KR19990039267A - Structure and Compression / Restoration Method of Image Data Compression / Restoration Device Using Neural Network - Google Patents

Abstract

The present invention relates to a structure and a compression / restoration method of an image data compression / restoration apparatus using neural networks. In general, when processing all the image data using a single neural network, if the image content is significantly different from the sample used for the neural network training, the quality of the recovered image cannot be guaranteed. Therefore, the image data is classified according to the content. By separately employing neural networks that are in charge of processing image data, several neural networks are used in parallel. However, such a method of compressing and restoring image data applies only transform compression to the image data, and thus it is not possible to obtain a high quality of the compression ratio and the restored image quality of the image data. In order to solve this problem, in the present invention, the Karhunen-Loeve Transform (KLT) method, which is known as an optimal transform coding method, can be performed by a neural network, and the disadvantages of edge degradation can be compensated. Image data using neural networks that can compress and reconstruct image data requiring accuracy by using classification Karhnen-Loeve Transform (DKLT) method that combines classification vector quantization that performs specialized vector quantization according to characteristics. The structure and compression / restore method of the compression / restore device of is presented.

Description

Structure and Compression / Restoration Method of Image Data Compression / Restoration Device Using Neural Network

The present invention relates to a structure and a compression / restoration method of an image data compression / restoration device. In particular, a linear coding method, a Karhunen-Loeve Transform (hereinafter referred to as a KLT) method, and a classification with excellent recovery quality and compression ratio On the structure and compression / restore method of compression / restore device of image data compression / restore device using neural network that can improve recovery quality while maintaining high extrusion rate by using DKLT neural network, which is a KLT transform neural network with classification function combining quantization method will be.

In general, the KLT method is an encoding method that has superior characteristics to other linear transform functions in terms of energy concentration and correlation of transform coefficients. However, there is a problem in that a signal dependent characteristic in which the base matrix for transformation is changed according to the input data and the amount of computation is huge for the transformation. Due to the problems of KLT, the Discrete Cosine Transform (DCT) method, which has lower performance than KLT but enables high-speed conversion, is used for compression / restore of general image data requiring high speed while maintaining proper image quality. However, there is a problem that the DCT method is also not suitable for applications in which loss of image data is low while maintaining a high compression ratio such as a picture archiving and communication system (PACS). Accordingly, conventionally, a hybrid encoding method combining DCT and classification vector quantization or a transform encoding method using multiple KLT neural networks in parallel has been used for applications requiring accurate recovery images. The reason for employing the transform coding method that classifies an image in units of blocks and employs a KLT neural network in parallel with a separate basis matrix for each block type is as follows.

When compressing and restoring image data using a neural network, a compressed neural network having fewer output neurons than an input neuron is used for a compressor that compresses data and transmits the data to a channel. Also, in the reconstructor, as a reconstructed neural network corresponding to the compressed neural network, the number of input neurons of the reconstructed neural network is equal to the output neurons of the compressed neural network, and the number of output neurons of the reconstructed neural network is the same as the number of input neurons of the compressed neural network. The reconstructed neural network performs the inverse transformation function of the compressed neural network, and the compressed neural network and the reconstructed neural network are trained with a sample of input image data in advance. However, when all the image data is processed using one neural network, the image quality of the recovered image cannot be guaranteed if the image content is different from the sample used for the neural network training. For this reason, several neural networks must be configured in parallel by classifying image data according to its contents and employing neural networks that handle each type of image data separately.

In this case, when the transform coding method is applied to the KLT neural network, an input block is processed separately, and the output of each neural network is inversely transformed and restored in a block classifier and a compressor that separates the blocks, thereby selecting a neural network that can obtain the best reconstructed picture quality. An output selector is required. However, in this method, since the optimal energy concentration of the KLT transform coefficients is selected, an optimal hybrid coding characteristic cannot be provided because the optimal neural network is selected among a plurality of transform neural networks.

Accordingly, the present invention enables high-speed parallel processing of image data by determining the type of input image block and using the DKLT neural network for optimal transcoding, and adopts a vector quantizer specialized for each type of block. It is an object of the present invention to provide a structure and a compression / restore method of a compression / restoration device of image data using a neural network that can improve the quality of a recovery screen.

