KR20030021009A - Image compression method using block-based zerotree and quadtree - Google Patents

Abstract

PURPOSE: A method for compressing video using block-based zerotree and photograph tree structure is provided to efficiently encode insignificant coefficients covering a wide range and obtain an excellent property in a pyramid of a low level. CONSTITUTION: All blocks on a wavelet top level are repeatedly divided to find out significant blocks following a spacial orientation tree structure. A photograph tree structure division is executed while the significance of the found blocks related to a threshold value is evaluated, so that significant coefficients existing in a first block are found and coded until the block becomes small by the pixel. To the blocks which are judged as insignificant blocks, the photograph tree structure division is executed related to a current threshold value, in order of small size. Bits on a current bit plane of the significant coefficients are output, so that first found significant coefficients are restored to more accurate original values. The previous steps are repeated until a wanted bit rate or image quality is obtained, lowering the current threshold.

Description

Image Compression Method Using Block-Based Zero-Tree and Photo-Tree Structure {IMAGE COMPRESSION METHOD USING BLOCK-BASED ZEROTREE AND QUADTREE}

The present invention relates to an image compression method, and more particularly, to an image compression method that provides excellent compression performance by efficiently encoding a bit plane using a block-based zero tree and a photographic tree structure.

Recently, wavelet theory has been introduced, and many studies on image coding based on wavelet transform have been conducted. This is due to the wavelet transform providing several useful properties, such as energy compression and space-frequency localization.

The EZW (Embedded Zerotree Wavelet) method proposed by Shapiro uses two characteristics in terms of efficiency and complexity by using the characteristics of the image after wavelet transform, such as spatial-frequency localization, similarity between frequency bands, and size reduction of wavelet coefficients according to frequency bands. It is a video compression method to improve the side at the same time.

The Set Partitioning In Hierarchical Trees (SPIHT) method proposed by Said and Pearlman uses a new set partitioning method to encode the importance map, which outperforms EZW without using an arithmetic coder. .

Both the EZW and SPIHT methods use the correlation of non-significant wavelet coefficients between frequency bands using a zerotree, a structured set of coefficients, to encode the importance map.

On the other hand, the morphological representation of wavelet data (MRWD) method proposed by Servetto et al. And the signal-linked connected component analysis (SLCCA) method proposed by Chai et al. Are characterized by clustering of wavelet coefficients that are important in the frequency subband. By efficiently encoding the importance map using the characteristics of the important coefficients between the subbands, the performance is comparable to the zero tree-based methods.

The above two approaches have the common goal of efficiently encoding importance maps using the properties of wavelet transformed images, but they have developed algorithms from different aspects. In other words, the coding methods mentioned above focus on compressing and describing whether or not wavelet coefficients are important in a given bit plane. Zero-tree-based methods such as EZW and SPIHT have zero-zero coefficients in the form of zero trees. After encoding the non-critical coefficients effectively, the non-zero coefficients are located by excluding the regions where the coefficients are zero, and the MRWD and SLCCA methods encode only coefficients that surround the clustered non-zero coefficients. Identify coefficients as important, and identify other areas as noncritical.

The present invention uses block-based zerotree to efficiently encode non-essential coefficients that exist while showing correlations across frequency bands, and uses an image tree structure to find and encode important coefficients in a block, thereby compressing an image having excellent compression performance. The purpose is to provide a method.

An image compression method using a block-based zero tree and a photographic tree structure according to the present invention for realizing the above object is a first step of finding an important block by repeatedly dividing a spatial direction tree structure derived from a block on a wavelet top level. And a second step of finding and encoding the significant coefficients existing in the first block by repeatedly performing the photo-tree structure division on the block while evaluating the importance on the threshold. A third step of performing the second step with respect to the current threshold in order of decreasing magnitude, and the bits on the current bit plane of the significant coefficients found in the second step at the previous threshold; Outputs a more accurate circle of the important coefficients found first. And to the fourth step of restoring to a value, when achieving the desired bit rate and image quality, while reducing the current threshold value of a row and a fifth step of repeating the fourth step of the first stage to.

1 is a conceptual diagram showing a parent-child relationship in a spatial tree structure;

2 is a basic block structure diagram of an image compression method according to the present invention;

There may be a plurality of embodiments of the present invention. Hereinafter, preferred embodiments will be described in detail with reference to the accompanying drawings. This embodiment allows for a better understanding of the objects, features and advantages of the present invention.

