KR20230075248A - Device of compressing data, system of compressing data and method of compressing data - Google Patents

Abstract

A data compression device according to an embodiment of the present invention comprises: an input unit for acquiring data; a first encoder unit for generating first data in which data is lossy compressed; a probability distribution prediction unit previously trained to predict a plurality of parameters for the probability distribution of residual data using the residual data as input; and a second encoder unit for generating second data in which the residual data is losslessly compressed using the plurality of predicted parameters of the residual data. The first encoder unit and the second encoder unit may be previously trained so that the sum of the size of the first data and the size of the second data has a minimum value. Accordingly, a high compression ratio can be achieved even with small complexity.

Description

Data compression device, data compression system and data compression method

The present invention relates to an apparatus for compressing data, a data compression system, and a data compression method.

In order to differentiate from existing mobile communication systems, recent 5G and 6G mobile communication systems are required to be designed in a form capable of supporting high-quality network services such as ultra-realistic remote reality that guarantees high-level application performance. In order to guarantee the required performance of high-quality application services such as ultra-realistic remote reality, it is necessary to transmit high-quality volumetric data including 8K or higher resolution video and high-quality five-sensory data with ultra-low latency.

Transmission delay for ultra-high quality data is greatly affected by transmission time, which is defined as a value obtained by dividing a transmission amount (data size) by a transmission rate (data_rate) among various delay factors. When there is a physical transmission limit in the Internet bottleneck section, it is necessary to reduce the transmission amount itself, and an ultra-high efficiency data compression method is required to reduce the transmission amount.

On the other hand, compression codecs such as conventional PNG and GIF and deep neural network (DNN)-based compression codecs losslessly compress original data or losslessly compress original data, and losslessly restore original data to restore original and restored data. A method of losslessly compressing residual data defined by the difference of is used. In the prior art, a method for lossy compressing original data and a method for losslessly compressing residual data are performed independently, and thus an optimal compression rate cannot be achieved.

An object of the present invention is to provide a data compression device, a data compression system, and a data compression method having an improved data compression ratio by simultaneously considering lossy compressed data of original data and lossless compressed data of residual data.

However, the problem to be solved by the present invention is not limited to those mentioned above, and another problem to be solved that is not mentioned can be clearly understood by those skilled in the art from the description below. will be.

A data compression device according to an embodiment of the present invention includes an input unit for acquiring data; a first encoder unit generating first data obtained by lossy compression of the data; a probability distribution prediction unit pre-learned to predict a plurality of parameters for a probability distribution of the residual data by taking residual data obtained by subtracting loss-reconstructed data from the data as an input; and a second encoder unit generating second data obtained by lossless compression of the residual data using a plurality of parameters of the predicted residual data, wherein the size of the first data and the size of the second data The first encoder unit and the second encoder unit are pre-learned so that the sum of s has a minimum value.

The first encoder unit may include at least one of a generative adversarial network and an autoencoder.

The probability distribution prediction unit may be pre-learned to predict a plurality of parameters of a probability distribution of the residual data for learning using residual data for learning as an input and a plurality of parameters for a probability distribution of the residual data for learning as labels. .

The probability distribution of the residual data for learning includes at least one of a Gaussian distribution, a Poisson distribution, and a logistic distribution, and the plurality of parameters may be a multi-parameter mixture of the probability distribution.

The probability distribution prediction unit may include at least one of a convolutional neural network and a recurrent neural network.

The first encoder unit and the second encoder unit may be pre-learned such that a sum of the entropy of the first data and the entropy of the second data is minimized through joint-optimization.

The first encoder unit and the second encoding unit may determine the sum of the entropy of the first data and the cross-entropy between a plurality of parameters for probability distributions of the second data and the predicted residual data. It may be pre-learned to be minimized.

A data compression system according to another embodiment of the present invention includes an input unit for obtaining data; a first encoder unit generating first data obtained by lossy compression of the data; a probability distribution prediction unit pre-learned to predict a plurality of parameters for a probability distribution of the residual data by taking residual data of the data as an input; a second encoder unit generating second data obtained by lossless compression of the residual data using a plurality of parameters of the predicted residual data; And a decoder unit for obtaining and restoring the first data and the second data, respectively, and restoring the data without loss by adding loss-restored data from the first data and residual data restored from the second data; The first encoder unit and the second encoder unit may be pre-learned such that a sum of the size of the first data and the size of the second data has a minimum value.

