling unit 712 .

According to an embodiment, the AI setting unit 718 may determine the structure of the DNN corresponding to the downscale target, and obtain DNN setting information corresponding to the determined structure of the DNN, for example, parameters of the filter kernel.

The plurality of DNN configuration information for AI downscaling of the original image 105 may have an optimized value by performing joint training of the first DNN and the second DNN. Here, each DNN configuration information includes at least one of the number of convolutional layers included in the first DNN, the number of filter kernels for each convolutional layer, and parameters of each filter kernel.

The AI downscaling unit 712 sets the first DNN as the DNN setting information determined for the AI downscaling of the original image 105, and converts the first image 115 of a predetermined resolution and/or a predetermined quality to the first DNN. can be obtained through When DNN setting information for AI downscaling of the original image 105 is obtained among a plurality of DNN setting information, each layer in the first DNN may process input data based on information included in the DNN setting information. .

Hereinafter, a method for the AI setting unit 718 to determine the downscale target will be described. The downscale target may indicate, for example, how much the first image 115 with reduced resolution should be obtained from the original image 105 .

The AI setting unit 718 acquires one or more pieces of input information. In an embodiment, the input information includes a target resolution of the first image 115, a target bitrate of the image data, a bitrate type of the image data (eg, a variable bitrate type, a constant bitrate type, or an average bitrate type); Color format to which AI downscale is applied (luminance component, chrominance component, red component, green component, or blue component, etc.), codec type for first encoding of first image 115, compression history information, original image 105 at least one of a resolution of , and a type of the original image 105 .

One or more pieces of input information may include information previously stored in the AI encoding apparatus 700 or received from a user.

The AI setting unit 718 controls the operation of the AI downscale unit 712 based on the input information. In an embodiment, the AI setting unit 718 may determine a downscale target according to input information, and provide DNN setting information corresponding to the determined downscale target to the AI downscale unit 712 .

In an embodiment, the AI setting unit 718 transmits at least a portion of the input information to the first encoding unit 714 so that the first encoding unit 714 provides a bitrate of a specific value, a bitrate of a specific type, and a specific codec. may cause the first image 115 to be first encoded.

In an embodiment, the AI setting unit 718 may include a compression rate (eg, a difference in resolution between the original image 105 and the first image 115 , a target bitrate), and a compression quality (eg, a bitrate type). ), compression history information, and a downscale target may be determined based on at least one of the type of the original image 105 .

In an example, the AI setting unit 718 may determine a downscale target based on a preset or a compression rate or compression quality input from a user.

As another example, the AI setting unit 718 may determine the downscale target by using the compression history information stored in the AI encoding apparatus 700 . For example, according to the compression history information available to the AI encoding apparatus 700, the encoding quality or compression rate preferred by the user may be determined, and the downscale target may be determined according to the encoding quality determined based on the compression history information. can For example, the resolution and quality of the first image 115 may be determined according to the encoding quality that has been most frequently used according to the compression history information.

As another example, the AI setting unit 718 sets the encoding quality that has been used more than a predetermined threshold value (eg, average quality of encoding qualities that have been used more than a predetermined threshold value) according to the compression history information. A downscale target may be determined based on it.

As another example, the AI setting unit 718 may determine the downscale target based on the resolution and type (eg, file format) of the original image 105 .

In one embodiment, when the original image 105 is composed of a plurality of frames, the AI setting unit 718 independently acquires DNN setting information for each predetermined number of frames, and uses the independently acquired DNN setting information to AI downscaling. A portion 712 may be provided.

In one example, the AI setting unit 718 may divide the frames constituting the original image 105 into a predetermined number of groups, and independently obtain DNN setting information for each group. The same or different DNN configuration information may be obtained for each group. The number of frames included in the groups may be the same or different for each group.

In another example, the AI setting unit 718 may independently determine DNN setting information for each frame constituting the original image 105 . The same or different DNN configuration information may be obtained for each frame.

Hereinafter, an exemplary structure of the first DNN 800 as a basis for AI downscaling will be described.

8 is an exemplary diagram illustrating the first DNN 800 for AI downscaling of the original image 105 .

As shown in FIG. 8 , the original image 105 is input to the first convolutional layer 810 . The first convolutional layer 810 performs convolution processing on the original image 105 using 32 filter kernels having a size of 5×5. The 32 feature maps generated as a result of the convolution process are input to the first activation layer 820 .

The first activation layer 820 may provide non-linear characteristics to 32 feature maps.

The first activation layer 820 determines whether to transfer sample values of the feature maps output from the first convolution layer 810 to the second convolution layer 830 . For example, some sample values among the sample values of the feature maps are activated by the first activation layer 820 and transmitted to the second convolution layer 830 , and some sample values are activated by the first activation layer 820 . It is deactivated and is not transmitted to the second convolutional layer 830 . Information indicated by the feature maps output from the first convolution layer 810 is emphasized by the first activation layer 820 .

The output 825 of the first activation layer 820 is input to the second convolution layer 830 . The second convolution layer 830 performs convolution processing on input data by using 32 filter kernels having a size of 5×5. The 32 feature maps output as a result of the convolution process are input to the second activation layer 840 , and the second activation layer 840 may provide nonlinear characteristics to the 32 feature maps.

The output 845 of the second activation layer 840 is input to the third convolution layer 850 . The third convolution layer 850 performs convolution processing on input data using one filter kernel having a size of 5×5. As a result of the convolution processing, one image may be output from the third convolutional layer 850 . The third convolutional layer 850 is a layer for outputting a final image and obtains one output by using one filter kernel. According to the example of the present disclosure, the third convolution layer 850 may output the first image 115 through the convolution operation result.

DNN setting information indicating the number of filter kernels of the first convolutional layer 810, the second convolutional layer 830, and the third convolutional layer 850 of the first DNN 800, parameters of the filter kernel, etc. There may be a plurality, and the plurality of DNN configuration information should be associated with the plurality of DNN configuration information of the second DNN. The association between the plurality of DNN configuration information of the first DNN and the plurality of DNN configuration information of the second DNN may be implemented through association learning of the first DNN and the second DNN.

8 shows that the first DNN 800 includes three convolutional layers 810 , 830 , 850 and two activation layers 820 , 840 , but this is only an example, and the implementation Accordingly, the number of convolutional layers and activation layers may be variously changed. Also, according to an embodiment, the first DNN 800 may be implemented through a recurrent neural network (RNN). In this case, it means changing the CNN structure of the first DNN 800 according to the example of the present disclosure to the RNN structure.

In an embodiment, the AI downscaling unit 712 may include at least one ALU for a convolution operation and an activation layer operation. The ALU may be implemented as a processor. For the convolution operation, the ALU may include a multiplier that performs a multiplication operation between the sample values of the feature map output from the original image 105 or the previous layer and the sample values of the filter kernel, and an adder that adds the result values of the multiplication. have. In addition, for the calculation of the activation layer, the ALU is a multiplier that multiplies an input sample value by a weight used in a predetermined sigmoid function, a Tanh function, or a ReLU function, and compares the multiplication result with a predetermined value to calculate the input sample value It may include a comparator that determines whether to transfer to the next layer.

Referring back to FIG. 7 , the AI setting unit 718 transmits AI data to the data processing unit 716 . The AI data includes information enabling the AI upscaling unit 236 to AI upscale the second image 135 to an upscale target corresponding to the downscale target of the first DNN.

The first encoder 714 receiving the first image 115 from the AI downscaler 712 first encodes the first image 115 according to the frequency transformation-based image compression method to obtain the first image 115 . ) can reduce the amount of information it has. As a result of the first encoding through a predetermined codec (eg, MPEG-2, H.264, MPEG-4, HEVC, VC-1, VP8, VP9, or AV1), image data is obtained. The image data is obtained according to a rule of a predetermined codec, that is, a syntax. For example, the image data includes residual data that is a difference between the first image 115 and the prediction data of the first image 115 , prediction mode information used to first encode the first image 115 , motion information, and Information related to a quantization parameter used to first encode the first image 115 may be included.

The image data obtained as a result of the first encoding by the first encoder 714 is provided to the data processor 716 .

The data processing unit 716 generates AI encoded data including the image data received from the first encoding unit 714 and the AI data received from the AI setting unit 718 .

In an embodiment, the data processing unit 716 may generate AI-encoded data including image data and AI data in a separate state. As an example, AI data may be included in a Vendor Specific InfoFrame (VSIF) in the HDMI stream.

In another embodiment, the data processing unit 716 may include AI data in image data obtained as a result of the first encoding by the first encoding unit 714 and generate AI encoded data including the corresponding image data. . For example, the data processing unit 716 may generate image data in the form of one bitstream by combining a bitstream corresponding to image data and a bitstream corresponding to AI data. To this end, the data processing unit 716 may express AI data as bits having a value of 0 or 1, that is, a bitstream. In an embodiment, the data processing unit 716 may include a bitstream corresponding to AI data in supplemental enhancement information (SEI), which is an additional information area of a bitstream obtained as a result of the first encoding.

AI-encoded data is transmitted to the transmitter 730 . The transmitter 730 transmits the AI-encoded data obtained as a result of the AI-encoding through a network.

In one embodiment, the AI encoded data is a hard disk, a magnetic medium such as a floppy disk and magnetic tape, an optical recording medium such as CD-ROM and DVD, a magneto-optical medium such as a floppy disk. medium) may be stored in a data storage medium including

9 is a diagram illustrating a structure of AI-encoded data 900 according to an embodiment.

As described above, the AI data 912 and the image data 932 may be separately included in the AI encoded data 900 . Here, the AI-encoded data 900 may be in a container format such as MP4, AVI, MKV, or FLV. The AI-encoded data 900 may include a metadata box 910 and a media data box 930 .

The metadata box 910 includes information about the image data 932 included in the media data box 930 . For example, the metadata box 910 may include information on the type of the first image 115 , the type of codec used to encode the first image 115 , and the playback time of the first image 115 . have. Also, the metadata box 910 may include AI data 912 . The AI data 912 may be encoded according to an encoding method provided by a predetermined container format and stored in the metadata box 910 .

The media data box 930 may include image data 932 generated according to the syntax of a predetermined image compression method.

10 is a diagram illustrating a structure of AI encoded data 1000 according to another embodiment.

Referring to FIG. 10 , AI data 1034 may be included in image data 1032 . The AI-encoded data 1000 may include a metadata box 1010 and a media data box 1030 . When the AI data 1034 is included in the image data 1032 , the metadata box 1010 contains AI Data 1034 may not be included.

The media data box 1030 includes image data 1032 including AI data 1034 . For example, the AI data 1034 may be included in the additional information area of the image data 1032 .

Hereinafter, a method of jointly training the first DNN 800 and the second DNN 300 will be described with reference to FIG. 11 .

11 is a diagram for explaining a method of training the first DNN 800 and the second DNN 300 .

In an embodiment, the AI-encoded original image 105 through the AI encoding process is restored to the third image 145 through the AI decoding process. The third image 145 and the original image 105 obtained as a result of AI decoding In order to maintain the similarity with the AI encoding process and the AI decoding process, correlation is required. That is, information lost in the AI encoding process should be able to be restored in the AI decoding process, and for this purpose, joint training between the first DNN 800 and the second DNN 300 is required.

For accurate AI decoding, it is ultimately necessary to reduce the quality loss information 1130 corresponding to the comparison result between the third training image 1104 and the original training image 1101 shown in FIG. 11 . Accordingly, the quality loss information 1130 is used for both training of the first DNN 800 and the second DNN 300 .

First, the training process shown in FIG. 11 will be described.

In FIG. 11 , an original training image 1101 is an image subject to AI downscale, and a first training image 1102 is an AI downscaled image from the original training image 1101 . it's a video Also, the third training image 1104 is an AI upscaled image from the first training image 1102 .

