KR100216600B1 - Method of transmitting multi-code for vector quantizer of image signal - Google Patents

Abstract

1. TECHNICAL FIELD OF THE INVENTION

Multiple Codeword Transmission Method for Image Signal Vector Quantizer.

2. The technical problem to be solved by the invention

We want to improve the performance of vector quantization by assigning one or more codewords to input images that are different from the training data and statistical characteristics, and interpolating them on the receiving side.

3. Summary of Solution to Invention

The present invention improves the performance of vector quantization by using a plurality of codewords to encode one input data pattern and reconstructing the original pattern by using the plurality of codewords received at the receiver.

4. Important uses of the invention

Applied to vector quantizer for compression and transmission of video signal.

Description

Multiple Codeword Transmission Method for Video Signal Vector Quantizer

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a multiple codeword transmission method as a method for improving performance in vector quantization, which is one of the important element techniques used to compress a video signal.

In general, the purpose of image data compression is to reduce the number of bits required to transmit or store an image without losing visual information. Coding techniques for image compression are based on information theory, and are based on first-generation image coding techniques for reducing image data by effectively using redundancy in time or space, and using human visual characteristics or pattern recognition techniques. It can be classified as a second generation video encoding technique.

Most of the classical data compression methods belonging to the first generation video encoding scheme are largely divided into two types of prediction methods and transform methods. Prediction method removes mutual redundancy between consecutive samples and quantizes only error information. Redundancy refers to a property in which a current sample value can be predicted by a previous sample value.

Therefore, instead of directly quantizing a pixel, the prediction method quantizes the difference between the predicted value obtained from each pixel and the previously encoded pixel. In the coding method, a given image is first concentrated in a small number of samples. Change the image as much as possible and quantize its transform coefficients. In general, various image data compression methods in the form of generalizing or combining these two methods have been proposed and studied. Second-generation image coding techniques use the characteristics of the human visual system (HVS) and include pyramid coding, directional decomposition-based, and segmentation-based coding. A common drawback of this coding method is that quantization is done in a scalar fashion on individual real values of waveforms or image pixels.

Vector quantization is a method of encoding an input block using a codebook composed of prototype patterns, which may be referred to as a pattern matching technique. According to Shannon's rate-distortion theory, vector quantization is always better than scalar quantization, and there is a gap between the conventional sub-optimal techniques and the theoretical limitations of source coding proposed by rate-distortion theory. To fill.

When the vector quantization is applied to video compression on a line having a low data rate, the codebook is small in size, and thus, it is difficult to prevent a significant deterioration in image quality of an input image having different statistical characteristics from training data.

In other words, in vector quantization, the input data pattern is represented by a codeword, which is a prototype pattern included in the codebook, where the reproduction quality is determined by the codebook and has input data having statistical characteristics similar to the training data used to learn the codebook. Only patterns can be properly represented. Therefore, in high compression coding where the size of the codebook should be small, information such as edges, which occupy a relatively small portion of the input image, cannot be properly represented.

An object of the present invention is to provide a multiple codeword transmission method for improving performance of vector quantization by allocating one or more codewords to an input image having different characteristics from training data and transmitting the interpolated codewords.

1 is a vector quantization system block diagram,

2 is a conceptual diagram of image vector quantization according to the present invention;

3 is a conceptual diagram of a multiple codeword transmission method according to the present invention;

4 is a processing flowchart of a multiple codeword transmission method according to the present invention.

In order to achieve the above object, the present invention provides a first step of selecting a codeword having a minimum distortion by searching a codebook when an input vector is input, and having a next smaller distortion when the distortion is larger than a predetermined threshold. A second step of selecting two codewords and interpolating them by an average and calculating the distortion between the interpolated result vector and the input vector and three codes if the value of the distortion error is smaller than the distortion error of the first selected codeword. A third step of transmitting all indexes of the word, otherwise only the index of the first selected codeword, and a fourth step of restoring information transmitted from the receiving side.

Hereinafter, with reference to the accompanying drawings will be described an embodiment of the present invention;

1 and 2 show a block diagram and a conceptual diagram of a vector quantization system.

In a vector quantization system, encoders and decoders must store the same codebook in advance.

