KR102317205B1 - Method and apparatus for estimating parameters of compression algorithm - Google Patents

Abstract

In a method for reconstructing compressed data, a bit stream of data is divided into groups each including a plurality of bits, a frequency distribution of values corresponding to each of the groups is generated, and the frequency distribution is Determining a compression algorithm used when compressing data based on, obtaining a plurality of binary images from a bit stream, summing a plurality of binary images to obtain a final image, and inputting the final image into a neural network to The determined parameters of the compression algorithm may be estimated, and the data may be decompressed and restored using the parameters.

Description

Method and apparatus for restoring compressed data {Method and apparatus for estimating parameters of compression algorithm}

The present disclosure relates to a method and apparatus for recovering compressed data.

Various data compression algorithms are being developed to compress data in order to efficiently write the data to less storage space.

On the other hand, the compressed data may include a header containing information on the type of compression algorithm used for data compression and the parameter values set at the time of compression. However, if the header is damaged, the compressed data cannot be decompressed. have.

In the past, there was a method of attempting decompression through a combination of various compression algorithms and possible parameters of each compression algorithm until decompression using brute-force search to solve this problem. There was a limitation in that a large amount of calculation and a long time were required.

Various embodiments are directed to providing a method and apparatus for restoring compressed data. The technical problems to be achieved by the present disclosure are not limited to the technical problems as described above, and other technical problems may be inferred from the following embodiments.

According to an aspect of the present disclosure, there is provided a method for reconstructing compressed data, the method comprising: dividing a bit stream of the data into groups each including a plurality of bits; generating a frequency distribution of values corresponding to each of the groups; determining a compression algorithm used when compressing the data based on the frequency distribution; obtaining a plurality of binary images from the bit stream; obtaining a final image by summing the plurality of binary images, and estimating the determined parameters of the compression algorithm by inputting the final image into a neural network; and decompressing and restoring the data using the parameter.

In addition, the generating may include: generating an array including values from 0 to N (N is a natural number) based on the number of bits included in each of the groups; and storing the frequency number of values corresponding to each of the groups in the array. may further include.

In addition, in the determining step, a ratio value of the sum of the frequencies of values within a preset section among the values corresponding to each of the groups to the sum of the frequencies of values corresponding to each of the groups is a threshold value. It can be determined whether the value is less than or not.

In addition, in the determining step, if the ratio is less than the threshold, it is determined that the compression algorithm used for compressing the data corresponds to the LZ (Lempel-Ziv) 77 compression algorithm, and the parameter is a search buffer ) may correspond to the size value of

In addition, the obtaining may include: dividing the bit stream into 1*M (M is a natural number) dimensional vectors each including a plurality of bits; and converting each of the vectors into m*m (m is a natural number) dimensional vectors to obtain the plurality of binary images. may include.

Also, the final image may be determined based on the number of bits included in each of the vectors and the determined parameter of the compression algorithm.

According to another aspect of the present disclosure, in a computer-readable recording medium recorded with a program for implementing a method of restoring compressed data, the method includes a plurality of bit streams of the data. dividing the bits into groups comprising bits; generating a frequency distribution of values corresponding to each of the groups; determining a compression algorithm used when compressing the data based on the frequency distribution; obtaining a plurality of binary images from the bit stream; obtaining a final image by summing the plurality of binary images, and estimating the determined parameters of the compression algorithm by inputting the final image into a neural network; and decompressing and restoring the data using the parameter.

According to another aspect of the present disclosure, in combination with hardware, in a computer program stored in a medium to execute a method for restoring compressed data, the method includes: each of a plurality of bit streams of the data dividing the bits into groups comprising bits; generating a frequency distribution of values corresponding to each of the groups; determining a compression algorithm used when compressing the data based on the frequency distribution; obtaining a plurality of binary images from the bit stream; obtaining a final image by summing the plurality of binary images, and estimating the determined parameters of the compression algorithm by inputting the final image into a neural network; and decompressing and restoring the data using the parameter.

According to another aspect of the present disclosure, an apparatus for restoring compressed data includes: a memory for storing one or more instructions; and dividing a bit stream of compressed data into groups each comprising a plurality of bits by executing the one or more instructions, generating a frequency distribution of values corresponding to each of the groups, and determining a compression algorithm used in compressing the data based on a frequency distribution, obtaining a plurality of binary images from the bit stream, summing the plurality of binary images to obtain a final image, and the final image and a processor for estimating the determined parameters of the compression algorithm by inputting ? to the neural network, and decompressing and restoring the data using the parameters.

