KR20210035679A - Compressing apparatus and method of trained deep artificial neural networks for video coding unit - Google Patents

Abstract

Disclosed are a compression device of a trained deep artificial neural network of a video coding tool, and a method thereof. According to an embodiment of the present invention, the compression device comprises: a pruning unit for selectively pruning a trained deep neural network of a video coding tool; a quantization unit for quantizing the trained deep neural network pruned by the pruning unit; and an entropy coding unit for entropy-coding the trained deep neural network quantized by the quantization unit and outputting the same as a bitstream, wherein the pruning unit calculates the value for the elements describing the trained deep neural network by reflecting back propagation through the deep neural network.

Description

Compressing apparatus and method of trained deep artificial neural networks for video coding unit

The present invention relates to an artificial neural network (ANN), and more particularly, to a compressed expression technique of a trained deep artificial neural network for a video coding tool.

Attempts have been made to utilize artificial intelligence (AI) in various industries. In particular, the recent artificial intelligence technology uses a neural network (NN), which is an information processing system having a specific performance in common with biological neural networks, and its performance is greatly improved, and accordingly, application fields are rapidly increasing. have.

Such a neural network (NN) is also called an'artificial' neural network (ANN). An artificial neural network (ANN) is a distributed parallel information processing model that mimics the behavioral characteristics of animal neurons. ANN has a large number of nodes (also called neurons) that are connected to each other. An ANN has two characteristics: 1) Each neuron computes a weighted input from another adjacent neuron through a specific output function (also called an activation function). 2) The intensity of information transmission between neurons is measured by so-called "weights", and such weights can be adjusted by self-learning of specific algorithms.

ANNs can use different architectures to designate variables and topology relationships included in neural networks. A parameter included in the neural network may be a weight of a connection between neurons along with the activity of the neurons. There are feed forward networks and backward propagation neural networks as types of neural network topologies. In the former, nodes within each layer connected to each other in the same layer are supplied to the next stage, which includes some form of'learning rule' that modifies the weight of the connection according to the provided input pattern. The latter allows the propagation of errors in the reverse direction of the weighting adjustment, which is an advanced neural network than the former.

A deep neural network (DNN) responds to a neural network having multiple levels of interconnected nodes and can compactly express a highly nonlinear and highly variable function. Nevertheless, with the number of nodes associated with multiple layers, the computational complexity for the DNN increases rapidly. Until recently, efficient computation methods for learning or training such DNNs have been developed. As the learning speed of DNN increases dramatically, it has been successfully applied to various and complex tasks such as speech recognition, image segmentation, object detection, and facial recognition.

Multimedia content, such as video compression and decompression, is also one of the fields in which DNN is being applied. High Efficiency Video Coding (HEVC) as the current next-generation video coding was developed by the joint video project of the ITU-T (Video Coding Expert Group) and ISO/IEC MPEG (Video Expert Group) standardization organizations and adopted as an international standard. It has been used, and it is known that it is possible to further improve its performance by applying the DNN to a new video coding standard such as HEVC. One such attempt is disclosed in Korean Patent Laid-Open Publication No. 10-2018-0052651, "Method and apparatus of neural network-based processing in video coding".

However, the scale of neural networks has exploded due to rapid development in recent years. Some advanced neural network models will have hundreds of layers and billions of connections. And its implementation is both computation-centric and memory-centric.

As neural networks are getting bigger, it is very important to make the neural network model small in order to apply it to devices with limited storage and processor performance, such as mobile terminals, but this may degrade the performance of the neural network, so there is a limitation. . Particularly, in order to apply to a video coding application for the production and consumption of multimedia content used as an important application in a mobile terminal, a neural network model of a small size is essential. In addition, due to the characteristics of a video coding application, it is necessary to ensure compatibility between the encoding device and the decoding device.

Korean Patent Laid-Open Patent No. 10-2018-0052651, "Method and apparatus for processing based on neural networks in video coding"

One problem to be solved by the present invention is a trained deep neural network (trained deep neural network) of a video coding tool that can be applied to devices with limited performance of storage and processors such as mobile terminals while minimizing deterioration of the performance of a deep neural network. deep neural networks) compression apparatus and method.

The apparatus for compressing a trained deep artificial neural network according to an embodiment of the present invention for solving the above-described problem includes a pruning unit for selectively pruning a trained deep neural network of a coding tool constituting video coding, and the branch. A quantization unit for quantizing the trained deep neural network pruned by a pruning unit, and an entropy coding unit for entropy coding the trained deep neural network quantized by the quantization unit and outputting a bitstream, the branch The stroke unit, after identifying the type of the coding tool, calculates a value degree for the elements describing the trained deep neural network in consideration of the type of the identified coding tool, which propagates backward through the deep neural network. Reflecting the above value is calculated.