The original training image 1101 includes a still image or a moving image composed of a plurality of frames. In an embodiment, the original training image 1101 may include a still image or a luminance image extracted from a moving picture including a plurality of frames. Also, according to an embodiment, the original training image 1101 may include a still image or a patch image extracted from a moving image including a plurality of frames. When the original training image 1101 consists of a plurality of frames, the first training image 1102 , the second training image and the third training image 1104 also include a plurality of frames. When a plurality of frames of the original training image 1101 are sequentially input to the first DNN 800 , the first training image 1102 and the second training image are performed through the first DNN 800 and the second DNN 300 . and a plurality of frames of the third training image 1104 may be sequentially acquired.

For joint training of the first DNN 800 and the second DNN 300 , an original training image 1101 is input to the first DNN 800 . The original training image 1101 input to the first DNN 800 is AI downscaled and output as the first training image 1102 , and the first training image 1102 is input to the second DNN 300 . A third training image 1104 is output as a result of AI upscaling for the first training image 1102 .

Referring to FIG. 11 , a first training image 1102 is being input to the second DNN 300 . According to an embodiment, the first training image 1102 is obtained through a first encoding and a first decoding process. A second training image may be input to the second DNN 300 . In order to input the second training image to the second DNN, any one of MPEG-2, H.264, MPEG-4, HEVC, VC-1, VP8, VP9, and AV1 codecs may be used. Specifically, in the first encoding of the first training image 1102 and the first decoding of image data corresponding to the first training image 1102, MPEG-2, H.264, MPEG-4, HEVC, VC-1 , VP8, VP9, and any one codec of AV1 may be used.

Referring to FIG. 11 , the legacy downscaled reduced training image 1103 is obtained from the original training image 1101 separately from the output of the first training image 1102 through the first DNN 800 . Here, the legacy downscale may include at least one of a bilinear scale, a bicubic scale, a lanczos scale, and a stair step scale.

In order to prevent the structural features of the first image 115 from greatly deviate based on the structural features of the original image 105, a reduced training image 1103 that preserves the structural features of the original training image 1101 is acquired. .

Before the training proceeds, the first DNN 800 and the second DNN 300 may be set with predetermined DNN configuration information. As training progresses, structural loss information 1110 , complexity loss information 1120 , and quality loss information 1130 may be determined.

The structural loss information 1110 may be determined based on a comparison result between the reduced training image 1103 and the first training image 1102 . In one example, the structural loss information 1110 may correspond to a difference between the structural information of the reduced training image 1103 and the structural information of the first training image 1102 . The structural information may include various features that can be extracted from the image, such as luminance, contrast, and histogram of the image. The structural loss information 1110 indicates to what extent the structural information of the original training image 1101 is maintained in the first training image 1102 . As the structural loss information 1110 is smaller, the structural information of the first training image 1102 is similar to the structural information of the original training image 1101 .

The complexity loss information 1120 may be determined based on the spatial complexity of the first training image 1102 . In an example, as the spatial complexity, a total variance value of the first training image 1102 may be used. The complexity loss information 1120 is related to a bit rate of image data obtained by first encoding the first training image 1102 . The smaller the complexity loss information 1120 is, the smaller the bit rate of image data is defined.

The quality loss information 1130 may be determined based on a comparison result between the original training image 1101 and the third training image 1104 . The quality loss information 1130 includes an L1-norm value, an L2-norm value, a Structural Similarity (SSIM) value, and a Peak Signal-To (PSNR-HVS) value for the difference between the original training image 1101 and the third training image 1104 . It may include at least one of a Noise Ratio-Human Vision System) value, a Multiscale SSIM (MS-SSIM) value, a Variance Inflation Factor (VIF) value, and a Video Multimethod Assessment Fusion (VMAF) value. The quality loss information 1130 indicates how similar the third training image 1104 is to the original training image 1101 . As the quality loss information 1130 is smaller, the third training image 1104 is more similar to the original training image 1101 .

11, structural loss information 1110, complexity loss information 1120, and quality loss information 1130 are used for training of the first DNN 800, and the quality loss information 1130 is the second DNN ( 300) is used for training. That is, the quality loss information 1130 is used for both training of the first DNN 800 and the second DNN 300 .

The first DNN 800 may update parameters such that the final loss information determined based on the structural loss information 1110 , the complexity loss information 1120 , and the quality loss information 1130 is reduced or minimized. Also, the second DNN 300 may update the parameter so that the quality loss information 1130 is reduced or minimized.

Final loss information for training of the first DNN 800 and the second DNN 300 may be determined as in Equation 1 below.

[Equation 1]

In Equation 1, LossDS represents final loss information to be reduced or minimized for training of the first DNN 800, and LossUS represents final loss information to be reduced or minimized for training of the second DNN 300. indicates. Also, a, b, c, and d may correspond to predetermined weights.

That is, the first DNN 800 updates parameters in a direction in which LossDS of Equation 1 is decreased, and the second DNN 300 updates parameters in a direction in which LossUS is decreased. When the parameters of the first DNN 800 are updated according to the LossDS derived in the training process, the first training image 1102 obtained based on the updated parameters is different from the first training image 1102 in the previous training process. and, accordingly, the third training image 1104 is also different from the third training image 1104 in the previous training process. When the third training image 1104 is different from the third training image 1104 in the previous training process, the quality loss information 1130 is also newly determined, and accordingly, the second DNN 300 updates the parameters. When the quality loss information 1130 is newly determined, since LossDS is also newly determined, the first DNN 800 updates parameters according to the newly determined LossDS. That is, the parameter update of the first DNN 800 causes the parameter update of the second DNN 300 , and the parameter update of the second DNN 300 causes the parameter update of the first DNN 800 . In other words, since the first DNN 800 and the second DNN 300 are jointly trained through sharing of the quality loss information 1130 , the parameters of the first DNN 800 and the parameters of the second DNN 300 are They can be optimized in relation to each other.

Referring to Equation 1, it can be seen that LossUS is determined according to quality loss information 1130, but this is an example, and LossUS is at least one of structural loss information 1110 and complexity loss information 1120, It may be determined based on the quality loss information 1130 .

Previously, it has been described that the AI setting unit 238 of the AI decoding apparatus 200 and the AI setting unit 718 of the AI encoding apparatus 700 store a plurality of DNN setting information, the AI setting unit 238 and the AI A method of training each of the plurality of DNN configuration information stored in the setting unit 718 will be described.

As described in relation to Equation 1, in the case of the first DNN 800, the degree of similarity between the structural information of the first training image 1102 and the structural information of the original training image 1101 (structural loss information 1110) ), the bit rate (complexity loss information 1120) of the image data obtained as a result of the first encoding of the first training image 1102 and the difference between the third training image 1104 and the original training image 1101 (quality loss) The parameter is updated in consideration of the information 1130).

In detail, the first training image 1102 similar to the structural information of the original training image 1101 and having a small bit rate of image data obtained when the first encoding is performed can be obtained, and at the same time, the first training image The parameters of the first DNN 800 may be updated so that the second DNN 300 that AI upscales 1102 may acquire a third training image 1104 similar to the original training image 1101 .

By adjusting the weights of a, b, and c in Equation 1, the directions in which the parameters of the first DNN 800 are optimized are different. For example, when the weight of b is determined to be high, the parameter of the first DNN 800 may be updated with more importance on lowering the bit rate than the quality of the third training image 1104 . In addition, when the weight of c is determined to be high, it is more important to increase the quality of the third training image 1104 than to increase the bit rate or to maintain the structural information of the original training image 1101. 1 A parameter of the DNN 800 may be updated.

Also, the direction in which parameters of the first DNN 800 are optimized may be different according to the type of a codec used to first encode the first training image 1102 . This is because, depending on the type of codec, the second training image to be input to the second DNN 300 may vary.

That is, the parameters of the first DNN 800 and the parameters of the second DNN 300 are linked based on the weight a, the weight b, the weight c, and the type of the codec for the first encoding of the first training image 1102 . so it can be updated. Therefore, when each of the weight a, the weight b, and the weight c is determined as a predetermined value and the type of the codec is determined as the predetermined type, the first DNN 800 and the second DNN 300 are trained, they are linked to each other. Optimized parameters of the first DNN 800 and parameters of the second DNN 300 may be determined.

Then, when the first DNN 800 and the second DNN 300 are trained after changing the weight a, the weight b, the weight c, and the type of the codec, the parameters of the first DNN 800 are optimized in connection with each other. and parameters of the second DNN 300 may be determined. In other words, when the first DNN 800 and the second DNN 300 are trained while changing the respective values of the weight a, the weight b, the weight c, and the type of the codec, the plurality of DNN setting information trained in connection with each other is transmitted to the first It may be determined from the DNN 800 and the second DNN 300 .

As described above with reference to FIG. 5 , the plurality of DNN configuration information of the first DNN 800 and the second DNN 300 may be mapped to the first image related information. To establish such a mapping relationship, a first training image 1102 output from the first DNN 800 is first encoded with a specific codec according to a specific bit rate, and a bitstream obtained as a result of the first encoding is first decoded. The obtained second training image may be input to the second DNN 300 . That is, after setting the environment so that the first training image 1102 of a specific resolution is first encoded at a specific bit rate by a specific codec, by training the first DNN 800 and the second DNN 300 , the first DNN setting mapped to the resolution of the training image 1102, the type of codec used for the first encoding of the first training image 1102, and the bitrate of the bitstream obtained as a result of the first encoding of the first training image 1102 The information pair can be determined. The resolution of the first training image 1102, the type of codec used for the first encoding of the first training image 1102, and the bitrate of the bitstream obtained according to the first encoding of the first training image 1102 are varied. , a mapping relationship between a plurality of DNN configuration information of the first DNN 800 and the second DNN 300 and the first image related information may be determined.

12 is a diagram for explaining a training process of the first DNN 800 and the second DNN 300 by the training device 1200 .

The training of the first DNN 800 and the second DNN 300 described with reference to FIG. 11 may be performed by the training apparatus 1200 . The training device 1200 includes a first DNN 800 and a second DNN 300 . The training device 1200 may be, for example, the AI encoding device 700 or a separate server. The DNN setting information of the second DNN 300 obtained as a result of training is stored in the AI decoding apparatus 200 .

Referring to FIG. 12 , the training apparatus 1200 initially sets DNN configuration information of the first DNN 800 and the second DNN 300 ( S1240 , S1245 ). Accordingly, the first DNN 800 and the second DNN 300 may operate according to predetermined DNN configuration information. The DNN setting information includes at least the number of convolution layers included in the first DNN 800 and the second DNN 300, the number of filter kernels for each convolutional layer, the size of filter kernels for each convolutional layer, and parameters of each filter kernel. It can contain information about one.

The training apparatus 1200 inputs the original training image 1101 into the first DNN 800 (S1250). The original training image 1101 may include at least one frame constituting a still image or a moving image.

The first DNN 800 processes the original training image 1101 according to the initially set DNN setting information, and outputs the AI downscaled first training image 1102 from the original training image 1101 ( S1255 ). 12 shows that the first training image 1102 output from the first DNN 800 is directly input to the second DNN 300 , but the first training image 1102 output from the first DNN 800 . ) may be input to the second DNN 300 by the training device 1200 . Also, the training apparatus 1200 may first encode and first decode the first training image 1102 using a predetermined codec, and then input the second training image to the second DNN 300 .

The second DNN 300 processes the first training image 1102 or the second training image according to the initially set DNN setting information, and the AI upscaled third from the first training image 1102 or the second training image. A training image 1104 is output (S1260).

The training apparatus 1200 calculates complexity loss information 1120 based on the first training image 1102 (S1265).

The training apparatus 1200 calculates structural loss information 1110 by comparing the reduced training image 1103 with the first training image 1102 ( S1270 ).

The training apparatus 1200 calculates quality loss information 1130 by comparing the original training image 1101 and the third training image 1104 ( S1275 ).

The first DNN 800 updates the initially set DNN configuration information through a back propagation process based on the final loss information (S1280). The training apparatus 1200 may calculate final loss information for training the first DNN 800 based on the complexity loss information 1120 , the structural loss information 1110 , and the quality loss information 1130 .