A method for compressing data according to an aspect of the present invention includes obtaining the data; generating first data obtained by lossy compression of the obtained data; predicting a plurality of parameters of a probability distribution of the residual data by using residual data of the data as an input using a pre-learned probability distribution prediction neural network; and generating second data obtained by lossless compression of the residual data using a plurality of parameters of the predicted residual data, wherein the size of the first data and the size of the second data are The sum is pre-learned to have a minimum value.

A computer readable recording medium according to another aspect of the present invention includes a computer program, the computer program includes instructions for causing a processor to perform a method of compressing data, the method comprising: acquiring the data; generating first data obtained by lossy compression of the obtained data; predicting a plurality of parameters of a probability distribution of the residual data by using residual data of the data as an input using a pre-learned probability distribution prediction neural network; and generating second data obtained by lossless compression of the residual data using a plurality of parameters of the predicted residual data, wherein the size of the first data and the size of the second data are The sum is pre-learned to have a minimum value.

A computer readable recording medium according to another aspect of the present invention includes a computer program, the computer program includes instructions for causing a processor to perform a method of compressing data, the method comprising: obtaining the data; ; generating first data obtained by lossy compression of the obtained data; predicting a plurality of parameters of a probability distribution of the residual data by using residual data of the data as an input using a pre-learned probability distribution prediction neural network; and generating second data obtained by lossless compression of the residual data using a plurality of parameters of the predicted residual data, wherein the size of the first data and the size of the second data are The sum is pre-learned to have a minimum value.

According to an embodiment of the present invention, since the amount of compressed data can be drastically reduced while losslessly compressing original data, high-quality data can be transmitted in the shortest time, so that various types of information can be rapidly accumulated as data in a next-generation network environment. However, it has the effect of realizing high value-added application services such as ultra-realistic remote reality.

In addition, according to an embodiment of the present invention, since the complexity of compressing and restoring data is low, there is an effect of achieving a high compression rate even with low complexity.

1 is a block diagram conceptually illustrating the functions of a data compression apparatus according to an embodiment of the present invention.
2 shows a transmission amount for compressing and transmitting data as a picture.
3 is a block diagram conceptually illustrating functions of a data compression system according to an embodiment of the present invention.
4 is a block diagram conceptually illustrating a flow of compressing and restoring data by a data compression system according to an embodiment of the present invention.
5 is a graph comparing performance of a data compression method according to an embodiment of the present invention and other data compression methods.
6 is a block diagram illustrating a data compression device according to an embodiment of the present invention in terms of hardware.
7 is a flowchart illustrating a data compression method according to an embodiment of the present invention.

Advantages and features of the present invention, and methods of achieving them, will become clear with reference to the detailed description of the following embodiments taken in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms, and only these embodiments make the disclosure of the present invention complete, and common knowledge in the art to which the present invention belongs. It is provided to completely inform the person who has the scope of the invention, and the present invention is only defined by the scope of the claims.

In describing the embodiments of the present invention, if it is determined that a detailed description of a known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description will be omitted. In addition, terms to be described later are terms defined in consideration of functions in the embodiment of the present invention, which may vary according to the intention or custom of a user or operator. Therefore, the definition should be made based on the contents throughout this specification.

1 is a block diagram conceptually illustrating the functions of a data compression apparatus according to an embodiment of the present invention.

Referring to FIG. 1 , the data compression device 100 may include an input unit 110, a first encoder unit 120, a second encoder unit 130, and a probability distribution prediction unit 140. However, the configuration of the data compression device 100 is not limited to that shown in FIG. 1 . For example, although not shown in FIG. 1 , the data compression device 100 may include a communication unit for communicating with other devices, a display unit for displaying compression results, and the like.

Depending on the embodiment, the data compression apparatus 100 may receive data to be compressed, loss-compress original data, and losslessly compress residual data obtained by subtracting loss-reconstructed data from original data. Here, the data compression apparatus 100 may determine a compression rate such that the sum of lossy compressed data and lossless compressed data is minimized.