According to an aspect of the embodiment, the trained deep neural network includes a plurality of output characteristics, a plurality of neurons, and a plurality of convolution kernels, and one or more of the elements describing the trained deep neural network are, It may include at least one of the plurality of output characteristics, the plurality of neurons, and the plurality of convolution kernels. In this case, the trained deep neural network may include a plurality of neurons present in the full connection layer.

A method for compressing a trained deep artificial neural network according to an embodiment of the present invention for solving the above-described problem includes a pruning step for selectively pruning a trained deep neural network of a coding tool constituting video coding, the branch A quantization step for quantizing the trained deep neural network pruned in the pruning step, and an entropy coding step for entropy coding the trained deep neural network quantized in the quantization step and outputting a bitstream, the pruning step In, after identifying the type of the coding tool, the value of the elements describing the trained deep neural network is calculated in consideration of the type of the identified coding tool, reflecting that propagating backward through the deep neural network. Calculate the above value.

According to an embodiment of the present invention, a trained deep neural network can be applied to an application using a video even in a device having limited storage and processor performance, such as a mobile terminal, while minimizing deterioration of the performance of a neural network.

1 is a block diagram showing a detailed configuration of a computer system in which an apparatus for compressing a trained deep artificial neural network of a video coding tool according to an embodiment of the present invention is implemented.
2 is a block diagram showing an example of a configuration of a deep neural network compression apparatus.
3 is a diagram showing an example of a trained deep neural network according to an embodiment of the present invention, in the case of a convolutional neural network (CNN).
4 is a flowchart illustrating an example of a pruning method in a pruning unit of a compression apparatus of a trained deep artificial neural network of a video coding tool according to an embodiment of the present invention.
5 is a flowchart illustrating an example of a pruning method in a pruning unit of a compression apparatus of a trained deep artificial neural network of a video coding tool according to another embodiment of the present invention.
6 is an example of a quantization process in a quantization unit, and is a flowchart showing an example of an adaptive quantization process.
7 is a flowchart showing a quantization process according to another embodiment of the present invention.

Hereinafter, preferred embodiments and examples of the present invention will be described with reference to the drawings. However, the following embodiments and examples are merely illustrative of preferred configurations of the present invention, and the scope of the present invention is not limited to these configurations. And in the following description, the hardware configuration and software configuration of the device, processing flow, manufacturing conditions, size, material, shape, etc. are not intended to limit the scope of the present invention to this, unless specifically stated otherwise. .

1 is a block diagram showing a detailed configuration of a computer system 100 in which an apparatus for compressing a trained deep artificial neural network of a video coding tool according to an embodiment of the present invention is implemented.

Referring to FIG. 1, the computer system 100 includes one or more processors 110, an input/output device interface 120, a network interface 130, an interconnector (BUS, 140), a memory 150, and a storage 160. Includes. The computer system 100 may be one specific device composed of a single computing device, or may be a plurality of devices including one or more processors and one or more associated memories.

The processor 110 fetches and executes programming instructions stored in the memory 150 or the storage 160. Likewise, the processor 110 stores or retrieves application data in the memory 150. The input/output device interface 120 is for connecting the input/output device 12 such as a keyboard, a display, and a mouse device to the computer system 100. The network interface 130 is for communicating with an external network such as an internal network (intranet), the Internet, or a wireless communication network through wired or wireless communication, and transmits data through the data communication network 14.

The interconnector 140 performs a function of transmitting programming commands and application data between the processor 110 and the input/output device interface 120, the storage 160, the network interface 130, and the memory 150, respectively. do. The interconnector 140 may be one or more buses. The processor 110 may be implemented as a single central processing unit (CPU) or as a single CPU having multiple CPUs, and multiple processing cores in various implementations. According to one aspect, the processor 110 may be a digital signal processor (DSP).

The memory 150 generally includes a memory random access memory such as static random access memory (SRAM), dynamic random access memory (DRAM), or flash. Storage 160 is generally a hard disk drive, solid state device (SSD), removable memory card, optical storage, flash memory device, such as connections to a network attached storage (NAS) or a storage area device (SAN). Includes non-volatile memory.