The second DNN 300 updates the initially set DNN configuration information through a reverse transcription process based on quality loss information or final loss information (S1285). The training apparatus 1200 may calculate final loss information for training of the second DNN 300 based on the quality loss information 1130 .

Thereafter, the training apparatus 1200, the first DNN 800, and the second DNN 300 update the DNN configuration information while repeating processes S1250 to S1285 until the final loss information is minimized. At this time, during each iteration process, the first DNN 800 and the second DNN 300 operate according to the DNN configuration information updated in the previous process.

Table 1 below shows the effects of AI encoding and AI decoding of the original image 105 and encoding and decoding of the original image 105 using HEVC according to an embodiment of the present disclosure.

[Table 1]

As can be seen from Table 1, the subjective picture quality when AI encoding and AI decoding of contents consisting of 300 frames of 8K resolution according to an embodiment of the present disclosure is higher than the subjective picture quality when encoding and decoding using HEVC. , it can be seen that the bit rate is reduced by more than 50%.

Meanwhile, the aforementioned AI encoding apparatus 700 processes the original image 105 as a first DNN to obtain a first image 115 , and the AI decoding apparatus 200 converts the second image 135 into a second DNN to obtain a third image 145 . Since the original image 105 and the second image 135, which are the objects of AI encoding and AI decoding, are directly input to the first DNN and the second DNN, the amount of calculations to be performed in the first DNN and the second DNN may increase. . For example, the first DNN needs to reduce the resolution of the original image 105 to acquire the first image 115 , and the second DNN requires the second image 135 to acquire the third image 145 . resolution should be increased. That is, the first DNN requires an operation to decrease the resolution of the original image 105 , and the second DNN requires an operation to increase the resolution of the second image 135 . These calculations may be omitted from the AI downscaling and AI upscaling processes based on preprocessed data, which will be described later.

In addition, since the first DNN and the second DNN described above both receive and process one image (that is, the original image 105 and the second image 135), the image to be subjected to AI downscale and AI upscale It is difficult to quickly identify the characteristics of Therefore, in order to perform AI downscaling and AI upscaling that accurately reflect image characteristics, a layer for image characteristic extraction must be included in the first DNN and the second DNN, which are included in the first DNN and the second DNN. This may cause an increase in the number of layers.

That is, according to the AI downscaling and AI upscaling process described above, the number of layers to be included in the first DNN and the second DNN increases and the amount of computation may be large, so that the first DNN and the second DNN are reduced in complexity. need to be implemented.

Hereinafter, AI downscaling and AI upscaling processes using the first DNN and the second DNN of low complexity will be described with reference to FIGS. 13 to 21 .

13 is a block diagram illustrating a configuration of an AI encoding apparatus according to another embodiment.

Referring to FIG. 13 , the AI encoding apparatus 1300 may include an AI encoding unit 1310 and a transmission unit 1330 . The AI encoder 1310 may include an AI downscaler 1312 , a first encoder 1314 , a data processor 1316 , and an AI setting unit 1318 .

The functions of the AI encoder 1310 and the transmitter 1330 may be the same as those of the AI encoder 710 and the transmitter 730 described above with reference to FIG. 7 . That is, the AI downscaler 1312 of the AI encoder 1310 AI downscales the original image 105 to obtain the first image 115 , and the first encoder 1314 performs the AI downscale of the first image 115 . ) can be first encoded. AI data related to AI downscaling and image data obtained as a result of the first encoding are transmitted to the data processing unit 1316 . The AI setting unit 1318 acquires DNN setting information corresponding to the downscale target among a plurality of DNN setting information and transmits it to the AI downscale unit 1312 .

The data processing unit 1316 generates AI-encoded data including image data and AI data and outputs the generated AI-encoded data to the transmitter 1330 , and the transmission unit 1330 outputs the AI-encoded data. Since the functions of the AI encoder 1310 and the transmitter 1330 have been described in detail with reference to FIG. 7 , a detailed description thereof will be omitted here.

Comparing the AI downscale unit 1312 with the AI downscale unit 712 shown in FIG. 7 , the AI downscale unit 1312 includes a preprocessor 1313 .

The preprocessor 1313 preprocesses the original image 105 and inputs data obtained as a result of the preprocessing into the first DNN.

In an embodiment, the data obtained as a result of the preprocessing may include a plurality of images having a resolution smaller than that of the original image 105 . The resolution of the plurality of images may be smaller than the resolution of the original image 105 and greater than the resolution of the first image 115 . Alternatively, the resolution of the plurality of images may be the same as the resolution of the first image 115 .

Since a plurality of images having a resolution smaller than that of the original image 105 are input to the first DNN, the amount of computation to be performed in the first DNN is reduced compared to a case in which a single original image 105 is input to the first DNN. can be reduced In other words, when the original image 105 is input to the first DNN, an operation for reducing the resolution of the original image 105 to that of the first image 115 is required, but a resolution smaller than that of the original image 105 is required. When the images having the images are input to the first DNN, the operation for changing the resolution of the image may not be necessary or the amount of computation may be reduced.

Since the number of layers included in the first DNN is proportional to the amount of computation, the number of layers included in the first DNN may be reduced as the amount of computation to be performed in the first DNN is reduced.

Some of the plurality of images obtained as a result of the preprocessing may include a feature map of the original image 105 . The feature map indicates unique characteristics of the original image 105 , for example, a vertical direction characteristic, a horizontal direction characteristic, or an edge characteristic. The first DNN quickly grasps the characteristics of the original image 105 by using the feature map input from the preprocessor 1313 without having to directly obtain a feature map representing the characteristics of the original image 105 from the original image 105 . can

That is, according to an embodiment, since a plurality of images partially having a feature map while having a resolution smaller than that of the original image 105 are input to the first DNN, the structure of the first DNN can be simplified.

The preprocessor 1313 may preprocess the original image 105 by a rule-based method (also referred to as a legacy method) instead of a neural network using parameters obtained as a result of training for calculation, which requires a large amount of computation. have. The reason for using the legacy method for preprocessing is that preprocessing the original image 105 based on a neural network may not have a significant difference in the amount of computation compared to directly inputting the original image 105 into the first DNN.

In an embodiment, the data obtained as a result of the preprocessing may include a first reduced image downscaled from the original image 105 and a reduced feature map corresponding to the original image 105 . The first reduced image may be obtained by downscaling the original image 105 using a legacy scale method. In one embodiment, the legacy scale method is a near neighbor scale method, a bilinear scale method, a bicubic scale method, a lanczos scale method, and a stair step method. It may include at least one of a scale method.

The resolution of the first reduced image and the reduced feature map is smaller than the resolution of the original image 105 . For example, the resolution of the first reduced image and the reduced feature map may be the same as the resolution of the first image 115 .

The first DNN processes the first reduced image output from the preprocessor 1313 and the reduced feature map to obtain the first image 115 . As described above, the first DNN may operate according to the DNN configuration information corresponding to the downscale target among the plurality of DNN configuration information.

Hereinafter, the configuration of the preprocessor 1313 and the structure of the first DNN for processing the preprocessed data will be described in detail.

14 is a diagram for explaining an AI downscaling process using the first DNN 2400 according to an embodiment.

The preprocessor 1313 includes a first downscaler 1410 , a second downscaler 1430 , and a subtraction unit 1450 for preprocessing the original image 105 .

Each of the first downscaler 1410 and the second downscaler 1430 may downscale the original image 105 according to a predetermined scaling method. The first downscaler 1410 and the second downscaler 1430 downscale the original image 105 using a rule-based scaling method (also referred to as a legacy scaling method).

In one embodiment, the legacy scale method is a near neighbor scale method, a bilinear scale method, a bicubic scale method, a lanczos scale method, and a stair step method. It may include at least one of a scale method.

The first downscaler 1410 and the second downscaler 1430 may downscale the original image 105 using different scaling methods. For example, the first downscaler 1410 downscales the original image 105 using a near-neighbor scale method, and the second downscaler 1430 uses a bicubic scale method for the original image. (105) can be downscaled.

The first downscaler 1410 downscales the original image 105 to obtain a first reduced image 1412 . The second downscaler 1430 downscales the original image 105 to obtain a second reduced image 1432 .

The subtractor 1450 obtains a residual image between the second reduced image 1432 and the first reduced image 1412 as the reduced feature map 1452 . The residual image may include difference values between pixel values of one of the first and second reduced images 1412 and 1432 and pixel values of the other image.

The first reduced image 1412 and the reduced feature map 1452 are concatenated 1470 and input to the first DNN 2400 . The sum of the number of first reduced images 1412 and the number of reduced feature maps 1452 should be equal to the number of input channels of the first layer of the first DNN 2400 . In FIG. 14 , 5 x 5 x 2n x k of the first convolutional layer 2410, which is the first layer of the first DNN 2400, is 2n with k filter kernels having a size of 5 x 5 in the first convolutional layer 2410. This means that k images are processed and k feature maps are output. where n and k are real numbers greater than zero.

Since the number of input channels of the first convolutional layer 2410 is 2n, the sum of the number of first reduced images 1412 and the number of reduced feature maps 1452 should also be 2n. That is, if the number of first reduced images 1412 is p, the number of reduced feature maps 1452 should be 2n-p.

In an embodiment, the first downscaler 1410 downscales the original image 105 using different p (p is a natural number) scale method to obtain p first reduced images 1412 . Then, the second downscaler 1430 downscales the original image 105 using q different scale methods (where q is a natural number) to obtain q second reduced images 1432 . The subtractor 1450 may acquire 2n-p residual images between the p first reduced images 1412 and q second reduced images 1432 . The p first reduced images 1412 and 2n-p residual images may be input to the first DNN 2400 .

In another embodiment, the first downscaler 1410 downscales the original image 105 using n different scaling methods to obtain n first reduced images 1412 . Then, the second downscaler 1430 downscales the original image 105 by any one of the scaling methods to obtain one second reduced image 1432 . The subtractor 1450 may obtain n residual images between each of the n first reduced images 1412 and one second reduced images 1432 . The n first reduced images 1412 and n residual images may be input to the first DNN 2400 .

In another embodiment, the first downscaler 1410 downscales the original image 105 using n different scale methods to obtain n first reduced images 1412 . In addition, the second downscaler 1430 downscales the original image 105 using n different scaling methods to obtain n second reduced images 1432 . The subtractor 1450 may acquire n residual images between the n first reduced images 1412 and n second reduced images 1432 . For example, the subtractor 1450 maps n first reduced images 1412 and n second reduced images 1432 on a one-to-one basis, and maps the first reduced images 1412 and the second reduced images 1412 to each other. Among the images 1432 , n residual images may be obtained. The n first reduced images 1412 and n residual images may be input to the first DNN 2400 .

In another embodiment, the first downscaler 1410 obtains n different first reduced images 1412 from the original image 105 according to the near-near scale method, and the second downscaler 1430 One second reduced image 1432 is obtained by downscaling the original image 105 by any one of the scaling methods. In addition, the subtractor 1450 may obtain n residual images between each of the n first reduced images 1412 and one second reduced images 1432 . The n first reduced images 1412 and n residual images may be input to the first DNN 2400 .

According to the near list neighbor scale method, the first downscaler 1410 determines pixel groups including n pixels in the original image 105 to obtain n first reduced images 1412, and each The n first reduced images 1412 including pixels located at different points in the pixel group may be acquired. Pixel groups including n pixels may have the form of a square block or a rectangular block.

For example, when the first downscaler 1410 is to acquire four first reduced images 1412 , the first downscaler 1410 includes four pixels adjacent to each other in the original image 105 . Determine pixel groups. Here, each pixel group including four pixels may have the shape of a square block. The first downscaler 1410 includes a first reduced image 1412 including pixels located on the upper left and a first reduced image 1412 including pixels located on the upper right among four pixels included in each of the pixel groups. ), a first reduced image 1412 including pixels located on the lower left side and a first reduced image 1412 including pixels located on the lower right side may be obtained.