The input unit 110 may receive data to be compressed from an external device separate from the data compression device 100 . For data input, the input unit 110 may receive data using wired or wireless communication, or may directly receive data from a user, or may receive data stored in the data compression device 100 through the input unit 110. can be entered through A method for the input unit 110 to receive data to be compressed is not limited thereto.

In FIG. 2 , first, the data transmission amount for the data compression device 100 to compress data will be described.

(a) of FIG. 2 shows the amount of data transmission for lossless compression and transmission of data.

Referring further to (a) of FIG. 2 , the size of data obtained by compressing the original data (compressed data) may be expressed as a product of the number of symbols of lossless compressed data N and the average symbol length L of the lossless compressed data. In addition, in order to restore compressed data, a probability distribution for symbols of lossless compressed data is required, and if the probability distribution is defined as p, the actual data transmission amount (T) required to restore the original data is It can be defined as the size (NL) and the data I(p) necessary to convey the probability distribution.

에 대한 데이터를 전달할 수 있다.Here, since a large amount of data (bits) is required to deliver the probability distribution p as it is, instead of delivering the accurate probability distribution p, the artificial neural network estimated

data can be transmitted.

(b) of FIG. 2 shows the amount of data transmission for transmission after loss-compressing data and losslessly compressing residual data.

Further referring to FIGS. 1 and 2(b), the data compression device 100 transmits data through lossless compression as shown in FIG. Since the original data and the residual data of the original data can be lossy compressed and lossless compressed, respectively, it is possible to compensate for the wide distribution range and uncertain characteristics of the original data.

To this end, the first encoder unit 120 may loss-compress the data without losslessly compressing the data, and the second encoder unit 130 loss-restores the compressed data obtained from the first encoder unit in the original data. The residual data after subtracting can be losslessly compressed.

The first encoder unit 120 may include a neural network that compresses data into a small size in order to perform lossy compression. According to an embodiment, the first encoder unit 120 may include at least one of a generative adversarial network and an autoencoder.

In this case, the data transmission amount (T) compressed by the data compression device 100 is the size of data obtained by compressing the original data with loss (Compressed data with loss) and the size of data obtained by compressing the residual data without loss (Compressed Data without loss) and It is the sum of all the data sizes of the predicted probability distribution (Distribution information) of the residual data predicted by the probability distribution predictor 140 to be described later. The amount of data transmission (T) compressed by the data compression device 100 can be expressed as in Equation 1 below.

은 잔차 데이터의 확률 분포를 의미하고,

은 추정된 잔차 데이터의 확률 분포를 의미한다. Here, x corresponds to the original data, E() is a lossy compressor, and b() means the amount of transmission data for transmitting the input of the function. also,

is the probability distribution of the residual data,

denotes the probability distribution of the estimated residual data.

The probability distribution predictor 140 may include a pre-learned artificial neural network in order to predict the probability distribution of residual data.

More specifically, the probability distribution predictor 140 takes residual data for learning as an input and labels a plurality of parameters for the probability distribution of the residual data for learning to predict a plurality of parameters for the probability distribution of the residual data for learning. can be learned in advance.

Also, according to an embodiment, the probability distribution predictor 140 includes at least one of a convolutional neural network and a recurrent neural network in order to predict the plurality of parameters of residual data for learning. can do. However, the artificial neural network included in the probability distribution predictor 140 is merely exemplary, and the plurality of parameters of the residual data for learning may be predicted using other known artificial neural network models.

The probability distribution of residual data for learning input to the probability distribution prediction unit 140 includes at least one of a Gaussian distribution, a Poisson distribution, and a logistic distribution, and the plurality of parameters are a multi-parameter mixture of the probability distribution ) can be.

Here, the data compression device 100 can minimize the data transmission amount (T) while performing lossless compression using a joint optimization method in order to minimize the data transmission amount (T) of Equation 1. there is.

If the data encoded by the lossy compressor is z and the residual data is r, the data compression apparatus 100 calculates the entropy (H(p _z )) of the probability distribution of the encoded data z and the entropy of the probability distribution of the residual data The data transfer amount T can be minimized by taking the sum of (H( _pr )) as a common loss function and learning the first encoder unit 110 and the second encoder unit 120 to minimize it.

