RU2316908C2 - Method for combined usage of vector quantization and fractal encoding of images - Google Patents

Abstract

FIELD: digital processing of images, possible use during transmission of images through low speed communication channels.

SUBSTANCE: in accordance to the invention, technical effect is achieved by combined usage of vector quantization and fractal encoding of data during image processing. During that, identical code books are formed in advance in encoder and decoder same as for vector quantization. Then fractal compression method is applied to original image. As domain blocks, either fragments of image being encoded, or code book vector, are used. If rank block is better indicated by domain block, the index of corresponding domain is transmitted to decoder, in opposite case the index of vector, on basis of which the required fragment, used to fill the restored image after fractal decoding, is selected from code book during receipt, is transmitted to decoder.

EFFECT: ensured high quality of transmitted information at high degree of its compression.

3 dwg

Description

The invention relates to the field of digital image processing and can be used when transmitting images on low-speed communication channels. The technical result of the invention is the development of a method for sharing vector quantization and fractal coding of images, ensuring high quality of transmitted information with a high degree of compression.

In [1], the classical algorithm for fractal coding of images is presented, which consists in the following. For each ranking block of the source image, find the domain and the corresponding transformation that best covers the ranking block. This is usually an affine transformation. The affine transformation consists of three stages. First, one of eight basic rotations / reflections (four rotations of 90 degrees and mirror reflection in each orientation) is applied to the selected domain. Secondly, the rotated domain region is compressed to fit the size of the ranking region. And finally, the least squares method calculates the parameters of contrast α and brightness β, corresponding to the optimal values at which expression (1) is minimized

where E (R, D) is the metric between the rank block R and the domain block D reduced to it in size;

n is the number of rows in the processed rank block;

m is the number of columns in the processed rank block.

If the exact match did not work out, then the rank blocks are divided into smaller ones. Continue this process until they reach an acceptable match, or the size of the rank blocks reaches a certain predetermined limit.

In [2], a vector quantization method is presented, which is one of the varieties of efficient block coding. As applied to images, vector quantization is as follows. The original image is divided into blocks of elements. Blocks are identified with vectors whose dimension is determined by the number of elements included in the block. Next, these source vectors are compared with the code vectors making up the code book, and one of the set of code vectors is selected as the quantized vector, at which the adopted measure of distortions characterizing the distance between the input and code vectors is minimized. When the codebook is generated, the code vector index uniquely determines the quantized vector, since there is the same code book on the receiving side (it is transmitted to the receiver in advance), and when the index number is received on the receiving side, the corresponding code vector is selected from the code book and the image location is filled with it that matches the location of the original vector.

Thus, the vector quantization method is based on the preliminary construction of code books based on the learning process of specific image blocks. In this case, the used codebooks do not preserve the topology of the encoded data, which leads to significant coding errors.

The closest in technical essence to the claimed method of sharing vector quantization and fractal coding of images is a coding and decoding device using vector quantization according to US patent No. 5010401, H 04 N 7/13, 1991 [3].

The image conversion in the specified prototype is as follows. The image is divided into blocks of κ samples in each (κ is a positive integer), inter-frame difference blocks are formed and stored, a standard library (code book) is formed for specific image blocks, the image block code determines the address of the block in the code book, and this block is read during restoration in the buffer memory with the subsequent formation of the restored image,

The disadvantage of the prototype is that the codebook constructed in this way is adapted only to a certain class of images and cannot be used to encode other types of images.

The aim of the present invention is to develop a method of sharing vector quantization and fractal coding of images, providing improved image recovery quality and reduced decoding time compared to the fractal compression method for any type of image while maintaining the transmission speed and frame format length.

The essence of the proposed method is to combine the fractal compression method with vector quantization. First, the fractal compression algorithm described in [1] is applied to the original image. After that, the search for the most suitable domain is carried out from the codebook. This is due to the fact that for ranking areas with a large dynamic range it can be difficult to select a domain with an even larger dynamic range, therefore, a codebook is used to encode such areas. The codebook is formed by an iterative cluster algorithm, also known as the k-means algorithm described in [2]. In accordance with this algorithm, a certain sequence of images belonging to different classes is selected. Next, this sequence is converted into a sequence of input vectors or a training sequence by which a code book is formed. In this case, the input vectors should have a large dynamic range, since the codebook should contain blocks that display sharp edges between fragments of the image, since it is in such cases that problems may arise with finding a suitable domain. Such a situation usually arises due to restrictions on the dynamic range of the desired domain, which, as a rule, should be greater than the dynamic range of the rank domain. In areas with sharp boundaries, the dynamic range is very large and it often turns out that it reaches its maximum value. In this case, there is no domain with a strictly large dynamic range and the rank area is replaced by the most similar block from the codebook. Blocks are chosen so that their structure does not depend on resolution, and this gives it an additional advantage in combination with the fractal method.