The computer system 100 may include one or more operating systems (OS) 164. A part of the operating system 164 may be stored in the memory 150 and a part of the operating system 164 may be stored in the storage 160. Alternatively, the entire operating system 164 may be stored in the memory 150 or in the storage 160. The operating system 164 provides an interface between various hardware resources such as the processor 110, the input/output device interface 110, the network interface 130, and the like. In addition, the operating system 164 provides common services such as a time function for an application program.

The deep neural network compression apparatus 152 compresses a trained deep neural network of a coding tool, for example, a trained deep convolutional neural network, and outputs a compressed bitstream. That is, the deep neural network compression apparatus 152 is a means for compressing a deep neural network of a learned or trained coding tool and describing it in a compatible format. According to this, the resulting compressed deep neural network corresponds to an interoperable compressed representation of neural networks of a coding tool. In order to compress the trained deep neural network of the coding tool, the deep neural network compression device 152 is a series of processes including pruning, quantization, and entropy coding for the trained deep neural network. To output a bitstream.

2 is a block diagram showing an example of the configuration of the deep neural network compression apparatus 152. Referring to FIG. 2, the deep neural network compression apparatus 152 includes a pruning unit 22, a quantization unit 24, and an entropy coding unit 26. As described above, the deep neural network compression apparatus 152 compresses the trained deep neural network of the coding tool and outputs it as a bitstream.

The input to the deep neural network compression device 152 includes various information and parameters for describing the trained deep neural network of the coding tool.

First, the input to the deep neural network compression device 152 includes various high-level information. For example, the high-level information may include information indicating the type of a corresponding deep neural network (DNN)-based coding tool. There are two types of DNN-based coding tools. More specifically, there are two types of DNN-based coding tools: a first type coding tool that implements a function essential to both an encoder and a decoder, and a second type coding tool that implements a function essential to only one of the encoder and the decoder. This is, in image/video coding, some coding tools, i.e., the first type coding tool, are required for both the encoder and the decoder, and the rest of the coding tools, i. This is because it is a function to be used. For example, in the video coding process, the in-loop filtering process corresponds to the function of the first type coding tool performed at both the encoder and the decoder, but the intra mode prediction process is the function of the second type coding tool performed only at the encoder. And only determined prediction mode information is sent to the decoder. Therefore, two types of DNN-based coding tools must be considered, and the type of these DNN-based coding tools must be indicated as high-level information of the trained DNN-based coding tools.

And, the high-level information includes information on the overall configuration of the DNN-based coding tool. More specifically, as high-level information related to the composition of the trained DNN-based tool, targets viewed from the perspective of basic functions of the neural network such as recognition, classification, generation, and discrimination. Information on the target application, information indicating the type of trained DNN-based coding tool, applying the trained tool neural network to the inference engine when the encoder performs a specific coding process, and a standardized image or video coding tool. Information indicating what is selected from among those to be applied, information on customized content type, autoencoder, CNN (Convolutional Neural Network), GAN (Generative Adversarial Network), RNN (Recurrent Neural Network) ), etc., basic information about the algorithm of the trained DNN-based neural network, basic information about training data and/or test data, information about the capability required by the inference engine in terms of memory capacity and computing power, information about model compression. And the like.

The input to the deep neural network compression device 152 includes various parameters describing the trained deep neural network. Various parameters include kernel, neuron, and connection weights. Hereinafter, with reference to the architecture of a convolutional neural network (CNN), which is an example of a deep neural network, this will be described in more detail.

3 is a diagram showing an example of a trained deep neural network according to an embodiment of the present invention, in the case of a convolutional neural network (CNN). Referring to FIG. 3, the CNN 200 includes an input layer (imput layer, 210), a convolutional layer (215), a subsampling layer (220), a convolutional layer (225), and a subsampling layer (230). , A fully connected (FC) layer (235, 240) and an output layer (245). In the example shown in FIG. 3, the input layer 210 is configured to accept a 32×32 pixel image, and the convolution layer 215 generates six 28×28 feature maps from the input layer.

As such, although the configuration of a specific CNN is shown in FIG. 3, more generally, the CNN includes one or more convolutional layers each having a sub-sampling step, and one or more fully-connected layers. In general, the illustrated CNN architecture is designed to use a two-dimensional structure of an input image. For example, CNN achieves this by using local connections and connected weights followed by some form of pooling. In general, compared to all-connected networks with a similar number of hidden units, CNNs are easier to train and tend to have fewer parameters.