Also, the processor generates an array including values from 0 to N (N is a natural number) based on the number of bits included in each of the groups, and a value corresponding to each of the groups on the array It is possible to store the number of frequencies.

In addition, the processor may be configured such that a ratio value of a sum of the frequency numbers of values within a preset section among the values corresponding to each of the groups to the sum of the frequencies of values corresponding to each of the groups is less than the threshold value. It can be determined whether or not

In addition, when the ratio is less than the threshold, the processor determines that the compression algorithm used for compressing the data corresponds to the LZ (Lempel-Ziv) 77 compression algorithm, and the parameter is It may correspond to a size value.

In addition, the processor divides the bit stream into 1*M (M is a natural number) dimensional vectors each including a plurality of bits, and divides each of the vectors into an m*m (m is a natural number) dimensional vector. , to obtain the plurality of binary images.

Also, the final image may be determined based on the number of bits included in each of the vectors and the determined parameter of the compression algorithm.

1 is a block diagram illustrating an example of a hardware configuration of an apparatus for restoring compressed data.
2 is a flowchart illustrating an example of a method of restoring compressed data.
3 is an example of a schematic diagram of a frequency distribution generated from uncompressed data.
4 is a schematic example of a frequency distribution generated from data compressed by the LZ77 compression algorithm.
5 is a diagram illustrating an example of obtaining a final image from compressed data.
6 is a diagram illustrating an example of a convolutional neural network applied to a method of estimating a parameter of a compression algorithm.
7 is a diagram illustrating an example of a support vector machine (SVM) classifier applied to a method of estimating a parameter of a compression algorithm.

Hereinafter, exemplary embodiments will be described in detail with reference to the accompanying drawings. It goes without saying that the following description is only for specifying the embodiments and does not limit or limit the scope of the invention. What can be easily inferred by an expert in the art from the detailed description and examples is construed as belonging to the scope of the right.

As used herein, terms such as 'consisting' or 'comprising' should not be construed as necessarily including all of the various components or various steps described in the specification, and some components or some steps thereof. It should be construed that they may not be included, or may further include additional components or steps.

Also, terms including an ordinal number such as 'first' or 'second' used in this specification may be used to describe various elements, but the elements should not be limited by the terms. The above terms are used only for the purpose of distinguishing one component from another.

The terms used in this specification have been selected as currently widely used general terms as possible while considering the functions in the present invention, which may vary depending on the intention or precedent of a person skilled in the art, the emergence of new technology, and the like. In addition, in a specific case, there is a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the corresponding invention. Therefore, the term used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than the name of a simple term.

The present embodiments relate to a method and an apparatus for estimating parameters of a compression algorithm, and detailed descriptions of matters widely known to those of ordinary skill in the art to which the following embodiments belong will be omitted.

1 is a block diagram illustrating an example of a hardware configuration of an apparatus for restoring compressed data.

The apparatus 100 for restoring compressed data may be implemented with various types of devices such as a personal computer (PC), a server device, a mobile device, and an embedded device. Furthermore, the apparatus 100 for restoring compressed data may correspond to a dedicated hardware accelerator mounted on the above device, and the apparatus 100 for restoring compressed data is a dedicated module for driving a neural network. It may be a hardware accelerator such as a neural processing unit (NPU), a tensor processing unit (TPU), a neural engine, or the like, but is not limited thereto.

Referring to FIG. 1 , the apparatus 100 for restoring compressed data includes a processor 110 and a memory 120 . Only the components related to the present embodiments are shown in the apparatus 100 for restoring compressed data shown in FIG. 1 . Accordingly, it is apparent to those skilled in the art that the apparatus 100 for restoring compressed data may further include other general-purpose components in addition to the components shown in FIG. 1 .

The processor 110 serves to control overall functions for executing the apparatus 100 for restoring compressed data. For example, the processor 110 generally controls the apparatus 100 for restoring compressed data by executing one or more instructions or programs stored in the memory 120 in the apparatus 100 for restoring compressed data. . The processor 110 may be implemented as a central processing unit (CPU), a graphics processing unit (GPU), an application processor (AP), etc. provided in the apparatus 100 for restoring compressed data, but is not limited thereto.

The memory 120 is hardware for storing various data processed in the apparatus 100 for restoring compressed data, for example, the memory 120 is data processed by the apparatus 100 for restoring compressed data. and data to be processed. In addition, the memory 120 may store applications, drivers, etc. to be driven by the apparatus 100 for restoring compressed data. The memory 120 includes random access memory (RAM), such as dynamic random access memory (DRAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), CD- It may include ROM, Blu-ray or other optical disk storage, a hard disk drive (HDD), a solid state drive (SSD), or flash memory 120 .