As another example, when each pixel group including 4 pixels has the shape of a rectangular block, the first downscaler 1410 is a pixel located at the top (or the leftmost) among 4 pixels included in each of the pixel groups. A first reduced image 1412 consisting of The formed first reduced image 1412 and the first reduced image 1412 including pixels located at the lowermost (or rightmost) side may be obtained.

The preprocessor 1313 may determine the resolutions of the first reduced image 1412 and the second reduced image 1432 according to the target resolution of the first image 115 . For example, when the ratio between the target resolution of the first image 115 and the resolution of the original image 105 is 1/2, the preprocessor 1313 may perform the first reduced image 1412 and the second reduced image 1432 . ) may be determined to be 1/2 the resolution of the original image 105 . In addition, the first downscaler 1410 and the second downscaler 1430 obtain a first reduced image 1412 and a second reduced image 1432, each having a resolution reduced by 1/2 times from the original image 105 , respectively. can do.

As another example, when the ratio between the target resolution of the first image 115 and the resolution of the original image 105 is 1/4, the preprocessor 1313 may perform the first reduced image 1412 and the second reduced image 1432 . ) may be determined to be 1/2 the resolution of the original image 105 . In addition, the first downscaler 1410 and the second downscaler 1430 obtain a first reduced image 1412 and a second reduced image 1432, each having a resolution reduced by 1/2 times from the original image 105 , respectively. can do. In this case, the first DNN 2400 uses the first reduced image 1412 and the reduced feature map 1452 to obtain the first image 115 whose resolution is reduced by a factor of 1/4 compared to the original image 105 . An operation for reducing the resolution by 1/2 can be performed.

The first reduced image 1412 and the reduced feature map 1452 are input to the first convolutional layer 2410 . The first convolutional layer 2410 performs convolution processing on the first reduced image 1412 and the reduced feature map 1452 using k filter kernels having a size of 5×5. The k feature maps generated as a result of the convolution process are input to the first activation layer 2420 .

The first activation layer 2420 may provide a non-linear characteristic to the k feature maps. The first activation layer 2420 determines whether to transfer sample values of the feature maps output from the first convolution layer 2410 to the second convolution layer 2430 . For example, some sample values among the sample values of the feature maps are activated by the first activation layer 2420 and transmitted to the second convolution layer 2430 , and some sample values are activated by the first activation layer 2420 . It is deactivated and is not transmitted to the second convolutional layer 2430 . Information indicated by the feature maps output from the first convolutional layer 2410 is emphasized by the first activation layer 2420 .

The output of the first activation layer 2420 is input to the second convolution layer 2430 . The second convolution layer 2430 performs convolution processing on k feature maps using one filter kernel having a size of 5×5. As a result of the convolution process, one image is output from the second convolutional layer 2430 .

The first image 115 may be obtained by combining the second reduced image 1432 obtained by the second downscaler 1430 with the image output from the second convolutional layer 2430 . Here, the summing of the second reduced image 1432 and the images output from the second convolutional layer 2430 may mean that one image is obtained by adding pixel values of the two images to each other. Concatenating two images and concatenating two images have different meanings. Specifically, combining two images means that the two images are input to a layer together, and combining two images means that pixel values of the two images are added to each other to generate one image.

In an embodiment, when the resolution of the second reduced image 1432 and the resolution of the image output from the second convolutional layer 2430 are different from each other, in accordance with the resolution of the image output from the second convolution layer 2430 The resolution of the second reduced image 1432 may be scaled.

According to an embodiment, the second reduced image 1432 is not combined with the image output from the second convolution layer 2430 , and the image output from the second convolution layer 2430 is converted to the first image 115 . can be decided.

14 illustrates that the first DNN 2400 includes two convolutional layers 2410 and 2430 and one activation layer 2420, but this is only an example, and according to an implementation, The number of solution layers and activation layers may be variously changed. Also, according to an embodiment, the first DNN 2400 may be implemented through a recurrent neural network (RNN). That is, the CNN structure of the first DNN 2400 according to an example of the present disclosure may be changed to an RNN structure.

14 shows that the filter kernels of the convolutional layers 2410 and 2430 have a size of 5x5, but this is only an example, and the size of the filter kernels used in each convolutional layer may vary depending on the implementation. have.

As described above, since the first reduced image 1412 and the reduced feature map 1452 input to the first DNN 2400 have a resolution smaller than that of the original image 105, the first reduced image 1412 and the amount of computation in the first DNN 2400 for processing the reduced feature map 1452 may be reduced. Specifically, when the filter kernel moves according to a predetermined stride and a convolution operation with input data is performed, if the resolution of the input data is high, the number of convolution operations is inevitably increased. However, in an embodiment, since the first reduced image 1412 and the reduced feature map 1452 whose resolution is reduced through preprocessing of the original image 105 are input to the first DNN 2400, the number of convolution operations is reduced. can be

Also, according to an embodiment, in addition to the first reduced image 1412 reduced from the original image 105 , intrinsic characteristics of the original image 105 , for example, a vertical direction characteristic, a horizontal direction characteristic, or an edge characteristic Since the reduced feature map 1452 representing the etc. is input to the first DNN 2400 , training of the first DNN 2400 may be simplified compared to using one piece of information (ie, an original image). In addition, since various data representing characteristics of the original image 105 are input to the first DNN 2400 , the first DNN 2400 including a small number of layers maintains the sameness with the original image 105 . A first image 115 and a third image 145 may be acquired.

15 is a diagram for explaining an AI downscaling process using the first DNN 2500 according to another embodiment.

The preprocessor 1313 includes a first downscaler 1510 , a second downscaler 1530 , an upscaler 1550 , a third downscaler 1570 and a subtraction unit 1590 for preprocessing the original image 105 . may include.

The first downscaler 1510 may downscale the original image 105 according to a predetermined scaling method. The first downscaler 1510 downscales the original image 105 using a rule-based scaling method.

The second downscaler 1530 and the upscaler 1550 may downscale and upscale the original image 105 according to a predetermined scaling method.

The scale method used in the first downscaler 1510 and the scale method used in the second downscaler 1530 and the upscaler 1550 may be the same or different from each other.

A modified image 1552 is obtained from the original image 105 according to the downscale and upscale of the second downscaler 1530 and the upscaler 1550 . The resolution of the modified image 1552 may be the same as the resolution of the original image 105 . The second downscaler 1530 and the upscaler 1550 are for quality degradation of the original image 105 , and at least some pixel values in the original image 105 are generated by the second downscaler 1530 and the upscaler 1530 . It may be changed by the scaler 1550 .

The third downscaler 1570 may downscale the deformed image 1552 according to a predetermined scaling method. The third downscaler 1570 downscales the transformed image 1552 using a rule-based scaling method.

The first downscaler 1510 and the third downscaler 1570 may downscale the original image 105 and the transformed image 1552 by the same scaling method, respectively. For example, the first downscaler 1510 and the third downscaler 1570 may downscale the original image 105 and the transformed image 1552 using a near-neighbor scale method, respectively.

According to an embodiment, the first downscaler 1510 and the third downscaler 1570 may downscale the original image 105 and the transformed image 1552 by different scaling methods, respectively. For example, the first downscaler 1510 downscales the original image 105 using a near-neighbor scale method, and the third downscaler 1570 uses a bicubic scale method to downscale the transformed image. (1552) can be downscaled.

The subtraction unit 1590 reduces the residual image between the first reduced image 1512 obtained by the first downscaler 1510 and the second reduced image 1572 obtained by the third downscaler 1570 . Acquired as a map (1592). The residual image may include difference values between pixel values of one of the first and second reduced images 1512 and 1572 and pixel values of the other image.

The first reduced image 1512 and the reduced feature map 1592 are concatenated 1595 and input to the first DNN 2500 . The sum of the number of first reduced images 1512 and the number of reduced feature maps 1592 should be equal to the number of input channels of the first layer of the first DNN 2500 . In FIG. 15 , 5 x 5 x 2n x k of the first convolutional layer 2510 that is the first layer of the first DNN 2500 is 2n with k filter kernels having a size of 5 x 5 in the first convolutional layer 2510 . This means that k images are processed and k feature maps are output. where n and k are real numbers greater than zero.

Since the number of input channels of the first convolutional layer 2510 is 2n, the sum of the number of first reduced images 1512 and the number of reduced feature maps 1592 should also be 2n. That is, if the number of first reduced images 1512 is p, the number of reduced feature maps 1592 should be 2n-p.

In an embodiment, the first downscaler 1510 downscales the original image 105 using different p (p is a natural number) scaling method to obtain p first reduced images 1512 . In addition, the third downscaler 1570 downscales the transformed image 1552 using q different scale methods (where q is a natural number) to obtain q second reduced images 1572 . The subtractor 1590 may acquire 2n-p residual images between the p first reduced images 1512 and q second reduced images 1572 . The p first reduced images 1512 and 2n-p residual images may be input to the first DNN 2500 .

In another embodiment, the first downscaler 1510 downscales the original image 105 using n different scale methods to obtain n first reduced images 1512 . Then, the third downscaler 1570 downscales the transformed image 1552 by any one scaling method to obtain one second reduced image 1572 . The subtractor 1590 may obtain n residual images between each of the n first reduced images 1512 and one second reduced images 1572 . The n first reduced images 1512 and n residual images may be input to the first DNN 2500 .

In another embodiment, the first downscaler 1510 downscales the original image 105 using n different scale methods to obtain n first reduced images 1512 . Then, the third downscaler 1570 downscales the transformed image 1552 using n different scale methods to obtain n second reduced images 1572 . The subtractor 1590 may obtain n residual images between the n first reduced images 1512 and n second reduced images 1572 . For example, the subtractor 1590 maps n first reduced images 1512 and n second reduced images 1572 one-to-one, and maps the first reduced images 1512 and the second reduced images 1572 to each other. Among the images 1572 , n residual images may be obtained. The n first reduced images 1512 and n residual images may be input to the first DNN 2500 .

In another embodiment, the first downscaler 1510 obtains n different first reduced images 1512 from the original image 105 according to the near list neighbor scale method, and the third downscaler 1570 According to the near list neighbor scale method, n different second reduced images 1572 are obtained from the transformed image 1552 . In addition, the subtractor 1590 may obtain n residual images between the n first reduced images 1512 and n second reduced images 1572 . The n first reduced images 1512 and n residual images may be input to the first DNN 2500 . Since the near list neighbor scale method has been described with reference to FIG. 14 , a detailed description thereof will be omitted.

The preprocessor 1313 may determine the resolutions of the first reduced image 1512 and the second reduced image 1572 according to the target resolution of the first image 115 . For example, when the ratio between the target resolution of the first image 115 and the resolution of the original image 105 is 1/2, the preprocessor 1313 may perform the first reduced image 1512 and the second reduced image 1572 . ) may be determined to be 1/2 the resolution of the original image 105 . In addition, the first downscaler 1510 and the third downscaler 1570 acquire a first reduced image 1512 and a second reduced image 1572, each having a resolution reduced by 1/2 compared to the original image 105 , respectively. can do.

As another example, when the ratio between the target resolution of the first image 115 and the resolution of the original image 105 is 1/4, the preprocessor 1313 may perform the first reduced image 1512 and the second reduced image 1572 . ) may be determined to be 1/2 the resolution of the original image 105 . In addition, the first downscaler 1510 and the third downscaler 1570 acquire a first reduced image 1512 and a second reduced image 1572, each having a resolution reduced by 1/2 compared to the original image 105 , respectively. can do. In this case, in order to obtain the first image 115 whose resolution is reduced by a factor of 1/4 compared to the original image 105 , the first DNN 2500 performs the first reduced image 1512 and the reduced feature map 1592 . An operation for reducing the resolution by 1/2 can be performed.