In general, the CNN includes a convolutional layer and a subsampling layer, followed by a full connection layer. According to an embodiment, the CNN receives an image of x×y×z as an input in the convolutional layer, where x and y represent the height and width of the image, respectively, and z represents a channel in the image. For example, an RGB image has a z=3 channel. The convolutional layer may include a filter (kernel) of size a×b×c, where a×b is less than x×ㅛ and c is less than or equal to z. In general, the size of the filter k results in a locally linked structure, which is convolved with the image to produce k feature maps. In addition, each feature map is subsampled over adjacent regions of various sizes.

The pruning unit 22 is for simplifying the configuration of the neural network by removing some of the nodes (neurons) constituting the trained deep neural network and/or their connection relationships. That is, the pruning unit 22 searches for an appropriate threshold according to an input pruning rate (search threshold), and processes all weights below the found threshold as '0' (pruning weight).

For example, the pruning unit 22 is a part of neurons constituting the trained neural network (left) according to a predetermined pruning rate (as a connection of neurons indicated by a green line, and the weight is random). Connections that are below the threshold) are removed, leaving only the connections of the neurons marked in blue. According to this, the expression of the trained neural network is simplified by substituting all weights whose weight values representing the trained neural network are less than or equal to a predetermined threshold with '0'. Here, the size of the threshold value may vary according to the pruning ratio. The pruning unit 22 finds a threshold value corresponding to the input value ratio, and processes weights below the threshold value as '0'.

A direct way to reduce the redundancy of a trained deep neural network is to prune the kernel of the convolutional layer and the neurons of the FC layer. Although the redundancy of the model generally increases the universality of the deep neural network, selectively reducing the redundancy of the deep neural network can appropriately reduce the redundancy, so that the model's predictive power, inference speed, memory usage, storage space and power A balance between consumption can be achieved. Although pruning kernels and neurons is beneficial, random pruning or randomly changing the number of neurons and kernels can lead to a major decline in predictive power.

From the perspective of feature extraction, the kernel and neurons of the deep neural network model can be regarded as feature extractors. Thus, when faced with a feature that is quite large, the pruning unit 22 uses a feature selection process to individualize irrelevant and/or whether or not features and avoid redundancy. In general, applying the feature selection/ranking method to the extracted features implies the importance of each feature extractor, so that the pruning unit 22 prunes the less important ones, between the predictive power of the deep neural network and the model overkill. To achieve a balance of. In doing so, the pruning unit 22 extracts the responses of each convolution and FC layer and ranks the kernel and neurons by their value from the viewpoint of feature selection, pruning the less valuable ones. The pruning unit 22 may also recover predictive power by performing fine tuning at a smaller learning rate using the selected valuable feature extractors as a starting point of a smaller deep neural network.

One obstacle to selective pruning of kernels and neurons is that the dimensions of the features extracted by the convolutional and FC layers are still large. Accordingly, the pruning unit 22 according to an exemplary embodiment of the present invention performs value score backward propagation in order to reduce the size of the deep neural network and maintain the predictive power at the same time. According to an embodiment, the pruning unit 22 may assign a value score to a higher level of the deep neural network, such as inputs of a classifier, and propagate the value score backward to the lower levels of the deep neural network. . Through this value score backward propagation, the pruning unit 22 not only can effectively measure the value of the feature extractors of the entire deep neural network, but also can perform consistently throughout the entire network.

4 is a flowchart showing an example of a pruning method in the pruning unit 22 of the compressed device 152 of a trained deep artificial neural network of a video coding tool according to an embodiment of the present invention.

Referring to FIG. 4, the pruning unit 22 identifies a coding tool described as a trained deep neural network (S11). That is, the pruning unit 22 first identifies what type of coding tool the trained deep neural network is. For example, the pruning unit 22 may identify whether the corresponding coding tool relates to a coding function commonly performed in encoding and decoding or a coding function applied only to encoding and not applied to decoding.

Then, the pruning unit 22 extracts characteristics of the trained deep neural network of the coding tool in consideration of the type of the identified coding tool (S12). This process corresponds to the process of extracting the deep neural network response from the trained deep neural network. Depending on whether the coding tool is applied to both encoding and decoding or only to encoding, a detailed algorithm applied when extracting features may vary.

Then, the pruning unit 22 calculates and measures a value degree score for each feature extractor (S13). In this step, the pruning unit 22 calculates a value score for the extracted features, and an example of the method will be described below. In the following description, a spatially square three-way tensor will be described, but it may be extended and applied to other configurations. For a convolutional layer having an output tensor size Y×Y×F (where Y is the spatial size and F is the number of output channels), the pruning unit 22 first determines the spatial position (i, j). The value of each position of the f-th output channel having the value IS _ijf is obtained. According to an embodiment, the pruning unit 22 calculates a value score for the f-th output channel using Equation 1.