The processor 110 may determine a compression algorithm used when data is compressed from the compressed data. The determined compression algorithm may correspond to, for example, a Lempel-Ziv (LZ) 77 compression algorithm.

When it is determined that the compression algorithm used during compression corresponds to the LZ77 compression algorithm, the processor 110 may estimate the parameters of the compression algorithm. As an embodiment, the processor 110 may estimate a parameter of a compression algorithm using a neural network. A neural network refers to an overall model that has problem-solving ability by changing the strength of synaptic bonding through learning in which artificial neurons that form a network through synaptic bonding.

Specifically, the processor 110 may effectively estimate the parameters of the compression algorithm using a deep neural network, which is a neural network structure having a plurality of hidden layers.

As another embodiment, the processor 110 may estimate parameters of the LZ77 compression algorithm by inputting feature values calculated from the compressed data into a support vector machine (SVM) classifier.

The LZ77 compression algorithm corresponds to a lossless compression algorithm, and when the LZ77 compression algorithm is applied, the user can set two parameters. The first parameter is the size of the search buffer, and the second parameter is the size of the lookahead buffer.

The LZ77 compression algorithm can compress data using a search buffer and a look-ahead buffer. For example, some data of uncompressed original data may be sequentially stacked in the lookahead buffer, and some data of compressed data may be sequentially stacked in the search buffer. The LZ77 compression algorithm may compare data present in each of the look-ahead buffer and the search buffer with each other, and store information on the start position of overlapping data between the two buffers and the number of overlapping codes in binary form. In this case, the length of bits expressed in binary numbers may be determined by Equation 1 below according to the size of the search buffer.

In Equation 1, S is the size of the search buffer, and L may correspond to the length of bits expressed when binarizing information about a start position of overlapping data and the number of overlapping codes. For example, if the size of the search buffer ^{is set to S=1024=2 10} , the L value may be 11, and information about the start position of overlapping data and the number of overlapping codes may be expressed in 11 bits. have.

When there is no overlapping data between data in each of the lookahead buffer and the search buffer, the LZ77 compression algorithm can binarize and store only one sign located at the front of the lookahead buffer. At this time, since the code is encoded based on the ASCII table, it can be expressed in 8 bits.

Decompression of the LZ77 compression algorithm may be performed in reverse of the above-described compression process. In order to perform decompression, it is necessary to know S corresponding to the size of the search buffer set when the compression algorithm is applied. If the size of the search buffer, S, is known, the value corresponding to L can be known, and accordingly, when decompressing, it is possible to determine how many bits of compressed data to cut and read.

Meanwhile, in the compressed data, there may be a header containing information on the type of compression algorithm used for data compression and parameter values set at the time of compression. However, there is a problem in that compressed data cannot be decompressed when the header is damaged due to a hacker attack, damage to a storage medium, data loss during data transmission, or the like.

In the present invention, when there is no header information, the type of compression algorithm used for data compression is determined, and when it is determined that the compression algorithm used corresponds to the LZ77 compression algorithm, the size of the search buffer is estimated using a neural network. suggest a way

Specifically, the processor 110 may divide a bit stream of compressed data into groups each including a plurality of bits, and generate a frequency distribution of values corresponding to each of the groups. The processor 110 may determine a compression algorithm used when data is compressed based on the frequency distribution.

Also, when it is determined that the compression algorithm used during compression corresponds to the LZ77 compression algorithm, the processor 110 may acquire a plurality of binary images from the bit stream of the compressed data.

The processor 110 may obtain a final image by summing a plurality of binary images, and may estimate a parameter of the determined compression algorithm by inputting the final image to the neural network.

Meanwhile, the processor 110 may train the neural network using learning data in various combinations of the number of bits included in each of the vectors and the parameters of the compression algorithm. The training data may be composed of a final image generated according to a combination of the number of bits included in each of the vectors and a parameter of the compression algorithm, and a label value indicating the size of a parameter set when the compression algorithm is applied.

As another embodiment, the processor 110 may train the SVM classifier using the training data in various combinations of feature values calculated from the bit stream of the compressed data and parameters of the compression algorithm. The training data may be composed of a feature vector including information on feature values calculated from a bit stream of compressed data and a label value indicating the size of a parameter set when a compression algorithm is applied. The SVM classifier corresponds to a non-stochastic binary linear classification model that determines which category the new data belongs to based on a given data set.