In addition to improving the quality of image recovery, the decoding time is reduced. If the codebook is composed of vectors having a large dynamic range that are contained in the encoded image, then from the first iteration, the blocks that were compressed using the codebook receive the final correct values. Accordingly, this affects the rate of convergence of the entire set of image points due to the fact that these areas, as a rule, serve as rank areas for other domain areas of the image, which already by the second iteration receive values quite close to the limit.

The claimed method is illustrated by drawings:

- Figure 1 image coding algorithm when sharing vector quantization and fractal coding of images;

- Figure 2 diagram of the sharing of vector quantization and fractal coding of images;

- Fig.3 dependence of the peak signal to noise ratio (PSNR) on the number of iterations when restoring the original image using the basic fractal compression algorithm and the algorithm for sharing vector quantization and fractal image encoding.

An image encoding algorithm for sharing vector quantization and fractal image encoding is proposed in FIG. The original image is divided into non-overlapping rank and domain blocks. Domains can be of different sizes, and usually their number is in the hundreds and thousands. After that, for each rank block, the domain and the corresponding transformation that best covers the rank block are found. This is usually an affine transformation. The domains may be domain regions of the source image or codebook blocks. Encoding is completed when each rank block is covered by a domain region with a given error.

Figure 2 shows a scheme for sharing vector quantization and fractal coding of images. Previously, in the encoder and decoder using vector quantization form identical code books. Next, a fractal compression method is applied to each ranking block of the source image. As a result of which, the algorithm compares the options for converting all domains to the selected rank block. Either fragments of the encoded image or codebook vectors are used as domain domains. If the rank block is better displayed by the domain block (E ₁ (R, D) _min <E ₂ (R, D) _min ), then the index of the corresponding domain is transmitted to the decoder, otherwise (E ₁ (R, D) _min > E ₂ ( R, D) _min ) the vector index is transmitted to the decoder, according to which the necessary fragment is selected from the codebook at the reception, with which, after fractal decoding, the restored image is filled.

Figure 3 shows the dependence of the peak signal to noise ratio (PSNR) on the number of iterations when restoring the original image using the basic fractal compressed algorithm and the algorithm for sharing vector quantization and fractal image encoding. As can be seen in figure 3, when using vector quantization, the maximum image quality is reached by the fourth iteration, while in the basic fractal algorithm, this image is stabilized only at the seventh iteration.

The proposed method for the joint use of vector quantization and fractal coding increases the computational complexity of the image coding process, which is insignificant given the reduction in decoding time and the current level of development of digital signal processing processors. The specified method can find its application in the transmission of images on low-speed communication channels

Literature

1. S. Welstead Fractals and wavelets for image compression in action. Tutorial. - M .: Triumph Publishing House, 2003.

2. Satellite television. New transmission methods. / N.G. Kharatishvili, E.I. Kumysh, V.Yu. Yepanechnikov, O.G. Zumburipze; Ed. N.G. Kharagishvili. - M: Radio and communications, 1993.

3. US Patent No. 5010401, IPC H 04 N 7/13, 1991

Claims (1)

A method of sharing vector quantization and fractal coding of images, which consists in the fact that in the process of vector quantization the original image is divided into blocks of k samples in each (k is a positive integer), identical code books are formed in the encoder and decoder, for each rank block of the source The fractal coding method is applied to the image; the variants of the fractal transformation of all domains are compared with the selected rank block, using either fragments of diruemogo image or vector codebook, wherein if the ranking unit more precisely displayed domain block E ₁ (R, D) _min <E ₂ (R, D) _min, the decoder transmitting index corresponding domain, otherwise for a large block dynamic range at E ₁ (R, D) _min > E ₂ (R, D) _{min the} decoder receives the index of the vector by which the corresponding index index is selected from the codebook at the reception, which fills the reconstructed image after fractal decoding, Where

- the metric between the rank block R and the domain block D reduced to it in size, n is the number of rows in the processed rank block, m is the number of columns in the processed rank block, E ₁ (R, D) - metric, where the domains are blocks of the original image, E ₂ (P, B) is a metric where the domains are blocks from the codebook.

Images