In general, there are three main categories of feature selection. The first is a wrapper that gives a score for a subset of features using a classifier, the second is an embedded method that potentially selects features in the training process of the classifier by a regularization method, and the third is It is a filter method that uses the unique characteristics of the data regardless of the classifier. According to an embodiment, the pruning unit 22 may be designed to perform feature selection based on a response of a previously trained model. In a particular embodiment, the pruning unit 22 may use an Infinite Feature Selection filter algorithm to perform feature ranking. In general, when performing the infinite feature selection filter algorithm, the pruning unit 22 correlates the feature selection problem to the affinity graph, where each vertex is a feature, and the edges between the vertices are a relationship. The importance of is defined by the function of the variable and the correlation of the vertex pairs. In the graph, each path (set of vertices and edges) is considered a subset of features, and the path cost is the sum of the edge weights. Therefore, in the case of performing the infinite feature selection filter algorithm, the pruning unit 22 evaluates the value of a given feature while considering a subset of all impersonated features, so that the score of each feature is assigned to all other features. Affected by

As described above, the pruning unit 22 measures the value of the feature extractor (S13). For example, the pruning unit 22 may consider the output according to the infinite feature selection filter algorithm as a value of each feature as a score. According to one embodiment, the response of each neuron is calculated by convolution. The pruning unit 22 may leverage the weight of the deep neural network to match the value of the feature extractor. In the case of neuron A, the neuron's value score propagates backwards in proportion to the weight of the connection, to neurons of the previous layer, either the FC layer or the convolutional layer, connected to neuron A. Given the value score vector for neurons in the FC layer, the pruning unit 22 may prune neurons based on the rank.

Once the value score is calculated, the pruning unit 22 prunes the less important kernels based on the rank of the value score of the output channel (S14). According to an embodiment, the pruning ratio used by the pruning unit 22 may be determined by a predetermined parameter in consideration of a balance between classification performance and model excess. Then, the pruning unit 22 performs necessary fine adjustment on the pruned deep neural network (S15).

5 is a flowchart showing an example of a pruning method in the pruning unit 22 of the compressed device 152 of a trained deep artificial neural network of a video coding tool according to another embodiment of the present invention. The pruning method according to the present embodiment is different from the above-described embodiment in that it uses the backward propagation of the value score.

Referring to FIG. 5, the pruning unit 22 identifies a coding tool described as a trained deep neural network (S21). That is, the pruning unit 22 first identifies what type of coding tool the trained deep neural network is. For example, the pruning unit 22 may identify whether the corresponding coding tool relates to a coding function commonly performed in encoding and decoding or a coding function applied only to encoding and not applied to decoding.

Then, the pruning unit 22 extracts a characteristic of the high-level layer of the trained deep neural network of the coding tool in consideration of the type of the identified coding tool (S22). This process corresponds to the process of extracting the deep neural network response from the trained deep neural network. Depending on whether the coding tool is applied to both encoding and decoding or only to encoding, a detailed algorithm applied when extracting features may vary.

And the pruning unit 22 calculates and measures a value degree score for each feature extractor (S23). For example, the pruning unit 22 may calculate a value degree score using Equation 1 described above. The calculated value score can be used to selectively prune the kernel and neurons of the deep neural network, and in this embodiment, pruning starts from the high-level layer.

Subsequently, the pruning unit 22 backpropagates the value score of the selected feature extractor (S24). In this case, the pruning unit 22 is pruned so that neurons and kernels that have already been removed can be ignored. If the deep neural network is a neural network without an FC layer before the final classifier, the pruning unit 22 may perform feature selection on the flattened response of the last convolutional layer.

In general, the pruning unit 22 may use the value back propagation in order to transfer the value from the downstream kernels and neurons of the deep neural network to the upstream kernels and neurons of the deep neural network. For example, when a specific neuron has a predetermined weight, the pruning unit 22 identifies the neuron previously used to calculate the activity of the neuron, and is proportional to the weight corresponding to the operation of the layer. It propagates the value of neurons backwards.

Then, the pruning unit 22 performs pruning based on the result of the backward propagated value score (S25), and then performs necessary fine adjustments on the pruned deep neural network (S26).