The processor 110 may decompress the compressed data using the estimated parameter.

2 is a flowchart illustrating an example of a method of restoring compressed data.

In operation 200 , the apparatus 100 for restoring compressed data may divide a bit stream of data into groups each including a plurality of bits.

For example, when the number of bits included in each of the groups is 8, the apparatus 100 for reconstructing compressed data may cut and read the bit stream of the compressed data by 8 bits.

In operation 210, the apparatus 100 for reconstructing compressed data may generate a frequency distribution of values corresponding to each of the groups.

The apparatus 100 for restoring compressed data may generate an array including values from 0 to N (N is a natural number) based on the number of bits included in each of the groups. When the number of bits included in each of the groups is n, ^{an array including values from 0 to 2 n} −1 may be generated. For example, when the number of bits included in each of the groups is 8, the apparatus 100 for reconstructing compressed data may generate an array including values from 0 to 255.

The apparatus 100 for restoring compressed data may store the frequency number of values corresponding to each of the groups in an array. For example, when a group including 8 bits corresponds to '00011010', the apparatus 100 for restoring compressed data may convert an 8-bit binary number '00011010' into a decimal value 26.

The apparatus 100 for restoring compressed data may increase a value stored in an element in an array corresponding to a corresponding decimal value by one. For example, when the 8-bit binary number '00011010' is converted into the decimal value 26, the apparatus 100 for restoring compressed data may increase the value of the element indexed by 26 in the array by 1.

This process can be repeated until the end of the input bit stream of the compressed data. When the bit length of the input bit stream is not a multiple of 8, the number of bits remaining after generating the last group may be less than 8. In this case, the remaining bits may be ignored.

The apparatus 100 for restoring compressed data may plot the frequency distribution with values from 0 to 255 as the x-axis and the values stored in each element in the array as the y-axis.

In operation 220 , the apparatus 100 for reconstructing compressed data may determine a compression algorithm used when compressing data based on a frequency distribution.

For example, the apparatus 100 for reconstructing compressed data is a value obtained by summing the frequencies of values within a preset section among values corresponding to each of the groups with respect to the sum of the frequencies of values corresponding to each of the groups. It can be determined whether the ratio of is less than a threshold value. Hereinafter, it will be described with reference to FIGS. 3 and 4 .

3 is an example of a schematic diagram of a frequency distribution generated from uncompressed data.

FIG. 3 is a graph in which a value stored in each element in the y-axis array is divided by the largest value among each element in the array and normalized to have a value between 0 and 1. Referring to FIG.

Referring to FIG. 3 , it can be seen that values from 10 to 129 exist with high frequency among values corresponding to each of the groups of the bit stream of uncompressed data. Referring to the ASCII Table, values from 10 to 129 are 'line feed', 'space', 'punctuation marks', 'Arabic numbers from 0 to 9 ( Digits)' and 'alphabet' correspond to symbols frequently used in general data or text files.

4 is a schematic example of a frequency distribution generated from data compressed by the LZ77 compression algorithm.

4 is a graph in which a value stored in each element in the y-axis array is divided by the largest value among each element in the array and normalized to have a value between 0 and 1. Referring to FIG.

4, the data compressed by the LZ77 compression algorithm has relatively various values from 0 to 255, and 0 is the highest frequency among values corresponding to each of the groups of the bit stream. Able to know.

Accordingly, the apparatus 100 for reconstructing compressed data is a ratio of a sum of the frequencies of values within a preset section among values corresponding to each of the groups to the sum of the frequencies of values corresponding to each of the groups. When the value is less than the threshold, it may be determined that the compression algorithm used for data compression corresponds to the LZ77 compression algorithm. The ratio value D(β) can be expressed as in Equation 2 below.

In Equation 2, β corresponds to a bit stream of compressed data, l corresponds to elements included in the generated array, and elements correspond to values from 0 to N (N is a natural number). d _l (β) is a value stored in the l-th element of the array, and may mean the frequency of a value corresponding to the l-th element of the array among values corresponding to each of the groups. Equation 1 indicates a ratio value when the number of bits included in each group is 8, and the ratio value D(β) is not limited to Equation 1.

As shown in Equation 3 below, when the ratio value D(β) is smaller than the threshold value α, it can be determined that the compression algorithm used for data compression corresponds to the LZ77 compression algorithm.