The present invention efficiently encodes the importance map using a block-based zero tree and a photo-tree structure. In the block-based zero tree used in the present invention, it is not a wavelet coefficient that constitutes each node of the hierarchical tree structure. A block containing coefficients. The compression method of the present invention encodes a wide range of zero coefficients by block-based zero tree and finds a block containing important coefficients. An important block found in the zero-tree coding process is to divide the block stepwise by the photo-tree structure method to find important coefficients located in the block. In other words, the compression method of the present invention utilizes a feature that effectively encodes coefficients that are not important for the zero tree method, divides an important block into a photo-tree structure, and effectively finds and encodes important coefficients generally present in a clustered form. Is to encode the map.

In order to help understand the compression method according to the present invention as described above, the basic method and various terms of zero-tree coding will be described first.

Zero-tree bit plane coding is one of the methods for efficiently encoding wavelet transformed coefficients. This method has advantages such as low complexity, embedded bit stream, multiple resolution, and accurate bit rate adjustment, in addition to excellent compression performance. Because of this property, zero-tree bit plane coding has been studied in recent years. The feature of this coding method is the use of interband similarity in the wavelet transformed image to predict the positions of important coefficients between frequency bands and the successive approximation of wavelet coefficients.

The terminology and encoding process used in the algorithm of zerotree bitplane coding are described in more detail.

Zero tree bit plane coding is a method applied to discrete signal wavelet transformed coefficients. In general, the discrete signal wavelet transform divides an input signal into hierarchical frequency subbands representing various resolutions. The sparse subbands (the top-level subbands) are approximations of the low frequency band of the original image, while the other subbands are the signals that make the approximate signal finer. Hierarchical band division systems, such as wavelet transforms, can associate all coefficients on one level with a set of coefficients in the same direction on a lower level. In this relationship, the coefficients on the sparse level are called parents, and the coefficients at the same location in the same direction on the lower level are called children. For example, FIG. 1 shows a wavelet tree structure derived from the coefficients on subband LL2. One parent will have four children at the next level, and each child will also have four children at the next level. Thus, for a given parent coefficient, four coefficients at one level lower, 16 coefficients at the next level, and so on can be established in this way, the related coefficients on these lower levels are called descendants. In this wavelet transformed image, a relationship is established between parent, child, and offspring, thus forming a tree structure. Such a structure is called a spatial orientation tree.

The zero-tree method introduced by EZW is based on the spatial direction tree structure. The zero-tree structure is important if a wavelet coefficient in the sparse band is not important for a given threshold T (i.e. expressed as 0), and the coefficients in the same position in the finer band below it are important for the same threshold T. Use points that are very unlikely. The EZW encodes all coefficients on the tree structure into one symbol using a zerotree symbol if all coefficients on the tree structure including the given coefficients are not important for the given threshold T. Unlike EZW, SPIHT determines the importance of each given coefficient and the tree structure derived from the coefficient. Since the threshold T is usually set to 2 ^ n, the successive threshold T becomes a bit plane according to n, and coding is performed by determining whether coefficients are important on this bit plane. In this case, representing the importance information of the coefficient on one bit plane is called a importance map. For coefficients determined to be important, the coefficients are coded so that the coefficients are restored to more accurate values for each bit plane which is continued by the continuous approximation quantization method.

Hereinafter, an image compression method using a block-based zero tree and a photo tree structure according to the present invention will be described with reference to FIG. 2.

As shown in the basic block structure diagram of FIG. 2, the image compression method according to the present invention comprises a zero tree step, a photo tree structure step, and a refinement step.

In the zero-tree step, the blocks are ordered by the importance test of the blocks along the spatial direction tree structure, and finally, the significant blocks are found.

When k is a nonnegative integer, the block Contains wavelet coefficients of magnitude. If a block B satisfies Equation 1 for n, the block is defined as important, and if the given block does not satisfy Equation 1, the block is not important.

In Equation 1 above Is a wavelet coefficient, and (i, j) represents the coordinates of that coefficient.

The spatial tree structure is introduced to evaluate the importance of a block group by using the similarity between the bands of the wavelet transformed image and localization of coefficients, and has a structure shown in FIG. 1. In a spatial tree structure, each node consists of 2 × 2 adjacent blocks, and each block of nodes has four child blocks. At the top level of the wavelet pyramid, however, only one block in the node (the block marked with * in Figure 1) has no child blocks.

In practical implementation of the algorithm, the present invention maintains four lists, each of which has a list of unimportant blocks (LIB), a list of unimportant block sets (LIBS), a list of important coefficients (LSP), and a photograph tree. Structure Applied Block List (LQB).