In addition, when considering the lower limit of the joint loss function, the data compression apparatus 100 instead of the sum of the entropy (H(p _z )) of the encoded data z and the entropy (H( _pr )) of the residual data, the encoded data z The sum of the entropy of data z (H(p _z )), the probability distribution of actual residual data, and the cross-entropy of the probability distribution of residual data predicted by the probability distribution predictor 140 is used as the loss function , the first encoder unit 110 and the second encoder unit 120 may be trained to minimize this.

Hereinafter, a process of compressing and restoring data will be described in detail.

3 is a block diagram conceptually showing functions of a data compression system according to an embodiment of the present invention, and FIG. 4 is a block diagram conceptually showing a flow of data compression and restoration by the data compression system according to an embodiment of the present invention. am.

Referring to FIG. 3 , the data compression system 10 may include a data compression device 100 and a data restoration device 200 . Here, each function of the data compression device 100 is as described above with reference to FIG. 1 .

The data recovery apparatus 200 may include a data receiver 210 and a decoder 220 .

Referring further to FIG. 4 , the data compression apparatus 100 may receive data (x) to be compressed and loss-compress it by an encoder (ENC) corresponding to the first encoder 120 . In addition, the second encoder 130 loss-compresses the data (x) to be compressed from the original data by the encoder (ENC), and then losslessly compresses the residual data (r), which is the difference between the restored data restored by the decoder (DEC). And, by the probability distribution prediction unit 140, the residual data (r) can be encoded together with the predicted probability distribution data.

In FIG. 4, for convenience of explanation, it is assumed that the first encoder 120 is implemented as an autoencoder, but is not limited thereto. For example, the first encoder 120 may be implemented as a generative adversarial network (GAN).

Although not shown in FIG. 4 , the data compression apparatus 100 may further include a process of quantizing the lossy compressed data and entropy-coding the quantized data. Here, entropy coding is a coding method in which the length of a code representing a symbol is varied according to the probability of a symbol appearing, and methods such as Huffman coding, range coding, and arithmetic coding may be used.

The data receiver 210 may receive compressed data from the data compression device 100 . Here, the method for the data receiving unit 210 to receive the compressed data from the data compression device 100 may be received using wired or wireless communication, and when the data recovery system 10 is implemented in one device, the internal Compressed data may be received through signaling. A method for the data receiving unit 210 to obtain compressed data from the data compression device 100 is not limited thereto.

Here, the compressed data received by the data receiving unit 210 may be original data loss-compressed data, residual data loss-free compression data, and data about a predicted probability distribution of the residual data.

)로 복원할 수 있다. 또한, 디코더부(220)는 잔차 데이터가 무손실 압축된 데이터 및 잔차 데이터의 예측된 확률 분포에 대한 데이터를 이용하여, 무손실 디코더(Lossless Decoder)에 의해 디코딩하여, 잔차 데이터(r)를 복원할 수 있다. 최종적으로 디코더부(220)는 손실 데이터(

)와 잔차 데이터(r)를 더하여 원본 데이터(x)를 복원할 수 있다.The decoder unit 220 converts the received original data to lossy compressed data using a decoder (DEC), and loses data (

) can be restored. In addition, the decoder unit 220 may restore the residual data r by decoding the residual data by a lossless decoder using lossless compressed data and data about a predicted probability distribution of the residual data. there is. Finally, the decoder unit 220 loses data (

) and the residual data (r), the original data (x) can be restored.

5(a) shows the result of compressing and restoring data having an image size of 512x512 using a plurality of data compression methods, and FIG. 5(b) shows data having a size of 4K UHD using a plurality of data compression methods. Shows the result of compression and restoration using .

Referring to (a) and (b) of FIG. 5 , compared to other data compression methods, the data compression method according to the embodiment of the present invention (Ours) takes time to perform compression and restoration regardless of the amount of data to be compressed. (Encoding+Decoding Time) can also be confirmed as the shortest. In addition, compared to other data compression methods, the data compression method according to the embodiment of the present invention (Ours) has the lowest number of bits per pixel regardless of the amount of data to be compressed, so the proposed data compression method It can be seen that this shows the most efficient compression ratio.