On the other hand, when the ratio value D(β) is equal to or greater than the threshold value α as shown in Equation 4 below, it may be determined that the compression algorithm used for data compression does not correspond to the LZ77 compression algorithm.

Referring back to FIG. 2 , the apparatus 100 for reconstructing compressed data in operation 230 may obtain a plurality of binary images from a bit stream.

For example, the apparatus 100 for reconstructing compressed data may divide the bit stream into 1*M (M is a natural number) dimensional vectors each including a plurality of bits. Also, the apparatus 100 for reconstructing compressed data may obtain a plurality of binary images by converting each of the vectors into m*m (m is a natural number) dimensional vectors.

5 is a diagram illustrating an example of obtaining a final image from compressed data.

Referring to FIG. 5 , the apparatus 100 for reconstructing compressed data may divide the compressed data 500 into 1*M-dimensional vectors 510 each including M bits. In this case, the 1*M-dimensional vector 510 may be expressed as in Equation 5 below.

In Equation 5, K(i) may correspond to the i-th 1*M-dimensional vector 510 . b _j (i)(j∈{1,2...M}) are bits included in the 1*M-dimensional vector 510 and may correspond to 0 or 1.

The number of 1*M-dimensional vectors 510 generated through the division process may be N in total. When the total number of bits included in the compressed data 500 is not a multiple of M, the number of bits remaining after generating the N-th 1*M-dimensional vector 510 may be smaller than M. In this case, the remaining bits may be ignored.

Next, the apparatus 100 for reconstructing compressed data may generate a plurality of binary images by converting each of the vectors into m*m (m is a natural number) dimensional vectors.

The apparatus 100 for reconstructing compressed data may deform the structure so that each of the generated N 1*M-dimensional vectors 510 has a two-dimensional arrangement. In this case, the two-dimensional array may correspond to the m*m-dimensional vector 520, and is expressed as in Equation 6 below.

In Equation 6, B(i) may correspond to an m*m-dimensional vector 520 obtained when the i-th 1*M-dimensional vector 510 is transformed into a two-dimensional array structure. Referring to FIG. 5 , an m*m-dimensional vector 520 may correspond to a binary image.

Accordingly, N binary images may be generated from the N 1*M-dimensional vectors 510 .

Referring back to FIG. 2 , in step 240 , the apparatus 100 for reconstructing compressed data acquires a final image by summing a plurality of binary images, and inputs the final image to a neural network to estimate parameters of the determined compression algorithm can do.

For example, the apparatus 100 for reconstructing compressed data may add N binary images generated from N 1*M-dimensional vectors 510 as shown in Equation 7 below.

은 복수의 이진 영상들 각각에 해당하는 m*m(m은 자연수)차원의 벡터들의 동일 위치에 있는 요소(element)들끼리 더하는 요소 별 합 연산자일 수 있다. Z는 N개의 이진 영상들을 모두 합하여 생성되는 이진 영상(530)에 해당할 수 있다.Above, in Equation 7

may be an element-by-element sum operator that adds elements at the same position of m*m (m is a natural number) dimensional vectors corresponding to each of the plurality of binary images. Z may correspond to the binary image 530 generated by summing all N binary images.

Also, the apparatus 100 for reconstructing compressed data may normalize the generated binary image 530 Z by Equation 8 below.

In Equation 8, I corresponds to the final image 540 obtained when the binary image 530 and Z are normalized. Through normalization, each of the values included in the vector corresponding to I of the final image 540 may have a value between 0 and 1. w may mean the smallest value among the values included in the vector corresponding to Z of the binary image 530 , and x may mean the largest value among values included in the vector corresponding to Z of the binary image 530 .

The apparatus 100 for reconstructing compressed data may estimate parameters of the determined compression algorithm by inputting the final image to the neural network. For example, the apparatus 100 for restoring compressed data may estimate a size value of a search buffer set when the LZ77 compression algorithm is applied.

6 is a diagram illustrating an example of a convolutional neural network applied to a method of estimating a parameter of a compression algorithm.

Referring to FIG. 6 , an example of a convolutional neural network for estimating the size of a search buffer set when a compression algorithm is applied by inputting a final image generated from compressed data is illustrated.

In FIG. 6 , the final image input to the convolutional neural network corresponds to an image obtained when compressed data is divided into 1*900 dimensional vectors each including 900 bits. Accordingly, the final input image may correspond to a 30*30 dimensional vector.

A convolutional neural network may include three convolutional layers and one fully connected layer.