In the initialization phase, the LIB is placed at the top level of the wavelet pyramid. With blocks of size, LIBS is initialized with all blocks at the top level except blocks that have no descendants, and LSP and LQB are set to empty lists.

The basic action of the zerotree phase is to iteratively partition the block sets on the LIBS, find each non-critical block and a set of non-critical blocks of smaller size than before, and list each location in LIB and LIBS as appropriate. In the end, it's about finding important blocks. Once an important block is found, the block tree segmentation process is immediately applied to that block.

The photo-tree block partitioning process performs the photo-tree structure division on the block while evaluating the importance of the block against the threshold n. If a given block is important, the block is divided into four smaller blocks with one quarter the size of the block. These four blocks are each set to a given block again and repeatedly processed until they are reduced pixel by pixel. This iterative process finds and encodes the significant coefficients that exist in the first block, and the significant coefficients are found on the LSP. Blocks that are determined to be insignificant during this process will be listed on the LQB and reevaluated for the next lower threshold at the photographic tree structure stage.

In the photo-tree structure step, the photo-tree structure block division is performed on the non-critical blocks listed on the LQB with respect to the current threshold. There may be blocks of size (where k = 0, 1, ..., i-1). These blocks are processed in order of decreasing size. In other words, the 1 × 1 block is processed first, followed by 2 × 2 blocks. The reason for processing the blocks in the order of the smallest size is that all the blocks listed on the LQB are generated when there is a significant coefficient in adjacent blocks of the same size as the block. Therefore, the smaller the block size, the higher the probability that the block contains important coefficients at successive thresholds.

The refinement step after the photo-tree structure step outputs the bits on the current bit plane of the coefficients listed on the LSP at the previous threshold to restore the first found significant coefficients to more accurate original values. These three steps are repeated until the desired bit rate or image quality is achieved while continuously lowering the current threshold twice. At this time, the bits output in each step may be a bit stream without passing through the entropy encoder, and the bit rate may be further lowered through the entropy encoder.

The image compression method of the present invention as described above is summarized as follows.

Initialization Steps:

First course- Set n to be greater than the absolute value of all these wavelet coefficients.

Step 2-LSP and LQB are set to empty lists, LIB and LIBS contain all blocks on the top level of the wavelet. LIBS excludes blocks without descendants.

Picture of wood structure steps:

Step 3-The blocks listed on the LQB are sent to the photo-tree block dividing step in the order of the smallest block size.

Zero Tree Steps:

Step 4-Make an importance determination on all the blocks on the LIB. If a block is important, delete the block from the LIB and send it to the split picture tree block.

Step 5-For every block on LIBS, the importance of the descendants of that block is determined. If any of the descendants are important, the children of that block are evaluated for importance. Important blocks of the children are sent to the photo-tree block division stage, and non-critical blocks of the children are listed at the end of the LIB.

Step 6-Delete the block from LIBS, and if there are children of the children, add the children to LIBS.

Picture of wood block dividing step:

Step 7-Divide a given block into four smaller blocks of the same shape and make an importance determination on each small block. If the significant block is a coefficient, it is sent to the eighth step, otherwise it is sent to the seventh step. Insignificant blocks are listed on the LQB.

Step 8-Output the sign of the coefficients and list them in the LSP.

Refinement Steps:

Step 9-Output the bits on the n-th bit plane of the coefficient size except the coefficients listed in the last zero tree step among the coefficients listed on the LSP.

Quantization Size Update Steps:

Step 10-Reduce n by 1 and go to the photo-tree structure step.

On the other hand, due to the structure of the algorithm, it is natural that the photo-tree structure step comes after the zero-tree step, because the photo-tree structure step processes small blocks that are generated from the important blocks found in the zero-tree step. However, in the actual structure as shown in Figure 2, the photographic tree structure step is placed before the zero tree step. Since there are no blocks on the LQB in the first bit plane, the phototree structure step outputs no bits, while the subsequent bit planes process the blocks that occur in the zero tree phase in the previous bit plane. This order is arranged because the bit that occurs in the photo-tree structure stage has a higher performance improvement than the bit in the other stages in terms of bit rate-distortion, so that the overall bit rate-distortion curve can be improved by positioning it first. .