The structure of the apparatus for compressing and restoring image data using a neural network according to the present invention for achieving the above object is to compress the input image block data by inputting the vectorized image block data in the same dimension as the number of pixels in the image block. A plurality of classification Karhnen-Rube transformed neural networks for performing the Karhnen-Rube transformed to output the Karhnen-Lube transformed coefficients and the classification Karhnen-Rube transformed coefficients The codebook address of the vector having the most similar value to the input vector is compared with the coefficient vectors in the vector codebook by receiving each of the Karhnen-Rube transform coefficients output from the plurality of classification Karchenen-Lube transform neural networks. A plurality of vector quantizers to be output, and to each of the classification Karchenen-Rube transform neural networks. An output selector for generating an output selection control code by determining which classification Karhnen-Lube transform neural network is a vector of the input image block by receiving the classification Karhnen-Lube transform coefficients output from the The codebook contains the same codebook as the one used in the vector quantizer, receives the codebook address, outputs the Karchenn-Lube transform coefficients stored in the address, and corresponds to the vector quantizer. A vector table and an inverse Karchenen-Rube transform neural network that is paired with the plurality of vector tables and receives KLT coefficients output from the plurality of vector tables and inversely converts the outputted image block data. It features.

In addition, the compression / restoration method of the image data using the neural network according to the present invention for achieving the above object is when the image block data is input to the classification Karhnen-Lube transform neural network, Karhnen-Lube transform the input data and determine the block type Performing a linear discriminant analysis required for the step of outputting the Karhnen-Lube transform coefficients and the classification Karhnen-Lube transform coefficients, and inputting the output Karhnen-Lube transform coefficients to the respective vector quantizers. Determining a classification Karhnen-Rube transform neural network capable of optimally converting the input image block data by outputting Rube transform coefficients to an output selector and outputting a result, and vector quantization of the determined classification Karhnen-Lube transform neural network. Outputting a selection codebook address from the device; Transmitting the address of the vector quantizer designated by the output selection codebook and the codebook address to a remote location having a reconstructor through a communication channel, and a vector table to be used for image block recovery using the vector quantizer address and the inverse thereof. Specifying a Karhnen-Rube transform neural network, applying an output of a vector quantizer, the remainder of the compressed data, to the vector table to retrieve inverse quantized Karhnen-Lube transform coefficients, and the retrieved Karhnen- And inputting the Rube transform coefficients to the inverse Carchenne-Rube transform neural network, and inversely transforming the final block to output the recovered image block data.

1 is a structural diagram of an image data compression / restoration apparatus according to the present invention;

2 is a structural diagram of a classification Karchenen-Rube transform neural network for image data compression applied to the present invention.

3 is a structural diagram of an inverse Karchenen-Rube transform neural network for image data restoration applied to the present invention.

4 is a flowchart illustrating a training method of a Karhnen-Rube transform neural network.

5 is a flowchart illustrating a method of compressing and restoring video data according to the present invention.

<Description of the symbols for the main parts of the drawings>

11: compressor 12: restorer

13: communication channel

111-1 to 111-n: DKLT neural network 112: output selector

113-1 to 113-n: Vector quantizers 121-1 to 121-n: Vector table

122-1 to 122-n: inverse KLT neural network

Hereinafter, with reference to the accompanying drawings will be described in detail the present invention.

1 is a structural diagram of an image data compression / restoration apparatus according to the present invention. The image data compression / restoration apparatus is largely composed of a compressor 11 and a decompressor 12, and the compressor 11 determines a type of an input image block and performs a plurality of DKLT neural networks 111-1 to 111- that are in charge of KLT conversion. k), an output selector 112 and an output of the DKLT neural networks 111-1 to 111-k that combine the input image block discrimination results of the respective DKLT neural networks 111-1 to 111-k to determine a final discrimination result. And the KLT coefficients transformed by the compressor 11 and the vector tables 121-1 through 121-k for inverse transformation of the vector quantization. Is composed of inverse KLT transform neural networks 122-1 to 122-k. More detailed description is as follows.

First, referring to the configuration of the compressor 11, the DKLT neural networks 111-1 to 111-k perform KLT transformation, which is an optimal transformation for image compression, and output KLT transformation coefficients while pre-training the KLT transformation coefficients. This function also outputs the DKLT coefficients that can be used to determine the type of input image data that has been KLT-converted by projecting onto space. In addition, the vector quantizers 113-1 to 113-k are specialized to compress the result of the KLT transformation according to each type of image block, and the output selector 112 outputs the DKLT coefficients output from the respective DKLT neural networks. Selects an output from which the input image block data is optimally converted from the

The output of the compressor 11 is transmitted to the decompressor 12 via the communication channel 13, and the decompressor 12 corresponds to each vector quantizer of the compressor and is responsible for inverse quantization. 121-k) and inverse KLT neural networks 122-1 through 122-k that perform inverse KLT transformation of the KLT coefficients retrieved from the vector tables 121-1 through 121-k.