In the first convolution layer, feature maps may be generated by performing a convolution operation between the input final image and four different kernels having a size of 5X5. In this case, the number of generated feature maps may be the same as the number of filters used. An activation map may be generated by applying an activation function to each of the four generated feature maps. The activation function may correspond to a Rectified Linear Unit function (ReLU), but is not limited thereto, and may correspond to a step function, a sigmoid function, or the like. Meanwhile, the ReLU function can be expressed as in Equation 9 below.

In addition, pooling may be performed using a kernel having a size of 2*2 for each of the four generated activation maps. Pooling may be performed to reduce the size of data of the activation map, which is a result of the convolution operation, and a specific value may be extracted from values in the feature map in which the kernel is located during the pooling operation. An example of pooling to simplify information is max-pooling. In this case, the maximum value can be extracted from values in the feature map in which the kernel is located during the pooling operation.

Specifically, the maximum pooling operation may move the 2*2 kernel in a zigzag manner on the activation map, and extract the largest value among 4 pixels existing in the 2*2 kernel as a representative value. As a result, the size of the activation map can be reduced from 30*30 to 15*15.

In the second convolution layer, a convolution operation can be performed using 8 different filters with a size of 5*5 for 4 activation maps of size 15*15. As a result, the activation map can be generated by applying the ReLU function to each of the generated eight feature maps. Max pooling may be performed using a filter having a size of 2*2 for each of the generated eight activation maps. The size of the activation map that has been max-pooled can be reduced from 15*15 to 8*8.

In the third convolution layer, a feature map can be generated by performing a convolution operation on eight 8*8 activation maps using eight different filters having a size of 5*5. An activation map can be generated by applying the ReLU function to each of the eight generated feature maps. Finally, 8 activation maps of size 8*8 are generated, which can be converted into 512 (=8*8**8) one-dimensional vectors.

Meanwhile, a dropout layer may be added between the third convolutional layer and the fully connected layer. The drop-out layer is to prevent the neural network from overfitting the training data during the training process.

In the fully connected layer, the 512 generated nodes can be fully connected with F output nodes. In this case, the number F of output nodes may correspond to the number of different search buffer sizes that the convolutional neural network can estimate.

In FIG. 6, F is set to 4, so the number of different search buffer sizes that the convolutional neural network can estimate is 4. For example, the sizes of the different search buffers may be 1024, 2048, 4096, and 8192, respectively. In this case, the size of the search buffer 1024 may be set to class 0, the size of the search buffer 2048 may be set to class 1, the size of the search buffer 4096 may be set to class 2, and the size of the search buffer 8192 may be set to class 3. And in the training process of the convolutional neural network, the classes corresponding to the size of each search buffer are [1, 0, 0 0], [0, 1, 0, 0], [0, 0, 1, 0], [0]. , 0, 0, 1] may be encoded and used as a one-hot vector.

Activation values output from the fully connected layer may be normalized with a softmax function. The softmax function is as shown in Equation 10 below.

는 클래스 p에 대응되는 정규화된 활성화 값이며,

y_p는 클래스 p에 대응되는 정규화되기 전의 활성화 값이다..In Equation 10 above

is the normalized activation value corresponding to class p,

y _p is the activation value before normalization corresponding to class p.

Meanwhile, a loss function may be set to train the convolutional neural network. The loss function may correspond to any one of a mean squared error and a cross entropy error, but is not limited thereto. The cross entropy error can be expressed as Equation 11 below.

In Equation 11, E is the cross entropy error value, and T is the total number of classes. Y _t may mean the value of the t-th element in the vector when the class is expressed as a one-hot vector.

The convolutional neural network can find parameter values that make the value of the loss function as small as possible during the training process. In this case, the convolutional neural network may repeat the process of calculating the differential value of the parameter and updating the parameter value based on the calculated differential value.

On the other hand, as another embodiment of estimating the determined parameters of the compression algorithm, the apparatus 100 for reconstructing compressed data includes a first feature value indicating a ratio of 0 and 1 in the bit stream of the compressed data and the same in the bit stream. A second characteristic value indicating the degree of continuity of bits may be calculated.

For example, when it is determined that the compression algorithm used during compression corresponds to the LZ77 compression algorithm, the apparatus 100 for restoring compressed data may calculate feature values from the bit stream of the compressed data.

The apparatus 100 for reconstructing compressed data may perform a first statistical test for calculating a first feature value and a second statistical test for calculating a second feature value.

In the first statistical test, the apparatus 100 for reconstructing compressed data may convert bit '0' present in the bit stream β of the compressed data to '-1', and bit '1 present in the bit stream ' can be converted to '+1'.