The first reduced image 1512 and the reduced feature map 1592 are input to the first convolutional layer 2510 . The first convolutional layer 2510 performs convolution processing on the first reduced image 1512 and the reduced feature map 1592 using k filter kernels having a size of 5×5. The k feature maps generated as a result of the convolution process are input to the first activation layer 2520 .

The first activation layer 2520 may provide a non-linear characteristic to the k feature maps. An output of the first activation layer 2520 is input to the second convolution layer 2530 .

The second convolution layer 2530 performs convolution processing on k feature maps using one filter kernel having a size of 5×5. As a result of the convolution process, one image is output from the second convolutional layer 2530 .

The first image 115 may be obtained by combining the reduced image obtained by the second downscaler 1530 with an image output from the second convolution layer 2530 . In an embodiment, when the resolution of the reduced image obtained by the second downscaler 1530 and the resolution of the image output from the second convolution layer 2530 are different from each other, output from the second convolution layer 2530 The resolution of the reduced image obtained by the second downscaler 1530 may be scaled according to the resolution of the converted image. According to an embodiment, the second downscaler 1530 may downscale the original image 105 to match the resolution of the first image 115 . In this case, the upscaler 1550 may upscale the reduced image acquired by the second downscaler 1530 according to the resolution of the original image 105 .

According to an embodiment, the reduced image obtained by the second downscaler 1530 is not combined with the image output from the second convolution layer 2530, and the image output from the second convolution layer 2530 is One image 115 may be determined.

15 shows that the first DNN 2500 includes two convolutional layers 2510 and 2530 and one activation layer 2520, but this is only an example, and according to an implementation, The number of solution layers and activation layers may be variously changed. Also, according to an implementation, the first DNN 2500 may be implemented through an RNN. That is, the CNN structure of the first DNN 2500 according to an example of the present disclosure may be changed to an RNN structure.

15 shows that the size of the filter kernel of the convolutional layers 2510 and 2530 is 5x5, this is only an example, and the size of the filter kernel used in each convolutional layer may vary depending on the implementation. have.

16 is a diagram for explaining an AI downscaling process using the first DNN 2600 according to another embodiment.

The preprocessor 1313 may include a first downscaler 1610 , a feature map obtainer 1630 , and a second downscaler 1650 for preprocessing the original image 105 . Depending on the embodiment, the second downscaler 1650 may not be included in the preprocessor 1313 .

The first downscaler 1610 may downscale the original image 105 according to a predetermined scaling method. The first downscaler 1610 downscales the original image 105 using a rule-based scaling method.

The feature map obtainer 1630 generates a reduced feature map 1632 from the original image 105 according to a predetermined feature map extraction algorithm. When the image generated according to the feature map extraction algorithm is different from the resolution of the first reduced image 1612 acquired by the first downscaler 1610, the feature map acquisition unit 1630 is generated according to the feature map extraction algorithm. The resolution of the image may be scaled according to the resolution of the first reduced image 1612 .

In an embodiment, the feature map acquirer 1630 may acquire the edge map generated according to the edge detection algorithm as the reduced feature map 1632 . The edge detection algorithm may include, but is not limited to, a Sobel algorithm, a prewitt algorithm, a Roberts algorithm, a compass algorithm, or a canny algorithm.

The edge map shows the edge characteristics of the original image 105, for example, the amount of brightness change in the vertical direction, the amount of brightness change in the left and right directions, etc., so this edge map is the reduced feature map 1632 and the first DNN 2600 As input to , AI downscaling in consideration of image characteristics can be performed more quickly.

In an embodiment, the feature map obtainer 1630 may obtain the reduced feature map 1632 by processing the original image 105 with a neural network. For example, the feature map acquisition unit 1630 may acquire the reduced feature map 1632 using a neural network such as VGGnet.

The first reduced image 1612 and the reduced feature map 1632 obtained by the feature map obtainer 1630 are concatenated 1670 and input to the first DNN 2600 .

The sum of the number of first reduced images 1612 and the number of reduced feature maps 1632 should be equal to the number of input channels of the first layer of the first DNN 2600 . In FIG. 16 , 5 x 5 x 2n x k of the first convolutional layer 2610 that is the first layer of the first DNN 2600 is 2n with k filter kernels having a size of 5 x 5 in the first convolutional layer 2610 . This means that k images are processed and k feature maps are output. where n and k are real numbers greater than zero.

Since the number of input channels of the first convolutional layer 2610 is 2n, the sum of the number of first reduced images 1612 and the number of reduced feature maps 1632 should also be 2n. That is, if the number of first reduced images 1612 is p, the number of reduced feature maps 1632 should be 2n-p.

In an embodiment, the first downscaler 1610 may acquire p first reduced images 1612 by downscaling the original image 105 using different p scale methods (p is a natural number). The feature map acquisition unit 1630 may acquire 2n-p reduced feature maps 1632 according to the 2n-p feature map extraction algorithm.

In another embodiment, the first downscaler 1610 may acquire different p number of first reduced images 1612 from the original image 105 according to the near list neighbor scale method. The feature map acquisition unit 1630 may acquire 2n-p reduced feature maps 1632 according to the 2n-p feature map extraction algorithm.

The preprocessor 1313 may determine the resolution of the first reduced image 1612 and the reduced feature map 1632 according to the target resolution of the first image 115 .

For example, when the ratio between the target resolution of the first image 115 and the resolution of the original image 105 is 1/2, the preprocessor 1313 may perform the first reduced image 1612 and the reduced feature map 1632 . may be determined to be 1/2 the resolution of the original image 105 . Then, the first downscaler 1610 and the feature map acquisition unit 1630 obtain the first reduced image 1612 and the reduced feature map 1632, each having a resolution reduced by 1/2 compared to the original image 105, respectively. can

As another example, when the ratio between the target resolution of the first image 115 and the resolution of the original image 105 is 1/4, the preprocessor 1313 may perform the first reduced image 1612 and the reduced feature map 1632 . may be determined to be 1/2 the resolution of the original image 105 . In addition, the first downscaler 1610 and the feature map acquisition unit 1630 obtain the first reduced image 1612 and the reduced feature map 1632, each having a resolution reduced by a factor of 1/2 from the original image 105 , respectively. can In this case, the first DNN 2600 performs the first reduced image 1612 and reduced feature map 1632 of the first reduced image 1612 and the reduced feature map 1632 to obtain the first image 115 whose resolution is reduced by a factor of 1/4 compared to the original image 105 . An operation for reducing the resolution by 1/2 can be performed.

The first reduced image 1612 and the reduced feature map 1632 are input to the first convolutional layer 2610 . The first convolutional layer 2610 performs convolution processing on the first reduced image 1612 and the reduced feature map 1632 using k filter kernels having a size of 5×5. The k feature maps generated as a result of the convolution process are input to the first activation layer 2620 .

The first activation layer 2620 may provide a non-linear characteristic to the k feature maps. The output of the first activation layer 2620 is input to the second convolution layer 2630 .

The second convolution layer 2630 performs convolution processing on k feature maps using one filter kernel having a size of 5×5. As a result of the convolution process, one image is output from the second convolutional layer 2630 . An image output from the second convolutional layer 2630 may be determined as the first image 115 .

As shown in FIG. 16 , the first image 115 is obtained by adding the image output from the second convolutional layer 2630 and the second reduced image 1652 obtained by the second downscaler 1650 . can The second downscaler 2630 may acquire the second reduced image 1650 by legacy downscaling the original image.

16 shows that the first DNN 2600 includes two convolutional layers 2610 and 2630 and one activation layer 2620, but this is only an example, and according to an implementation, The number of solution layers and activation layers may be variously changed. Also, according to an implementation, the first DNN 2600 may be implemented through an RNN. That is, the CNN structure of the first DNN 2600 according to an example of the present disclosure may be changed to an RNN structure.

Although FIG. 16 shows that the size of the filter kernel of the convolutional layers 2610 and 2630 is 5x5, this is only an example, and the size of the filter kernel used in each convolutional layer may vary depending on the implementation. have.

17 is an exemplary diagram illustrating a first DNN 2700 according to another embodiment.

The first DNN 2700 may replace the first DNNs 2400 , 2500 , and 2600 described above with reference to FIGS. 14 to 16 . The first DNN 2700 may include a plurality of convolutional layers 2710 , 2730 , and 2750 and a plurality of activation layers 2720 and 2740 , but the number of convolutional layers and activation layers may vary according to implementations. can be decided.

Referring to FIG. 17 , it can be seen that output data of any one convolutional layer 2710 or 2730 is combined with output data and input data of a previous layer and transferred to the next layer.

Specifically, the output of the first convolutional layer 2710 is combined with input data and transmitted to the first activation layer 2720 , and the output of the second convolution layer 2730 is the output of the first convolutional layer 2710 . and input data and transferred to the second activation layer 2740 . Also, the output of the third convolutional layer 2750 may be determined as output data of the first DNN 2700 . As described above, the output data of the first DNN 2700 is combined with the reduced images obtained by the second downscalers 1430, 1530, and 1650 of FIGS. 14 to 16 to obtain a first image 115. have.

Referring to FIG. 17 , the number of input channels and the number of output channels (the number of filter kernels) of the first convolution layer 2710 and the second convolution layer 2730 are both 2n, and the third convolution layer 2750 . The number of input channels is 2n. When 2n pieces of output data and 2n pieces of input data of the first convolution layer 2710 are summed, output data and element values of the input data for each channel corresponding to each other, that is, pixel values or sample values may be added. And, when 2n pieces of output data of the second convolutional layer 2730, 2n pieces of output data of the first convolutional layer 2710, and 2n pieces of input data are combined, output data and input data for each channel corresponding to each other are combined. Element values, ie, pixel values, may be added to each other.

The reason for adding the output data and input data of the previous layer to the output data of one layer is to reduce the number of input channels and the number of output channels to be processed in each layer. The reduction in the number of input channels of each convolutional layer means that the number of images or feature maps to be subjected to the convolution process is reduced, so that the amount of computation for the convolution process may be reduced.

18 is a block diagram illustrating a configuration of an AI decoding apparatus 1800 according to another embodiment.

Referring to FIG. 18 , the AI decoding apparatus 1800 may include a receiving unit 1810 and an AI decoding unit 1830 . The AI decoding unit 1830 may include a parsing unit 1832 , a first decoding unit 1834 , an AI upscaling unit 1836 , and an AI setting unit 1838 .

The functions of the receiving unit 1810 and the AI decoding unit 1830 may be the same as the functions of the receiving unit 210 and the AI decoding unit 230 described above with reference to FIG. 2 . That is, the receiver 1810 receives the AI encoded data and transmits it to the AI decoder 1830 . The parsing unit 1832 of the AI decoding unit 1830 parses the AI-encoded data, outputs image data included in the AI-encoded data to the first decoder 1834 , and outputs the AI data to the AI setting unit 1838 . do. The first decoder 1834 first decodes the image data to obtain a second image 135 , and transmits the second image 135 to the AI upscaler 1836 . The AI setting unit 1838 transmits DNN setting information obtained from among a plurality of DNN setting information based on the AI data to the AI upscaling unit 1836 . The AI upscaling unit 1836 acquires the third image 145 by AI upscaling the second image 135 according to the DNN setting information. Since the functions of the receiver 1810 and the AI decoder 1830 have been described in detail with reference to FIG. 2 , a detailed description thereof will be omitted here.

Compared with the AI upscaling unit 1836 and the AI upscaling unit 236 shown in FIG. 2 , the AI upscaling unit 1836 includes a preprocessing unit 1837 .

The preprocessor 1837 preprocesses the second image 135 and inputs data obtained as a result of the preprocessing into the second DNN.

In an embodiment, data obtained as a result of the preprocessing may include a plurality of images having a resolution greater than that of the second image 135 . The resolution of the plurality of images may be greater than the resolution of the second image 135 and smaller than the resolution of the third image 145 . Alternatively, the resolution of the plurality of images may be the same as the resolution of the third image 145 .