In the importance determination performed several times in the present invention, if there is a coefficient larger than the threshold value among the given blocks or descendants, '1' is outputted, otherwise '0' is outputted. When outputting the sign of a coefficient, only '0' and '1' are output, so these output bits can be transmitted and stored as a bit stream. However, the entropy coder can be used to further lower the bit rate. In the present invention, the adaptive arithmetic coder is used as the entropy coder to further reduce the bit rate using the correlation used during the encoding process. First, when determining the importance of each block and its descendants in the zero-tree step, arithmetic coding is performed by conditionally setting whether neighboring blocks and descendants of the blocks are determined to be important in the previous bit plane first, and also photographs. Even when performing tree block partitioning, arithmetic coding is performed by varying the conditional context depending on whether there are significant coefficients found first in adjacent blocks of the same size when determining the importance of the divided blocks.

The compression efficiency of the image compression method of the present invention as described above can be easily seen from the following experimental results.

For the experimental images, Barbara portrait and Lena portrait images, which are black and white images of 512 × 512 size, with 8 bits per pixel precision, were used as wavelets using 9-7 Daubechies filter. The conversion was done. The bit rate was calculated from the actual size of the actual compressed file, and the image quality was measured by the peak signal to noise ratio (PSNR) calculated from the actual reconstructed image.

The gist of the present invention is not only to take advantage of the zero tree that has insignificant coefficients, but also to take advantage of the property that important coefficients exist in clusters. In order to prove that the present invention is effective, we investigated how quickly the coefficients in the critical block turn out to be important coefficients in the course of successively decreasing the threshold.

Table 1 below shows the specific gravity of important coefficients in a block in two bit planes.

1st bit plane 2nd bit plane 3rd bit plane 4th bit plane 5th bit plane Zero Tree Step on First Bit Plane 70.0% 74.5% 49.6% 48.9% 47.5% One step lower bit plane picture of wood structure step 68.4% 62.9% 45.9% 46.0% 42.7% Total ratio of significant coefficients in two bit planes 90.5% 90.5% 72.7% 72.4% 69.9%

Table 1 shows the rate at which the coefficients in an important block are found to be important during the zero-tree phase on the threshold where the block is found, and the rate at the next lower threshold for the Barbara character image that is important during the photographic tree structure phase. About 45-70% of the coefficients in the critical block turn out to be important coefficients at the first threshold and about 40-70% of the remainder are important at the next lower threshold. Eventually, the coefficients of 70-90% in the critical blocks turn out to be important at the first or next lower threshold. Therefore, since the coefficients in the critical block once found are found as most important coefficients in the first two bit planes, the bit generation rate due to the importance determination at successive thresholds is not large. Therefore, we can find that the proposed algorithm effectively finds the important coefficients by using the block-based zero tree and the photo-tree structure.

Table 2 below is a table of image quality at each bit rate by conversion to block size.

Bit rate block size 0.125 bpp 0.25 bpp 0.5 bpp 1.0 bpp 2 x 2 24.39 dB 27.45 dB 31.56 dB 36.95 dB 4 × 4 24.57 dB 27.68 dB 31.70 dB 37.02 dB 8 × 8 24.62 dB 27.69 dB 31.71 dB 37.03 dB 16 × 16 24.72 dB 27.71 dB 31.72 dB 37.04 dB

Table 2 shows the performance of the encoder by changing the block size to various sizes after wavelet transforming the Barbara portrait image to three levels. As shown in Table 2, the larger the block size, the better the image quality of the image reconstructed at the same bit rate. It can also be seen that the performance difference at lower bit rates is greater than at higher bit rates. However, since the performance of the proposed encoder is almost saturated since the block size is 4 × 4, the performance improvement is not so great when the block size becomes larger. The reason for the large performance difference according to the block size at the low bit rate is as follows. At low bit rates, only a few of the most significant bit planes are encoded. In these bit planes, there are more noncritical coefficients than important coefficients, in which case the encoder consumes a large amount of bits to encode noncritical coefficients. In this case, the larger the block size is, the less significant coefficients can be encoded by using a smaller number of zero trees, and the bit amount can be reduced in encoding such coefficients. As it decreases, more bits can be used to encode important coefficients, which in turn improves the quality of the reconstructed image. However, as the bit rate increases, the ratio of important coefficients on the bit plane increases, which not only reduces the advantage of the larger block, but also requires more bits to encode the block by dividing the picture tree structure. Performance difference according to size is not big.

Table 3 below is a performance comparison table of the SPIHT and the present invention according to the wavelet pyramid level.