With continued reference to FIG. 2, the quantization unit 24 performs quantization on outputs of the pruning unit 22, that is, weights. To this end, quantization is performed based on input quantization bits, and a quantized weight is output by extracting a maximum/minimum value (Max/Min vaules).

According to an aspect of the present embodiment, in order to reduce a quantization error by considering a distribution of parameters input for quantization, the quantization unit 24 may perform adaptive quantization. In the adaptive quantization process, a compressed integer weight file and a codebook are input, and a quantization level (r _k ) and a quantization region boundary (d _k ) may be expressed by Equation (2).

An example of such an adaptive quantization process is shown in FIG. 6.

In the adaptive quantization process as shown in FIG. 6, if the quantization error is not low enough, non-uniform quantization may be replaced with uniform quantization. According to this, the input includes a configuration file indicating the number of layers. And, if the size of the quantization error is larger than the uniform quantization, uniform quantization may be used instead of non-uniform quantization.

7 is a flowchart showing a quantization process according to another embodiment of the present invention. The quantization process shown in FIG. 7 adds/supplements the existing quantization process for storing a quantized value (integer, integer) of a neuralnet weight value in a binarized form. According to the quantization process shown in FIG. 7, lossy coding may use not only quantization but also other techniques, and may be, for example, a method of dividing coding in a weight value matrix (with good coding efficiency (RD (Estimated by Rate distortion), etc.) Determine the segmentation map). In addition, the quantized values of the weight values are generated by entropy coding (arithmetic coding, CABAC, palette, index map coding, etc.), which is lossless coding. In addition, in the process of decoding (or decompression), reconstruction is performed by inputting a binary file (bitstream), etc., at this time, information on whether to perform the neuralnet model reconstruction in its entirety or only part of it may be included. have.

2, the entropy coding unit 26 encodes each of the quantized weights and indexes according to a predetermined algorithm (e.g., entropy coding), and as a result, a bit stream of a compressed deep neural network is output. . According to the present embodiment, there is no particular limitation on a specific process of entropy encoding, and as long as it is known in the art, it may be applied without limitation, unless an algorithm is essentially impossible to apply due to the nature of entropy encoding.

As described above, the above description is merely an example and should not be construed as being limited thereto. The technical idea of the present invention should be specified only by the invention described in the claims to be described later, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention. Therefore, it is obvious to a person skilled in the art that the above-described embodiments can be modified and implemented in various forms.

Claims (6)

A pruning unit for selectively pruning a trained deep neural network of a coding tool constituting video coding;
A quantization unit for quantizing the trained deep neural network pruned by the pruning unit; And
Entropy coding tool for entropy coding the trained deep neural network quantized by the quantization unit and outputting it as a bitstream,
The pruning unit, after identifying the type of the coding tool, calculates the value of the elements describing the trained deep neural network in consideration of the type of the identified coding tool, and propagates backward through the deep neural network. Compressing apparatus of a trained deep artificial neural network of a video coding tool, characterized in that to calculate the value by reflecting that.
The method of claim 1,
The trained deep neural network includes a plurality of output characteristics, a plurality of neurons, and a plurality of convolution kernels,
One or more of the elements describing the trained deep neural network include at least one of the plurality of output characteristics, the plurality of neurons, and the plurality of convolution kernels. Compression device of deep artificial neural network.
The method of claim 2,
The trained deep neural network is a compression apparatus for a trained deep artificial neural network of a video coding tool, characterized in that it includes a plurality of neurons present in a full connection layer.
A pruning step for selectively pruning a trained deep neural network of a coding tool constituting video coding;
A quantization step for quantizing the trained deep neural network pruned in the pruning step; And
Entropy coding the trained deep neural network quantized in the quantization step and outputting it as a bitstream,
In the pruning step, after the type of the coding tool is identified, the value of the elements describing the trained deep neural network is calculated in consideration of the type of the identified coding tool, but propagated backward through the deep neural network. A method for compressing a trained deep artificial neural network of a video coding tool, characterized in that to calculate the value by reflecting that.
The method of claim 4,
The trained deep neural network includes a plurality of output characteristics, a plurality of neurons, and a plurality of convolution kernels,
One or more of the elements describing the trained deep neural network include at least one of the plurality of output characteristics, the plurality of neurons, and the plurality of convolution kernels. Deep artificial neural network compression method.
The method of claim 5,
The trained deep neural network, a method for compressing a trained deep artificial neural network of a video coding tool, characterized in that it includes a plurality of neurons present in a full connection layer.

Images