Since a plurality of images having a resolution greater than that of the second image 135 are input to the second DNN, compared to a case in which a single second image 135 is input to the second DNN, the second DNN should be performed. The amount of computation can be reduced. In other words, when the second image 135 is input to the second DNN, an operation of increasing the resolution of the second image 135 to that of the third image 145 is required, but When images having a large resolution are input to the second DNN, the calculation for increasing the resolution of the image may be eliminated or the amount of calculation may be reduced.

Since the number of layers included in the second DNN is proportional to the amount of computation, the number of layers included in the second DNN may be reduced as the amount of computation to be performed in the second DNN is reduced.

Some of the plurality of images obtained as a result of the preprocessing may include a feature map of the second image 135 . The feature map indicates unique characteristics of the second image 135 , for example, a vertical direction characteristic, a horizontal direction characteristic, or an edge characteristic. The second DNN uses the feature map input from the preprocessor 1837 without having to directly obtain a feature map representing the characteristics of the second image 135 from the second image 135 to quickly generate the second image 135 . characteristics can be identified.

That is, according to an embodiment, since a plurality of images having a resolution greater than that of the second image 135 and a part of which are feature maps are input to the second DNN, the structure of the second DNN can be simplified.

The preprocessor 1837 may preprocess the second image 135 by a rule-based method instead of a neural network requiring a large amount of computation.

In an embodiment, data obtained as a result of the preprocessing may include a first enlarged image upscaled from the second image 135 and an enlarged feature map corresponding to the second image 135 . The first enlarged image may be obtained by upscaling the second image using a legacy scale method. In one embodiment, the legacy scale method is a near neighbor scale method, a bilinear scale method, a bicubic scale method, a lanczos scale method, and a stair step method. It may include at least one of a scale method.

The resolution of the first enlarged image and the enlarged feature map is greater than the resolution of the second image 135 . For example, the resolution of the first enlarged image and the enlarged feature map may be the same as the resolution of the third image 145 .

The second DNN processes the first enlarged image and the enlarged feature map output from the preprocessor 1837 to obtain the third image 145 . As described above, the second DNN may operate according to DNN configuration information selected based on AI data among a plurality of DNN configuration information.

Hereinafter, the configuration of the preprocessor 1837 and the structure of the second DNN for processing the preprocessed data will be described in detail.

19 is a diagram for explaining an AI upscaling process using the second DNN 2900 according to an embodiment.

The preprocessor 1837 includes a first upscaler 1910 , a second upscaler 1930 , and a subtraction unit 1950 for preprocessing the second image 135 .

Each of the first upscaler 1910 and the second upscaler 1930 may upscale the second image 135 according to a predetermined scaling method. The first upscaler 1910 and the second upscaler 1930 upscale the second image 135 using a rule-based scaling method.

In one embodiment, the rule-based scale method includes a near neighbor scale method, a bilinear scale method, a bicubic scale method, a lanczos scale method, and a stair step method. step) may include at least one of a scale method.

The first upscaler 1910 and the second upscaler 1930 may upscale the second image 135 using different scaling methods. For example, the first upscaler 1910 upscales the second image 135 using a near-neighbor scale method, and the second upscaler 1930 uses a bicubic scale method. 2 The image 135 may be upscaled.

The first upscaler 1910 acquires a first enlarged image 1912 by upscaling the second image 135 . The second upscaler 1930 upscales the second image 135 to obtain a second enlarged image 1932 .

The subtractor 1950 obtains a residual image between the second enlarged image 1932 and the first enlarged image 1912 as the enlarged feature map 1952 . The residual image may include difference values between pixel values of one of the first and second enlarged images 1912 and 1932 and pixel values of the other image.

The first enlarged image 1912 and the enlarged feature map 1952 are concatenated 1970 and input to the second DNN 2900 .

The sum of the number of first enlarged images 1912 and the number of enlarged feature maps 1952 should be equal to the number of input channels of the first layer of the second DNN 2900 . In FIG. 19 , 3 x 3 x 2n x k of the first convolutional layer 2910, which is the first layer of the second DNN 2900, is 2n with k filter kernels having a size of 3 x 3 in the first convolutional layer 2910. This means that k images are processed and k feature maps are output. where n and k are real numbers greater than zero.

Since the number of input channels of the first convolutional layer 2910 is 2n, the sum of the number of first enlarged images 1912 and the number of enlarged feature maps 1952 must also be 2n. That is, if the number of first enlarged images 1912 is p, the number of enlarged feature maps 1952 should be 2n-p.

In an embodiment, the first upscaler 1910 acquires p first enlarged images 1912 by upscaling the second image 135 using different p scale methods (p is a natural number). In addition, the second upscaler 1930 upscales the second image 135 using q different scale methods (where q is a natural number) to obtain q second enlarged images 1932 . The subtractor 1950 may acquire 2n-p residual images between the p first magnified images 1912 and q second magnified images 1932 . The p first enlarged images 1912 and 2n-p residual images may be input to the second DNN 2900 .

In another embodiment, the first upscaler 1910 acquires n first enlarged images 1912 by upscaling the second image 135 using n different scale methods. In addition, the second upscaler 1930 upscales the second image 135 by any one scaling method to obtain one second enlarged image 1932 . The subtractor 1950 may acquire n residual images between each of the n first magnified images 1912 and one second magnified image 1932 . The n first enlarged images 1912 and n residual images may be input to the second DNN 2900 .

In another embodiment, the first upscaler 1910 acquires n first enlarged images 1912 by upscaling the second image 135 using n different scale methods. In addition, the second upscaler 1930 upscales the second image 135 using n different scale methods to obtain n second enlarged images 1932 . The subtractor 1950 may acquire n residual images between the n first magnified images 1912 and n second magnified images 1932 . The n first enlarged images 1912 and n residual images may be input to the second DNN 2900 .

In another embodiment, the first upscaler 1910 obtains n different first enlarged images 1912 from the second image 135 according to the near-near scale method, and the second upscaler 1930 to acquire one second enlarged image 1932 by upscaling the second image 135 by any one of the scaling methods. Also, the subtractor 1950 may acquire n residual images between each of the n first magnified images 1912 and one second magnified image 1932 . The n first enlarged images 1912 and n residual images may be input to the second DNN 2900 .

According to the near list neighbor scale method, the first upscaler 1910 generates new pixels centering on each pixel included in the second image 135 to obtain n first enlarged images 1912, and , a first enlarged image 1912 including pixels included in the second image 135 and newly generated pixels may be acquired. In this case, in order to obtain n different first enlarged images 1912 , the first upscaler 1910 applies each pixel included in the second image 135 to different equations to generate new pixels. can create

The preprocessor 1837 may determine the resolutions of the first enlarged image 1912 and the second enlarged image 1932 according to the target resolution of the third image 145 . For example, when the target resolution of the third image 145 is twice the resolution of the second image 135 , the preprocessor 1837 may adjust the resolutions of the first enlarged image 1912 and the second enlarged image 1932 . It may be determined to be twice the size of the second image 135 . In addition, the first upscaler 1910 and the second upscaler 1930 each acquire a first enlarged image 1912 and a second enlarged image 1932 whose resolution is increased by two times from the second image 135. can

As another example, when the target resolution of the third image 145 is four times the resolution of the second image 135 , the preprocessor 1837 may set the resolutions of the first enlarged image 1912 and the second enlarged image 1932 . It may be determined to be twice the size of the second image 135 . In addition, the first upscaler 1910 and the second upscaler 1930 each acquire a first enlarged image 1912 and a second enlarged image 1932 whose resolution is increased by two times from the second image 135. can In this case, the second DNN 2900 performs the resolution of the first enlarged image 1912 and the enlarged feature map 1952 in order to obtain the third image 145 with an increased resolution of 4 times compared to the second image 135 . It is possible to perform an operation that increases by 2 times.

The first enlarged image 1912 and the enlarged feature map 1952 are input to the first convolutional layer 2910 . The first convolutional layer 2910 performs convolution processing on the first magnified image 1912 and the magnified feature map 1952 using k filter kernels having a size of 3×3. The k feature maps generated as a result of the convolution process are input to the first activation layer 2920 .

The first activation layer 2920 may provide a non-linear characteristic to the k feature maps. The first activation layer 2920 determines whether to transfer sample values of the feature maps output from the first convolution layer 2910 to the second convolution layer 2930 . For example, some sample values among the sample values of the feature maps are activated by the first activation layer 2920 and transmitted to the second convolution layer 2930 , and some sample values are activated by the first activation layer 2920 . It is deactivated and is not transmitted to the second convolutional layer 2930 . Information indicated by the feature maps output from the first convolutional layer 2910 is emphasized by the first activation layer 2920 .

The output of the first activation layer 2920 is input to the second convolution layer 2930 . The second convolution layer 2930 performs convolution processing on k feature maps using one filter kernel having a size of 3×3. As a result of the convolution processing, one image is output from the second convolutional layer 2930 .

A third image 145 is obtained by combining the second enlarged image 1932 obtained by the second upscaler 1930 with the image output from the second convolutional layer 2930 . Here, the summing of the second enlarged image 1932 and the images output from the second convolutional layer 2930 may mean that one image is obtained by adding pixel values of the two images to each other. In an embodiment, when the resolution of the second enlarged image 1932 and the resolution of the image output from the second convolutional layer 2930 are different from each other, in accordance with the resolution of the image output from the second convolutional layer 2930 The resolution of the second enlarged image 1932 may be scaled.

According to an embodiment, the second enlarged image 1932 is not combined with the image output from the second convolution layer 2930 , and the image output from the second convolution layer 2930 is converted to the third image 145 . may be decided.

19 illustrates that the second DNN 2900 includes two convolutional layers 2910 and 2930 and one activation layer 2920, but this is only an example, and according to an implementation, The number of solution layers and activation layers may be variously changed. Also, depending on the implementation, the second DNN 2900 may be implemented through a recurrent neural network (RNN). That is, the CNN structure of the second DNN 2900 according to an example of the present disclosure may be changed to an RNN structure.

19 shows that the size of the filter kernel of the convolutional layers 2910 and 2930 is 3x3, this is only an example, and the size of the filter kernel used in each convolutional layer may vary depending on the implementation. have.

20 is a diagram for explaining an AI upscaling process using the second DNN 3000 according to another embodiment.

The preprocessor 1837 includes a first upscaler 2010, a second upscaler 2030, a downscaler 2050, a third upscaler 2070, and a subtraction unit 2090 for preprocessing the second image 135. ) may be included.

The first upscaler 2010 may upscale the second image 135 according to a predetermined scaling method. The first upscaler 2010 upscales the second image 135 using a rule-based scaling method.

The second upscaler 2030 and the downscaler 2050 may upscale and downscale the second image 135 according to a predetermined scaling method.

The scale method used in the first upscaler 2010 and the scale method used in the second upscaler 2030 and the downscaler 2050 may be the same or different from each other.

A transformed image 2052 is obtained from the second image 135 according to the upscaling and downscaling of the second upscaler 2030 and the downscaler 2050 . The resolution of the deformed image 2052 may be the same as the resolution of the second image 135 . The second upscaler 2030 and the downscaler 2050 are for quality degradation of the second image 135 , and at least some pixel values in the second image 135 are determined by the second upscaler 2030 . and the downscaler 2050.

The third upscaler 2070 may upscale the deformed image 2052 according to a predetermined scaling method. The third upscaler 2070 upscales the transformed image 2052 using a rule-based scaling method.

The first upscaler 2010 and the third upscaler 2070 may respectively upscale the second image 135 and the transformed image 2052 by the same scaling method. For example, the first upscaler 2010 and the third upscaler 2070 may respectively upscale the second image 135 and the transformed image 2052 by using a near-neighbor scale method.

According to an embodiment, the first upscaler 2010 and the third upscaler 2070 may respectively upscale the second image 135 and the transformed image 2052 using different scaling methods. For example, the first upscaler 2010 upscales the second image 135 by a near neighbor scale method, and the third upscaler 2070 transforms it to a bicubic scale method The image 2052 may be upscaled.