Vector quantization encoding first calculates the distortion between the N codewords stored in the codebook and the K-dimensional input vector, and selects a codeword for which the distortion is minimized. This process is called exhaustive codebook search. In the global codebook search, find the best matched codeword and send the index. The decoder of vector quantization also stores a codebook and only requires a table look-up process. At this time, the output index of the received encoding is used as an address for extracting the corresponding codevector from the same codebook. This simple decoding process is useful for applications such as broadcasting, where the cost of the receiver must be lower than the transmitter.

In vector quantization, the input data pattern is represented by a codeword, which is a prototype pattern included in the codebook. In this case, only the input data pattern having a statistical characteristic similar to the training data used for learning the codebook may be properly represented. Therefore, in high compression coding where the size of the codebook should be small, information such as edges, which occupy a relatively small portion of the input image, cannot be properly represented.

The multiple codeword approach overcomes this problem by using multiple codewords to encode one input data pattern. The receiving side restores the original pattern using a plurality of codewords.

3 is a block diagram of an encoding and decoding process of an input vector by a multiple codeword method.

In the figure, P _i is a given input data vector and W _i = [W _i1 , W _i2 ... W _iM ] is an encoded codeword and R _i is a reproduction pattern at the receiving side. On the receiving side, R _i is calculated as in Equation 1.

Where a _ij is the weight vector for interpolation.

The reproduced pattern R _i thus reproduced is a new prototype pattern not included in the codebook. However, this method has to transmit the weight vector (aij) for interpolation together with the selected codeword, resulting in a decrease in the compression rate by increasing the amount of transmitted information. Therefore, there is a need to find a suitable compromise between playback quality and compression. The simplest method is to select three codewords (M = 3) and apply the method of interpolation by the average of the codewords at the receiving side instead of using the weight vector (aij). In this case, the weight vectors a _ij are all 1 / M.

When the neural network learning algorithm is used for the training of the codebook, it is very advantageous for the adaptive vector quantization which learns and updates the codebook whenever both sides of the transmission and reception are needed. The receiver recovers a single codeword or multiple codewords according to the indes information sent from the transmitter.

An internal processing procedure, that is, the encoding and decoding processes of the multiple codeword method is shown in FIG. 4.

When an input vector comes in (1), the codebook is first searched (2) to select the codeword with the least distortion. If the distortion at this time is larger than a predetermined threshold value, two codewords having the next smaller distortion are selected, interpolated by the average, and the distortion with the resulting vector is calculated (4 to 7). If the value of the distortion error is smaller than the distortion error with the first selected codeword, all the indexes of the three codewords are transmitted (8, 9). Otherwise, only the index of the first selected codeword is transmitted (8, 10). At this time, information indicating how to be encoded needs to be added to the transmitted index.

The receiver simply reads this information and restores the receiving index as an address if it is encoded by a single codeword, obtains three transmit vectors using the receiving index as an address if it is encoded by multiple codewords, and then interpolates them. Restore (11-14).

The present invention described above can be variously substituted, modified and changed within the scope without departing from the technical spirit of the present invention for those skilled in the art to which the present invention belongs, the above-described embodiment and drawings It is not limited to.

Accordingly, the present invention can improve the performance of vector quantization by using a plurality of codewords to encode one input data pattern and reconstructing the original pattern by using the plurality of codewords received at the receiver. In addition, this method has the advantage that it can be operated in a variable bit rate method in the situation that the characteristics of the line changes with time.

Claims (3)

The first step is to select a codeword with the least distortion by searching the codebook when the input vector comes in; and, if the distortion is greater than a predetermined threshold, select the next two codewords with the smallest distortion and add them to the average. The second step of calculating the distortion between the interpolated and interpolated result vector and the input vector, and if the distortion error value is smaller than the distortion error between the first selected codeword, all three codeword indices are transmitted. And a fourth step of transmitting only the index of the first selected codeword, and a fourth step of reconstructing information transmitted from the receiving side. The method of claim 1, wherein the fourth step is to simply read the encoding scheme indication information added to the transmitted index to restore the information. 2. The method of claim 1, wherein the fourth step restores the reception index to the address when the information from the transmitter is encoded by a single codeword, and sets the reception index as an address when encoded by multiple codewords. In the case of reconstruction and encoding by multiple codewords, three transmission vectors are obtained by using a reception index as an address, and then interpolated and reconstructed.