The apparatus 100 for reconstructing compressed data may add all values converted to '-1' or '+1' when the conversion process is completed for all bits in the bit stream β. For example, when the bit stream β is '1110', '1110' may be converted into '(+1)(+1)(+1)(-1)'. Therefore, the sum of the transformed values T _F may correspond to 2.

The apparatus 100 for reconstructing compressed data _{may calculate T F} /√N, which is a _{value obtained by dividing the sum of the transformed values T F} by the square root of the total length N of the bit stream β, as the first feature value. For example, when the bit stream β is '1110', since N is 4, the 1 feature value may be 1.

The second statistical test is a test of how successively the same bit appears in the bit stream β of the compressed data. On the other hand, the bit stream β of the compressed data is a 1*N-dimensional vector [β _1, β _2, β _{3, ??} ,β _N ] can be expressed.

The apparatus 100 for reconstructing compressed data may calculate R(β)/√N as the second characteristic value, and R(β) may be defined as in Equation 12 below.

In Equation 12, r(n) may be defined as Equation 13 below.

번째 비트 값에 해당할 수 있다. r(n)은 연속한 두 비트의 값이 동일할 경우에는 0, 연속한 두 비트의 값이 서로 다를 경우에는 1로 설정될 수 있다. In Equation 13, βn is the bit stream β of the compressed data.

It may correspond to the th bit value. r(n) may be set to 0 when values of two consecutive bits are the same, and set to 1 when values of two consecutive bits are different from each other.

For example, if the input bit stream β is '111010', then r(1)=0, r(2)=0, r(3)=1, r(4)=1, r(5)=1 can be calculated. Therefore, R(β) is 4.

The apparatus 100 for reconstructing compressed data may calculate R(β)/√N, which is a value obtained by dividing R(β) by the square root of the total length N of the bit stream β, as the second feature value. For example, when the bit stream β is '111010', N is 6, so the second feature value may be about 2.449 (=4/√6).

The apparatus 100 for reconstructing compressed data may estimate the determined parameters of the compression algorithm by inputting the first and second feature values to the neural network.

7 is a diagram illustrating an example of a support vector machine (SVM) classifier applied to a method of estimating a parameter of a compression algorithm.

First, the apparatus 100 for restoring compressed data may compress data by varying the search buffer size of the LZ77 compression algorithm to generate various bit streams. In this case, a label value indicating the size of the search buffer set when the bit stream is generated and bit stream information may be recorded in the database.

When the construction of the database is completed, learning of the SVM classifier for estimating the parameters of the compression algorithm may proceed.

In the learning step, information including bit streams for training the SVM classifier and a label value indicating a search buffer size set when generating each of the bit streams may be obtained from the database. In this case, the bit streams may be input to the feature vector extractor. The feature vector extractor may perform a first statistical test and a second statistical test on the input bit streams. The feature vector extractor may output a feature vector including first feature value and second feature value information for each of the input bit streams. Information including the output feature vector and label value is input to the SVM classifier, and the SVM classifier is trained to classify the feature vectors according to the label value.

In the test phase, the generated SVM classifier that has been trained can be used. In the test phase, bit streams not used in the learning phase may be retrieved from the database. The bit streams may be input to the feature vector extractor, and a feature vector including the first feature value and the second feature value information may be output through the first statistical test and the second statistical test. The output feature vector is input to a trained SVM classifier, and the SVM classifier may output an estimated label value for the input feature vector, and obtain a size of a search buffer corresponding to the corresponding label value.

Referring back to FIG. 2 , in step 250 , the apparatus 100 for restoring compressed data may decompress the data using parameters to restore the data.

The convolutional neural network trained in the above manner may receive the final image and output a class corresponding to the search buffer size value. Alternatively, the SVM classifier learned in the above manner may receive a feature vector and output a class corresponding to a search buffer size value. The size of the corresponding search buffer can be known from the output class, and the original data before compression can be restored by decompressing the compressed data using this.

Meanwhile, the above-described embodiments of the present invention can be written as a program that can be executed on a computer, and can be implemented in a general-purpose digital computer that operates the program using a computer-readable recording medium. In addition, the structure of the data used in the above-described embodiment of the present invention may be recorded in a computer-readable recording medium through various means. The computer-readable recording medium includes a storage medium such as a magnetic storage medium (eg, a ROM, a floppy disk, a hard disk, etc.) and an optically readable medium (eg, a CD-ROM, a DVD, etc.).