The subtraction unit 2090 enlarges the residual image between the first enlarged image 2012 acquired by the first upscaler 2010 and the second enlarged image 2072 acquired by the third upscaler 2070 . It is obtained as a map 2092 . The residual image may be formed of difference values between pixel values of one of the first and second enlarged images 2012 and 2072 and pixel values of the other image.

The first enlarged image 2012 and the enlarged feature map 2092 are concatenated 2095 and input to the second DNN 3000 .

The sum of the number of first enlarged images 2012 and the number of enlarged feature maps 2092 should be equal to the number of input channels of the first layer of the second DNN 3000 . In FIG. 20 , 3 x 3 x 2n x k of the first convolutional layer 3010, which is the first layer of the second DNN 3000, is 2n with k filter kernels having a size of 3 x 3 in the first convolutional layer 3010. This means that k images are processed and k feature maps are output. where n and k are real numbers greater than zero.

Since the number of input channels of the first convolutional layer 3010 is 2n, the sum of the number of first enlarged images 2012 and the number of enlarged feature maps 2092 should also be 2n. That is, if the number of first enlarged images 2012 is p, the number of enlarged feature maps 2092 should be 2n-p.

In an embodiment, the first upscaler 2010 acquires p first enlarged images 2012 by upscaling the second image 135 using different p (p is a natural number) scale method. Then, the third upscaler 2070 upscales the transformed image 2052 using q different scale methods (where q is a natural number) to obtain q second enlarged images 2072 . The subtractor 2090 may acquire 2n-p residual images between the p first magnified images 2012 and q second magnified images 2072 . The p first enlarged images 2012 and 2n-p residual images may be input to the second DNN 3000 .

In another embodiment, the first upscaler 2010 acquires n first enlarged images 2012 by upscaling the second image 135 using n different scale methods. In addition, the third upscaler 2070 upscales the deformed image 2052 by any one scaling method to obtain one second enlarged image 2072 . The subtractor 2090 may obtain n residual images between each of the n first magnified images 2012 and one second magnified image 2072 . The n first enlarged images 2012 and n residual images may be input to the second DNN 3000 .

In another embodiment, the first upscaler 2010 acquires n first enlarged images 2012 by upscaling the second image 135 using n different scale methods. In addition, the third upscaler 2070 upscales the transformed image 2052 using n different scale methods to obtain n second enlarged images 2072 . The subtractor 2090 may obtain n residual images between the n first magnified images 2012 and the n second magnified images 2072 . The n first enlarged images 2012 and n residual images may be input to the second DNN 3000 .

In another embodiment, the first upscaler 2010 obtains n different first magnified images 2012 from the second image 135 according to the near list neighbor scale method, and the third upscaler 2070 obtains n different second enlarged images 2072 from the deformed image 2052 according to the near list neighbor scale method. In addition, the subtractor 2090 may acquire n residual images between the n first magnified images 2012 and the n second magnified images 2072 . For example, the subtractor 2090 maps the n first magnified images 2012 and the n second magnified images 2072 one-to-one, and maps the first magnified images 2012 and the second magnified images to each other. Among the images 2072 , n residual images may be obtained. The n first enlarged images 2012 and n residual images may be input to the second DNN 3000 . Since the near list neighbor scale method has been described with reference to FIG. 19 , a detailed description thereof will be omitted.

The preprocessor 1837 may determine the resolutions of the first enlarged image 2012 and the second enlarged image 2072 according to the target resolution of the third image 145 . For example, when the target resolution of the third image 145 is twice the resolution of the second image 135 , the preprocessor 1837 may adjust the resolutions of the first enlarged image 2012 and the second enlarged image 2072 . It may be determined to be twice the size of the second image 135 . In addition, the first upscaler 2010 and the third upscaler 2070 are the first enlarged image 2012 and the second enlarged image in which the resolution is increased by two times from the second image 135 and the modified image 2052 , respectively. (2072) can be obtained.

As another example, when the target resolution of the third image 145 is 4 times the resolution of the second image 135 , the preprocessor 1837 may adjust the resolutions of the first enlarged image 2012 and the second enlarged image 2072 . It may be determined to be twice the size of the second image 135 and the modified image 2052 . In addition, the first upscaler 2010 and the third upscaler 2070 are the first enlarged image 2012 and the second enlarged image in which the resolution is increased by two times from the second image 135 and the modified image 2052 , respectively. (2072) can be obtained. In this case, the second DNN 3000 performs the resolution of the first magnified image 2012 and the magnified feature map 2092 to obtain the third image 145 with an increased resolution of 4 times compared to the second image 135 . It is possible to perform an operation to increase by 2 times.

The first enlarged image 2012 and the enlarged feature map 2092 are input to the first convolutional layer 3010 . The first convolution layer 3010 performs convolution processing on the first magnified image 2012 and the magnified feature map 2092 using k filter kernels having a size of 3×3. The k feature maps generated as a result of the convolution process are input to the first activation layer 3020 .

The first activation layer 3020 may provide non-linear characteristics to k feature maps.

An output of the first activation layer 3020 is input to the second convolution layer 3030 . The second convolution layer 3030 performs convolution processing on k feature maps using one filter kernel having a size of 3×3. As a result of the convolution process, one image is output from the second convolutional layer 3030 .

A third image 145 may be obtained by combining the enlarged image obtained by the second upscaler 2030 with an image output from the second convolutional layer 3030 . In an embodiment, when the resolution of the enlarged image obtained by the second upscaler 2030 and the resolution of the image output from the second convolution layer 3030 are different from each other, output from the second convolution layer 3030 The resolution of the enlarged image acquired by the second upscaler 2030 may be scaled according to the resolution of the converted image. Alternatively, the second upscaler 2030 may upscale the second image 135 to match the resolution of the third image 145 . In this case, the downscaler 2050 may downscale the enlarged image acquired by the second upscaler 2030 according to the resolution of the second image 135 .

According to an embodiment, the enlarged image obtained by the second upscaler 2030 is not combined with the image output from the second convolution layer 3030, and the image output from the second convolution layer 3030 is 3 images 145 may be determined.

20 illustrates that the second DNN 3000 includes two convolutional layers 3010 and 3030 and one activation layer 3020, but this is only an example, and according to an implementation, The number of solution layers and activation layers may be variously changed. Also, according to an implementation, the second DNN 3000 may be implemented through an RNN. That is, the CNN structure of the second DNN 3000 according to an example of the present disclosure may be changed to an RNN structure.

Although FIG. 20 shows that the size of the filter kernel of the convolutional layers 3010 and 3030 is 3x3, this is only an example, and the size of the filter kernel used in each convolutional layer may vary depending on the implementation. have.

21 is a diagram for explaining an AI upscaling process using the second DNN 3100 according to another embodiment.

The preprocessor 1837 may include a first upscaler 2110 , a feature map acquirer 2130 , and a second upscaler 2150 for preprocessing the second image 135 . Depending on the embodiment, the second upscaler 2150 may not be included in the preprocessor 1837 .

The first upscaler 2110 may upscale the second image 135 according to a predetermined scaling method. The first upscaler 2110 upscales the second image 135 using a rule-based scaling method.

The feature map acquisition unit 2130 generates an enlarged feature map 2132 from the second image 135 according to a predetermined feature map extraction algorithm. When the image generated according to the feature map extraction algorithm is different from the resolution of the first enlarged image 2112 , the feature map acquirer 2130 sets the resolution of the image generated according to the feature map extraction algorithm to the first enlarged image 2112 . can be scaled according to the resolution of

In an embodiment, the feature map acquisition unit 2130 may acquire an edge map generated according to an edge detection algorithm as the enlarged feature map 2132 . The edge detection algorithm may include, but is not limited to, a Sobel algorithm, a prewitt algorithm, a Roberts algorithm, a compass algorithm, or a canny algorithm. Since the edge map shows the edge characteristics of the second image 135 well, for example, the amount of brightness change in the vertical direction, the amount of brightness change in the left and right directions, etc., this edge map is used as the enlarged feature map 2132 for the second DNN 3100 ), AI upscaling in consideration of image characteristics can be performed more quickly.

In an embodiment, the feature map acquisition unit 2130 may obtain the enlarged feature map 2132 by processing the second image 135 with a neural network. For example, the feature map acquisition unit 2130 may acquire the enlarged feature map 2132 using a neural network such as VGGnet.

The first enlarged image 2112 and the enlarged feature map 2132 are concatenated 2170 and input to the second DNN 3100 .

The sum of the number of first enlarged images 2112 and the number of enlarged feature maps 2132 should be equal to the number of input channels of the first layer of the second DNN 3100 . In FIG. 21 , 3 x 3 x 2n x k of the first convolutional layer 3110, which is the first layer of the second DNN 3100, is 2n with k filter kernels having a size of 3 x 3 in the first convolutional layer 3110. This means that k images are processed and k feature maps are output. where n and k are real numbers greater than zero.

Since the number of input channels of the first convolutional layer 3110 is 2n, the sum of the number of first enlarged images 2112 and the number of enlarged feature maps 2132 should also be 2n. That is, if the number of first enlarged images 2112 is p, the number of enlarged feature maps 2132 should be 2n-p.

In an embodiment, the first upscaler 2110 may acquire p first enlarged images 2112 by upscaling the second image 135 using different p scale methods (p is a natural number). . The feature map acquisition unit 2130 may acquire 2n-p enlarged feature maps 2132 according to the 2n-p feature map extraction algorithm.

In another embodiment, the first upscaler 2110 may acquire different p number of first magnified images 2112 from the second image 135 according to the near list neighbor scale method. The feature map acquisition unit 2130 may acquire 2n-p enlarged feature maps 2132 according to the 2n-p feature map extraction algorithm.

The preprocessor 1837 may determine the resolutions of the first enlarged image 2112 and the enlarged feature map 2132 according to the target resolution of the third image 145 .

As an example, when the target resolution of the third image 145 is twice the resolution of the second image 135 , the preprocessor 1837 sets the resolutions of the first magnified image 2112 and the magnified feature map 2132 . 2 It can be determined to be twice the image 135 . In addition, the first upscaler 2110 and the feature map acquisition unit 2130 may acquire the first magnified image 2112 and the magnified feature map 2132, each having an increased resolution by 2 times compared to the second image 135 , respectively. have.

As another example, when the target resolution of the third image 145 is 4 times the resolution of the second image 135 , the preprocessor 1837 sets the resolutions of the first enlarged image 2112 and the enlarged feature map 2132 . 2 It can be determined to be twice the image 135 . In addition, the first upscaler 2110 and the feature map acquisition unit 2130 may acquire the first magnified image 2112 and the magnified feature map 2132, each having an increased resolution by 2 times compared to the second image 135 , respectively. have. In this case, the second DNN 3100 performs the resolution of the first enlarged image 2112 and the enlarged feature map 2132 in order to obtain the third image 145 with an increased resolution of 4 times compared to the second image 135 . It is possible to perform an operation to increase by 2 times.

The first enlarged image 2112 and the enlarged feature map 2132 are input to the first convolutional layer 3110 . The first convolution layer 3110 performs convolution processing on the first magnified image 2112 and the magnified feature map 2132 using k filter kernels having a size of 3×3. The k feature maps generated as a result of the convolution process are input to the first activation layer 3120 .

The first activation layer 3120 may provide a non-linear characteristic to the k feature maps. The output of the first activation layer 3120 is input to the second convolution layer 3130 .

The second convolutional layer 3130 performs convolution processing on k feature maps using one filter kernel having a size of 3×3. As a result of the convolution processing, one image is output from the second convolutional layer 3130 . An image output from the second convolutional layer 3130 may be determined as the third image 145 .

As shown in FIG. 21 , the image output from the second convolution layer 3130 and the second enlarged image 2152 obtained by the second upscaler 2150 are summed to obtain a third image 145 . can The second upscaler 2150 may legacy upscale the second image 135 to obtain a second enlarged image 2152 .