The DKLT neural networks 111-1 to 111-k receive an n-dimensional vector X in which image block data is vectorized, and output KLT coefficients and DKLT coefficients. Where n is the number of pixels in the video block. KLT coefficients and DKLT coefficients are also vectors, and the dimensions of these vectors vary depending on the compression performance and the number of DKLT neural networks 111-1 to 111-k. One vector quantizer 113-1 to 113-k exists at an output terminal of each of the DKLT neural networks 111-1 to 111-k, and the KLT coefficients output from the DKLT neural networks 111-1 to 111-k. Compares with the coefficient vectors in the vector codebook and outputs the codebook address of the vector having the most similar value to the input vector. Whether the codebook address is output is determined by the control code provided by the output selector. The output selector 12 receives the DKLT coefficients output from each of the DKLT neural networks 111-1 to 111-k, and the type of the vector of the input image block to be in charge of which DKLT neural networks 111-1 to 111-k. It determines whether the block is a block to generate an output selection control code. The compressed data transmitted to the decompressor 12 through the communication channel 13 is composed of an output control code provided from the output selector 13 and a codebook address output from the vector quantizers 113-1 to 113-k. . The decompressor 12 receives the compressed data, first checks the output control code, and designates the vector tables 121-1 to 121-k to be searched using the codebook address. The vector tables 121-1 to 121-k contain the same codebook as the contents used in the vector quantizers 113-1 to 113-k of the compressor 11, and receive a codebook address and store the codebook address in the address. Output the stored KLT coefficients. Each vector table 121-1 to 121-k is paired with an inverse KLT neural network 122-1 to 122-k that is responsible for inverse transformation of coefficients of the table. The inverse KLT neural networks 122-1 through 122-k receive the KLT coefficients output from the vector tables 121-1 through 121-k, and inversely transform the output KLT neural networks 122-1 through 122-k.

2 is a structural diagram of a classification Karchenen-Lube transform (DKLT) neural network for image data compression applied to the present invention.

The DKLT neural network is composed of four stages including the input stage, and each neural network is trained with test blocks corresponding to the type of the block to which each neural network is responsible for transformation. The trained DKLT neural networks determine whether the input image block is the type of block that is in charge of itself from the result of the transformation and also outputs the same. In this way, when the characteristics of the block are different from the training data used in the training, and the transform output is provided in two or more neural networks, the output selector finally determines the optimal output. Since the DKLT neural network is adopted for each classified block type, a classification vector quantization suitable for each block type can be used after transform coding. Referring to the drawings in more detail as follows. The number of neurons in the neural network is n (X1 to Xn), the dimension of the input image block, the number of neurons in the first hidden layer is m, and the number of neurons in the second hidden layer and the output layer is l. At this time, there is a relationship of n> m> l and the compression by the transformation is obtained by decreasing the number of coefficients n> m. The Y vector output from the first hidden layer is the KLT coefficient and the Z vector of the output layer is the DKLT coefficient. Each neuron simply outputs the product of the neuronal input plus the strength of the connection. In the DKLT neural network, the first layer is responsible for the KLT transformation, and the other two layers are responsible for the linear discriminant analysis required to determine the type of block.

3 is a structural diagram of an inverse Karchenen-Rube transform neural network for image data restoration applied to the present invention.

Inverse transformation is performed on the Y vector, the m-dimensional KLT coefficient, to output the n-dimensional recovered image block data X 'with n> m. The connection strength of the inverse KLT neural network uses the strength of the connection between the input layer and the first hidden layer in the training results of the DKLT neural network without any training.

4 is a flowchart illustrating a training method of a classification Karchenen-Rube transform neural network (DKLT).

The training of the DKLT neural network is performed step by step, using the Recursive Least Square (RLS) training method. First, the first layer connection strength between the input layer and the first hidden layer is initialized to a small random value and the training parameters for the RLS training are initialized (401). Next, input the training image block data composed of the representative image block data of the type of DKLT neural network to be trained to obtain the connection strength of the first layer by adjusting the connection strength according to the RLS training method (402). When the connection strength is obtained, the output from the first hidden layer is obtained by inputting the entire training data including all kinds of image block data (403). The output of the first hidden layer obtained in this way is divided into two data sets by dividing the output of the image block data of the type to be trained by the target DKLT neural network and the output of the other type of image block data into two data sets, and then combining the two data sets. The average value of and the average value of each data set are calculated to generate a total distributed data set and a class distributed data set (404). A globally distributed data set is a data set that combines the data of each data set from two data sets minus the average value of the entire data set into one data set. It is a data set that collects the data minus the average of the data set into one data set. When the data set is prepared, the second layer connection strength is first initialized to a small random value and the training parameters for RLS training are initialized (405). After that, the connection strength of the second layer is obtained using the class distribution data set according to the same RLS training process as that of the first layer (406). After obtaining the connection strength of the second layer, the entire distributed data set is input to the first hidden layer, and the values outputted to the second hidden layer are collected to obtain the connection strength of the third layer (407). The strength training course on the third floor is the same as the strength training course on the first floor. That is, the third layer connection strength is initialized to a small random value, the training parameters for the RLS training are initialized (408), and the third layer according to the RLS training method using the output of the entire distributed data set output from the second layer. The connection strength is obtained (409).