6 is a block diagram illustrating a data compression device according to an embodiment of the present invention in terms of hardware.

1 and 6, the data compression device 100 includes a storage device 191 for storing at least one command, a processor 192 for executing at least one command of the storage device 191, and a transmitting and receiving device. 193, an input interface device 194 and an output interface device 195.

Each of the components 191 , 192 , 193 , 194 , and 195 included in the data compression device 100 are connected by a data bus 196 to communicate with each other.

The storage device 191 may include at least one of a memory or a volatile storage medium and a non-volatile storage medium. For example, the storage device 191 may include at least one of read only memory (ROM) and random access memory (RAM).

The storage device 191 may further include at least one command to be executed by the processor 192 to be described later, and may store a data loss compression method, parameters of a probability distribution, and the like input from a user through the input interface device 194. there is.

The processor 192 may be a central processing unit (CPU), a graphics processing unit (GPU), a micro controller unit (MCU), or a dedicated processor on which methods according to embodiments of the present invention are performed. can mean

Referring further to FIG. 1 , as described above, the processor 192 operates the first encoder unit 120, the second encoder unit 130, and probability distribution prediction by at least one program command stored in the storage device 191. The functions of the unit 140 may be performed, and each of these may be stored in a memory in the form of at least one module and executed by a processor.

The transmitting/receiving device 193 may receive or transmit data from an internal device or an external device connected through communication, and may perform the function of the input unit 110 .

The input interface device 194 may receive at least one control signal or set value from a user. For example, the input interface device 194 may receive a user input such as a data transmission command or a compression command.

The output interface device 195 may output and visualize at least one piece of information including a compression rate for compressing data by the operation of the processor 192 .

7 is a flowchart illustrating a data compression method according to an embodiment of the present invention.

Referring further to FIG. 7 , first, the transmitting/receiving device 193 may obtain data (S100).

Subsequently, the processor 192 may generate first data obtained by loss-compressing the obtained data (S200).

In addition, the processor 192 may predict a plurality of parameters of the probability distribution of the residual data by using the pre-learned probability distribution prediction neural network by using the residual data of the data as an input (S300).

Subsequently, the processor 192 may generate second data obtained by losslessly compressing the residual data using a plurality of parameters of the predicted residual data (S400).

As such, since the data compression apparatus and method according to an embodiment of the present invention can losslessly compress original data while dramatically reducing the amount of compressed data, high-quality data can be transmitted in the shortest time, and various types of information can be stored as data. In the rapidly accumulating next-generation network environment, it has the effect of realizing high value-added application services such as ultra-realistic remote reality. In addition, since the complexity of compressing and restoring data is low, there is an effect of shortening the time for compressing and restoring data.

Combinations of each block of the block diagram and each step of the flowchart accompanying the present invention may be performed by computer program instructions. Since these computer program instructions may be loaded into an encoding processor of a general-purpose computer, special-purpose computer, or other programmable data processing equipment, the instructions executed by the encoding processor of the computer or other programmable data processing equipment are each block or block diagram of the block diagram. Each step in the flow chart creates means for performing the functions described. These computer program instructions may also be stored in a computer usable or computer readable memory that can be directed to a computer or other programmable data processing equipment to implement functionality in a particular way, such that the computer usable or computer readable memory It is also possible for the instructions stored in to produce an article of manufacture containing instruction means for performing the function described in each block of the block diagram or each step of the flow chart. The computer program instructions can also be loaded on a computer or other programmable data processing equipment, so that a series of operational steps are performed on the computer or other programmable data processing equipment to create a computer-executed process to generate computer or other programmable data processing equipment. It is also possible that the instructions performing the processing equipment provide steps for executing the functions described in each block of the block diagram and each step of the flowchart.

Further, each block or each step may represent a module, segment or portion of code that includes one or more executable instructions for executing specified logical function(s). It should also be noted that in some alternative embodiments, it is possible for the functions recited in the blocks or steps to occur out of order. For example, two blocks or steps shown in succession may in fact be performed substantially concurrently, or the blocks or steps may sometimes be performed in reverse order depending on their function.