21 illustrates that the second DNN 3100 includes two convolutional layers 3110 and 3130 and one activation layer 3120, but this is only an example. According to an embodiment, the second DNN 3100 may be implemented through a recurrent neural network (RNN). That is, the CNN structure of the second DNN 3100 according to an example of the present disclosure may be changed to an RNN structure.

21 shows that the size of the filter kernel of the convolutional layers 3110 and 3130 is 3x3, this is only an example, and the size of the filter kernel used in each convolutional layer may vary depending on the implementation. have.

In the second DNNs 2900, 3000, and 3100 described with reference to FIGS. 19 to 21 , the output of any one convolution layer is summed with the output of the previous layer and the input data of the second DNNs 2900, 3000, and 3100 to create the next It can be passed on to a layer. For example, as described in FIG. 17 showing the first DNN, the output of the first convolutional layer is combined with the input data of the second DNN and transferred to the next layer, and the output of the second convolutional layer is transmitted to the first It may be combined with the output of the convolution layer and the input data of the second DNN and transmitted to the next layer.

Hereinafter, a method of acquiring the residual image (ie, the reduced feature map or the enlarged feature map) described with reference to FIGS. 14, 15, 19, and 20 will be described with reference to FIGS. 22 to 24 .

22 to 24 are diagrams for explaining a method of obtaining a residual image using a first reduced image (or a first enlarged image) and a second reduced image (or a second enlarged image).

Hereinafter, the first reduced image, the second reduced image, and the first DNN will be mainly described, but the first reduced image, the second reduced image, and the first DNN are the first enlarged image, the second enlarged image, and the second DNN, respectively. It will be apparent that it can be replaced with

As described above, when the number of input channels of the first layer of the first DNN is 2n, the sum of the number of first reduced images input to the first DNN and the number of residual images should be 2n.

22 to 24 exemplify a case where 2n is 4.

22 , when the number of first reduced images 2201a and 2201b is two and the number of second reduced images 2202a, 2202b, and 2202c is three, first reduced images 2201a and 2201b Two residual images 2203a and 2203b must be obtained between the second reduced images 2202a, 2202b, and 2202c.

Specifically, one residual image 2203a may be obtained between any one of the two first reduced images 2201a and 2201b and any one of the three second reduced images 2202a, 2202b, and 2202c. In addition, the other residual image 2203b may be obtained between any one of the two first reduced images 2201a and 2201b and any one of the three second reduced images 2202a, 2202b, and 2202c. In this case, in order to prevent the case where the two residual images 2203a and 2203b become the same, the first reduced image and/or the second reduced image used to obtain the second residual image 2203b is the first residual image 2203a. ) may be different from the first reduced image and/or the second reduced image used to obtain the . For example, if the first reduced image 2201a and the second reduced image 2202a are used to obtain the first residual image 2203a, when the second residual image 2203b is obtained, the first reduced image 2201a ) and the second reduced image 2202b, the first reduced image 2201b and the second reduced image 2202a, or the first reduced image 2201b and the second reduced image 2202b can be used. have. That is, by generating the residual images 2203a and 2203b to be different from each other, the type of data input to the first DNN can be diversified.

Next, as shown in FIG. 23 , when the number of first reduced images 2201a and 2201b is two and the number of second reduced images 2202a and 2202b is two, two first reduced images 2201a , 2201b) and one residual image 2203a between any one of the two second reduced images 2202a and 2202b. In addition, the other residual image 2203b may be acquired between any one of the two first reduced images 2201a and 2201b and any one of the two second reduced images 2202a and 2202b. In this case, in order to prevent the case where the two residual images 2203a and 2203b become the same, the first reduced image and/or the second reduced image used to obtain the second residual image 2203b is the first residual image 2203a. ) may be different from the first reduced image and/or the second reduced image used to obtain the . For example, if the first reduced image 2201a and the second reduced image 2202a are used to obtain the first residual image 2203a, when the second residual image 2203b is obtained, the first reduced image 2201a ) and the second reduced image 2202b, the first reduced image 2201b and the second reduced image 2202a, or the first reduced image 2201b and the second reduced image 2202b can be used. have.

Next, as shown in FIG. 24 , when the number of first reduced images 2201a and 2201b is two and the number of second reduced images 2202 is one, two first reduced images 2201a and 2201b ) and one second reduced image 2202 , two residual images 2203a and 2203b may be obtained. That is, the residual image 2203a between the first reduced image 2201a and the second reduced image 2202 and the residual image 2203b between the first reduced image 2201b and the second reduced image 2202 are obtained. can be

22 to 24 only show an embodiment for matching the sum of the number of the first reduced image and the residual image to the number of input channels of the first layer of the first DNN. Methods other than the described methods may be used within the scope apparent to those skilled in the art.

Hereinafter, a training method of the first DNN (2400, 2500, 2600, or 2700) and the second DNN (2900, 3000, or 3100) of low complexity will be described with reference to FIG. 25 .

25 is a diagram for explaining a method of training the first DNN 2503 and the second DNN 2507 .

The first DNN 2503 shown in FIG. 25 may be the aforementioned first DNN (2400, 2500, 2600, or 2700), and the second DNN 2507 may be the aforementioned second DNN (2900, 3000, or 3100). .

Comparing FIG. 25 with the aforementioned FIG. 11, in the training process shown in FIG. 25, the original training image 2501 is preprocessed 2502 before being input to the first DNN 2503, and the first training image 2504 ( Alternatively, it can be seen that the second training image) is pre-processed ( 2506 ) before being input to the second DNN ( 2507 ).

Specifically, the original training image 2501 is pre-processed 2502 , and the reduced image and reduced feature map obtained as a result of the pre-processing 2502 are input to the first DNN 2503 . Here, the pre-processing 2502 may be the same as the process performed by the above-described pre-processing unit 1313 .

The first training image 2504 output by the first DNN 2503 or the first encoded/first decoded second training image from the first training image 2504 is pre-processed 2506 . The enlarged image and enlarged feature map obtained as a result of the preprocessing 2506 are input to the second DNN 2507 . Here, the preprocessing 2506 process may be the same as the process performed by the aforementioned preprocessing unit 1837 .

A third training image 2508 is obtained as a result of processing the enlarged image and the enlarged feature map by the second DNN 2507 .

A downscaled reduced training image 2505 is obtained from the original training image 2501 separately from the original training image 2501 being pre-processed 2502 , and between the first training image 2504 and the reduced training image 2505 . Structural loss information 2510 corresponding to the comparison result of .

Also, complexity loss information 2520 may be determined based on the spatial complexity of the first training image 2504 .

Also, quality loss information 2530 may be determined according to a comparison result between the original training image 2501 and the third training image 2508 .

Since the structural loss information 2510 , the complexity loss information 2520 , and the quality loss information 2530 have been described in detail with reference to FIG. 11 , their descriptions will be omitted herein.

25, structural loss information 2510, complexity loss information 2520, and quality loss information 2530 are used for training of a first DNN 2503, and quality loss information 2530 is a second DNN ( 2507). That is, the quality loss information 2530 is used for both training of the first DNN 2503 and the second DNN 2507 .

The first DNN 2503 may update a parameter such that the final loss information determined based on the structural loss information 2510 , the complexity loss information 2520 , and the quality loss information 2530 is reduced or minimized. Also, the second DNN 2507 may update the parameter so that the quality loss information 2530 is reduced or minimized. According to an implementation example, the second DNN 2507 may be trained based on at least one of the structural loss information 2510 and the complexity loss information 2520 , and the quality loss information 2530 .

The training process of the first DNN 2503 and the second DNN 2507 based on the final loss information is the same as that described with reference to FIG. 11, and thus a description thereof will be omitted here. It can be written as a program or instruction that can be executed in the , and the written program or instruction can be stored in a medium.

The medium may continuously store a program or instructions executable by a computer, or may be a temporary storage for execution or download. In addition, the medium may be various recording means or storage means in the form of a single or several hardware combined, it is not limited to a medium directly connected to any computer system, and may exist distributed on a network. Examples of the medium include a hard disk, a magnetic medium such as a floppy disk and a magnetic tape, an optical recording medium such as CD-ROM and DVD, a magneto-optical medium such as a floppy disk, and those configured to store program instructions, including ROM, RAM, flash memory, and the like. In addition, examples of other media may include recording media or storage media managed by an app store that distributes applications, sites that supply or distribute various other software, or servers.

Meanwhile, the above-described DNN-related model may be implemented as a software module. When implemented as a software module (eg, a program module including instructions), the DNN model may be stored in a computer-readable recording medium.

In addition, the DNN model may be integrated in the form of a hardware chip to become a part of the aforementioned AI decoding apparatus 200 or AI encoding apparatus 600 . For example, a DNN model may be built in the form of a dedicated hardware chip for artificial intelligence, or as part of an existing general-purpose processor (e.g., CPU or application processor) or graphics-only processor (e.g., GPU). it might be

In addition, the DNN model may be provided in the form of downloadable software. The computer program product may include a product (eg, a downloadable application) in the form of a software program distributed electronically through a manufacturer or an electronic market. For electronic distribution, at least a portion of the software program may be stored in a storage medium or may be temporarily generated. In this case, the storage medium may be a server of a manufacturer or an electronic market, or a storage medium of a relay server.

Above, the technical idea of the present disclosure has been described in detail with reference to preferred embodiments, but the technical idea of the present disclosure is not limited to the above embodiments, and those of ordinary skill in the art within the scope of the technical spirit of the present disclosure Various modifications and changes are possible by the person.

Claims (12)

In a server that provides an image using AI (artificial intelligence),
A processor that executes one or more instructions stored in the server;
The processor is
A first reduced image downscaled from the original image and a reduced feature map having a resolution smaller than the resolution of the original image are input as downscale DNN,
Obtaining an AI downscaled first image from the original image in the downscale DNN,
Encoding the first image to obtain image data,
A server that provides the image data to an electronic device.
According to claim 1,
The processor is
A server for acquiring a residual image between the downscaled second reduced image and the first reduced image from the original image as the reduced feature map.
3. The method of claim 2,
The processor is
obtaining a plurality of first reduced images composed of pixels located at different points within the pixel groups of the original image;
A server for acquiring a plurality of residual images between the plurality of first reduced images and the second reduced images as the reduced feature map.
4. The method of claim 3,
A sum of the number of the plurality of first reduced images and the number of the plurality of residual images is equal to the number of input channels of the first layer of the downscale DNN.
3. The method of claim 2,
The server, wherein the first image is obtained as the second reduced image and the output image of the last layer of the downscale DNN are summed.
According to claim 1,
The processor is
A server for obtaining an edge map corresponding to the original image as the reduced feature map.
7. The method of claim 6,
The server, wherein the first image is obtained as the third reduced image downscaled from the original image and the output image of the last layer of the downscale DNN are summed.
According to claim 1,
The processor is
Obtaining a downscaled and upscaled modified image from the original image,
A server for acquiring a residual image between the downscaled fourth reduced image and the first reduced image from the transformed image as the reduced feature map.
According to claim 1,
The output data of any one of the layers of the downscale DNN is added to the output data of the layers before the one of the layers and is input to the next layer.
In a method by a server that provides an image using AI (artificial intelligence),
inputting a first reduced image downscaled from the original image and a reduced feature map having a resolution smaller than the resolution of the original image into a downscale DNN;
acquiring an AI downscaled first image from the original image in the downscale DNN;
obtaining image data by encoding the first image; and
and providing the image data to an electronic device.
A computer-readable recording medium recording a program for executing the method of claim 10. In the electronic device for displaying an image using AI (artificial intelligence),
A processor for executing one or more instructions stored in the electronic device,
The processor is
Obtaining image data generated as a result of encoding the first image,
Decoding the image data to obtain a second image,
Input the first enlarged image upscaled from the second image and the enlarged feature map having a higher resolution than the resolution of the second image to the upscale DNN,
Obtaining an AI upscaled third image from the second image in the upscale DNN,
An electronic device that provides the third image to a display.

Images