When the strength of the three layers is obtained, the training of the DKLT neural network is completed, and the entire neural network becomes a DKLT neural network with KLT transformation and block type linear discrimination.

5 is a flowchart illustrating a method of compressing and restoring image data according to the present invention.

When the image block data is input 501, the KLT coefficients and the DKLT coefficients are output (502) by performing KLT transformation on the input data in all DKLT neural networks and performing linear discriminant analysis necessary for block type determination. The output KLT coefficients are input to the vector quantizer paired with each DKLT neural network, and the DKLT coefficients are input to the output selector to determine the DKLT neural network that can best convert the input image block data and output the result (503). An output selection codebook is output (504). Only the output of the vector quantizer specified by the output selection codebook is output to the compressor output and transmitted with the output selection code to a remote location with a reconstructor via a communication channel (505). The decompressor first separates the output selection code from the compressed data to designate a vector table for inverse quantization and a corresponding inverse KLT neural network (506). The output of the vector quantizer, which is the remainder of the compressed data, is applied to the vector table to retrieve the inverse quantized KLT coefficients, and the searched KLT coefficients are input to the inverse KLT neural network (507) to inversely transform and finally output the recovered image block data. (508).

As described above, according to the present invention, by using the DKLT neural network for determining the type of the input image block and the optimal transform coding transformation, it is possible to perform high-speed parallel processing of the image data, and specialized vector quantization according to the type of each block It can improve the compression rate and the image quality of the recovery screen by adopting the device, as well as the system such as PACS which needs to compress / restore the medical image, as well as the compression / restore of general image which requires higher compression rate and better image quality than the existing lossy compression. There is also an excellent effect that can be utilized.

Claims (2)

The Karhnen-Lube transform coefficient and the classification Karhnen-Lube transform coefficient are performed by performing the Karchenen-Lube transform for compressing the input image block data by inputting the image block data vectorized in the dimension of the number of pixels in the image block. A number of classification Karchenen-Rube transformed neural networks that output Receive the Karhnen-Rube transform coefficients, which are present at the output terminals of the plurality of classification Karhnen-Rube transformed neural networks, respectively, and are compared with the coefficient vectors in the vector codebook. A plurality of vector quantizers for outputting a codebook address of a vector having a value most similar to the vector, Selects outputs by judging which classification Karhnen-Rube transformed neural network is the vector of the input image block by inputting classification Karhnen-Lube transform coefficients outputted from the respective classification Karhnen-Lube transform neural networks An output selector for generating control codes, The codebook contains the same codebook as the one used in the vector quantizer, and receives the codebook address, outputs the Karhnen-Lube transform coefficients stored in the address, and corresponds to the vector quantizer while in charge of inverse quantization. Vector table, The neural network, which is paired with the plurality of vector tables and includes an inverse Karchenen-Rube transform neural network that receives KLT coefficients output from the plurality of vector tables and inversely converts the outputted image block data. Structure of Compression / Restoration of Image Data Using Digital Camera. When the image block data is input to the classification Karhnen-Rube transform neural network, the Karhnen-Lube transform coefficient and the classification Karhnen-Lube transform coefficient are converted by performing the Karhnen-Lube transform on the input data and performing linear discriminant analysis for block type determination. Outputting, Classified Karchenn which inputs the output Karchenn-Lube transform coefficients to each vector quantizer and inputs the sorted Karchenn-Rube transform coefficients to an output selector to optimally convert the input image block data and output a result. Determining a rube transform neural network, Outputting a selection codebook address from the vector quantizer of the determined classification Karchenen-Rube transform neural network; remind. Transmitting the address of the vector quantizer specified by the selected selection codebook and the codebook address to a remote location having a reconstructor through a communication channel; Designating a vector table to be used for image block recovery using the vector quantizer address and corresponding inverse Karchenen-Rube transform neural network; Applying an output of a vector quantizer, the remainder of the compressed data, to the vector table to retrieve inverse quantized Karchenen-Lube transform coefficients; Compressing and restoring image data using a neural network, comprising inputting the retrieved Karhnen-Lube transform coefficients to an inverse Karhnen-Lube transform neural network, and inversely transforming the final block and outputting the restored image block data. .