So far, with respect to the present invention, the preferred embodiments have been looked at. Those of ordinary skill in the art to which the present invention pertains will understand that the present invention can be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments are to be considered in an illustrative rather than a restrictive sense. The scope of the present invention is indicated in the claims rather than the foregoing description, and all differences within the scope equivalent thereto should be construed as being included in the present invention.

Claims (14)

A method for restoring compressed data, comprising:
dividing the bit stream of data into groups each comprising a plurality of bits;
generating a frequency distribution of values corresponding to each of the groups;
determining a compression algorithm used when compressing the data based on the frequency distribution;
obtaining a plurality of binary images from the bit stream;
obtaining a final image by summing the plurality of binary images, and estimating the determined parameter of the compression algorithm by inputting the final image into a neural network; and
and decompressing and restoring the data using the parameter. The method of claim 1,
The generating step is
generating an array including values from 0 to N (N is a natural number) based on the number of bits included in each of the groups; and
storing the frequency number of values corresponding to each of the groups in the array; How to include more. 3. The method of claim 2,
The determining step is
A method of determining whether a ratio of the sum of the frequencies of values corresponding to each of the groups to the sum of the frequencies of the values corresponding to each of the groups is less than a threshold value among the values corresponding to each of the groups . 4. The method of claim 3,
The determining step is
If the ratio is less than the threshold,
It is determined that the compression algorithm used when compressing the data corresponds to the LZ (Lempel-Ziv) 77 compression algorithm,
The parameter corresponds to a size value of a search buffer. The method of claim 1,
The obtaining step is
dividing the bit stream into 1*M (M is a natural number) dimensional vectors each including a plurality of bits; and
converting each of the vectors into m*m (m is a natural number) dimensional vectors to obtain the plurality of binary images; How to include. 6. The method of claim 5,
The final image is
A method determined based on the number of a plurality of bits included in each of the vectors and the determined parameter of the compression algorithm. In a computer-readable recording medium in which a program for implementing a method of restoring compressed data is recorded,
The method is
dividing the bit stream of data into groups each comprising a plurality of bits;
generating a frequency distribution of values corresponding to each of the groups;
determining a compression algorithm used when compressing the data based on the frequency distribution;
obtaining a plurality of binary images from the bit stream;
obtaining a final image by summing the plurality of binary images, and estimating the determined parameters of the compression algorithm by inputting the final image into a neural network; and
and decompressing and restoring the data using the parameter. In combination with hardware, a computer program stored in a medium to execute a method for restoring compressed data,
The method is
dividing the bit stream of data into groups each comprising a plurality of bits;
generating a frequency distribution of values corresponding to each of the groups;
determining a compression algorithm used when compressing the data based on the frequency distribution;
obtaining a plurality of binary images from the bit stream;
obtaining a final image by summing the plurality of binary images, and estimating the determined parameter of the compression algorithm by inputting the final image into a neural network; and
and decompressing and restoring the data using the parameters. a memory that stores one or more instructions; and
By executing the one or more instructions, a bit stream of compressed data is divided into groups each comprising a plurality of bits, generating a frequency distribution of values corresponding to each of the groups, the frequency determine a compression algorithm used in compressing the data based on the distribution, obtain a plurality of binary images from the bit stream, obtain a final image by summing the plurality of binary images, and obtain the final image and a processor for estimating the determined parameters of the compression algorithm by input to a neural network, and decompressing and restoring the data using the parameters. 10. The method of claim 9,
The processor is
Create an array including values from 0 to N (N is a natural number) based on the number of bits included in each of the groups,
an apparatus for storing a frequency count of values corresponding to each of the groups in the array. 10. The method of claim 9,
The processor is
Apparatus for determining whether a ratio value of the sum of the frequencies of values corresponding to each of the groups to the sum of the frequencies of values corresponding to each of the groups is less than a threshold value among the values corresponding to each of the groups . 12. The method of claim 11,
The processor is
If the ratio is less than the threshold,
It is determined that the compression algorithm used when compressing the data corresponds to the LZ (Lempel-Ziv) 77 compression algorithm,
The parameter corresponds to a size value of a search buffer. 10. The method of claim 9,
The processor is
The bit stream is divided into 1 * M (M is a natural number) dimensional vectors each including a plurality of bits,
An apparatus for obtaining the plurality of binary images by converting each of the vectors into m*m (m is a natural number) dimensional vectors. 14. The method of claim 13,
The final image is
The apparatus is determined based on the number of the plurality of bits included in each of the vectors and the determined parameter of the compression algorithm.

Images