The above description is merely an example of the technical idea of the present invention, and various modifications and variations can be made to those skilled in the art without departing from the essential qualities of the present invention. Therefore, the embodiments disclosed in the present invention are not intended to limit the technical idea of the present invention, but to explain, and the scope of the technical idea of the present invention is not limited by these embodiments. The protection scope of the present invention should be construed according to the claims below, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

10: Data Compression System
100: data compression device
200: data restoration device

Claims (11)

an input unit for obtaining data;
a first encoder unit generating first data obtained by lossy compression of the data;
a probability distribution prediction unit pre-learned to predict a plurality of parameters for a probability distribution of the residual data by taking residual data obtained by subtracting loss-reconstructed data from the data as an input; and
A second encoder unit generating second data obtained by lossless compression of the residual data using a plurality of parameters of the predicted residual data,
The first encoder unit and the second encoder unit are pre-learned so that the sum of the size of the first data and the size of the second data has a minimum value,
data compression device. According to claim 1,
The first encoder unit
At least one of a generative adversarial network and an autoencoder
data compression device. According to claim 1,
The probability distribution prediction unit,
Pre-learned to predict a plurality of parameters of the probability distribution of the residual data for learning using residual data for learning as input and a plurality of parameters for the probability distribution of the residual data for learning as labels
data compression device. According to claim 3,
The probability distribution of the residual data for learning is
At least one of a Gaussian distribution, a Poisson distribution, and a logistic distribution;
The plurality of parameters are
is a multi parameters mixture of the probability distribution.
data compression device. According to claim 1,
The probability distribution prediction unit
At least one of a convolutional neural network and a recurrent neural network
data compression device. According to claim 1,
The first encoder unit and the second encoder unit
Pre-learned to minimize the sum of the entropy of the first data and the entropy of the second data through joint-optimization.
data compression device. According to claim 1,
The first encoder unit and the second encoder unit,
Pre-learned to minimize the sum of cross-entropy between the entropy of the first data and a plurality of parameters for the probability distribution of the second data and the predicted residual data.
data compression device. an input unit for obtaining data;
a first encoder unit generating first data obtained by lossy compression of the data;
a probability distribution prediction unit pre-learned to predict a plurality of parameters for a probability distribution of the residual data by taking residual data of the data as an input;
a second encoder unit generating second data obtained by lossless compression of the residual data using a plurality of parameters of the predicted residual data; and
A decoder unit for obtaining and restoring the first data and the second data, respectively, and restoring the data without loss by adding loss-restored data from the first data and residual data restored from the second data;
The first encoder unit and the second encoder unit are pre-learned so that the sum of the size of the first data and the size of the second data has a minimum value,
data compression system. A method of compressing data performed by a data compression system, comprising:
acquiring the data;
generating first data obtained by lossy compression of the obtained data;
predicting a plurality of parameters of a probability distribution of the residual data by using residual data of the data as an input using a pre-learned probability distribution prediction neural network; and
Generating second data obtained by lossless compression of the residual data using a plurality of parameters of the predicted residual data,
The sum of the size of the first data and the size of the second data is pre-learned to have a minimum value.
Data compression method. A computer-readable recording medium storing a computer program,
The computer program,
Including instructions for causing a processor to perform a method of compressing data;
The method,
acquiring the data;
generating first data obtained by lossy compression of the obtained data;
predicting a plurality of parameters of a probability distribution of the residual data by using residual data of the data as an input using a pre-learned probability distribution prediction neural network; and
Generating second data obtained by lossless compression of the residual data using a plurality of parameters of the predicted residual data,
The sum of the size of the first data and the size of the second data is pre-learned to have a minimum value.
A computer-readable recording medium. A computer program stored on a computer readable recording medium,
The computer program,
Including instructions for causing a processor to perform a method of compressing data;
The method,
acquiring the data;
generating first data obtained by lossy compression of the obtained data;
predicting a plurality of parameters of a probability distribution of the residual data by using residual data of the data as an input using a pre-learned probability distribution prediction neural network; and
Generating second data obtained by lossless compression of the residual data using a plurality of parameters of the predicted residual data,
The sum of the size of the first data and the size of the second data is pre-learned to have a minimum value.
computer program.

Images