Bit rate pyramid level 0.125 bpp 0.25 bpp 0.5 bpp 1.0 bpp 3 level pyramid SPIHT 23.55 dB 26.46 dB 30.74 dB 36.32 dB The present invention 24.62 dB 27.69 dB 31.71 dB 37.03 dB 5 level pyramid SPIHT 24.95 dB 27.72 dB 31.65 dB 36.93 dB The present invention 25.46 dB 28.22 dB 32.03 dB 37.22 dB

Table 3 compares the performance of the proposed algorithm with 8 × 8 blocks and the performance of SPIHT after varying the levels of the divided pyramids when wavelet transforming Barbara portraits. At both levels, the proposed algorithm outperforms SPIHT. For the pyramid level 3, the proposed algorithm outperforms SPIHT. This is because SPIHT is an algorithm using only a zero tree using a parent-child relationship on a spatial tree structure, whereas the present invention uses not only a zero tree but also a photo tree structure method. If the pyramid level is small, the performance of SPHIT using only this relationship is reduced because the gain obtained by the zero tree using the spatial tree structure is small, but the present invention using the photo tree structure as well as the zero tree has a low pyramid level. Even if it shows excellent performance. In the case of a small image, the burden of increasing the pyramid level is not great when wavelet transforming. However, as the image is larger, the calculation amount of pyramid splitting at a higher level cannot be ignored.

Table 4 below is a performance evaluation table of the present invention and SPIHT with or without arithmetic encoder binding.

Bit rate experimental image and arithmetic encoder 0.125 bpp 0.25 bpp 0.5 bpp 1.0 bpp Barbara Water Video Arithmetic code SPIHT 23.55 dB 26.46 dB 30.74 dB 36.32 dB The present invention 24.62 dB 27.69 dB 31.71 dB 37.03 dB Arithmetic code base oil SPIHT 24.52 dB 27.59 dB 31.75 dB 37.23 dB The present invention 24.90 dB 28.09 dB 32.19 dB 37.56 dB Lena portrait video Arithmetic code SPIHT 27.53 dB 32.01 dB 36.12 dB 39.66 dB The present invention 29.72 dB 33.21 dB 36.62 dB 39.90 dB Arithmetic code base oil SPIHT 30.13 dB 33.54 dB 36.93 dB 40.28 dB The present invention 30.13 dB 33.59 dB 36.97 dB 40.27 dB

Table 4 shows the evaluation of the performance of the present invention, which has a block size of 8 × 8 and SPIHT, when combined with an arithmetic encoder and when the Barbara portrait image and the Lena portrait image are divided. The performance without combining the arithmetic code shows that the present invention is superior to the performance of the SPIHT due to the above-mentioned characteristics, and the performance difference is greater at a particular low bit rate. However, when the arithmetic encoders are combined, the performance difference of the two encoders is reduced, and in the case of Lena portrait images, the performances of the two encoders are almost the same. In the arithmetic code, SPIHT is arithmetic coded by referring to peripheral coefficients in pixel units so that the correlation with peripheral pixels can be effectively used, whereas the present invention uses the correlation with neighboring blocks in block units. It is not larger than that of the pixel unit, and the performance of the arithmetic encoder is not large. In another aspect, the present invention may be regarded as using the correlation between pixels by first using a block tree and a tree structure before passing through an arithmetic encoder. In addition, since the combination of the arithmetic encoder increases the complexity of the overall system, the present invention showing excellent performance in the state in which the arithmetic encoder is not coupled may have a useful aspect.

As described above, the present invention uses the block-based zero tree method to efficiently encode non-critical coefficients over a wide range by taking a zero-tree structure in which the elements of each node are blocks rather than coefficients. Because of its use, it shows excellent performance even at low level pyramids, especially at low bit rates, because it can allocate more bits to encode important coefficients.

Claims (2)

In the method of encoding the importance map of the wavelet transformed image, Firstly dividing all blocks on the wavelet top level to find important blocks along a spatial direction tree structure; The found significant block performs picture tree structure segmentation on the block while evaluating the importance of the threshold value, and finds and encodes the significant coefficients existing in the first block, until the given tree structure segment becomes smaller in pixels. Repeating a second step; A third step of performing the second step with respect to the current threshold in the order of the smallest size of the blocks determined to be insignificant in the second step; A fourth step of outputting a bit on the current bit plane of the significant coefficients found in the second step at a previous threshold to restore the first found significant coefficients to a more accurate original value; And And a fifth step of repeating the first to fourth steps until a desired bit rate or image quality is achieved while continuously lowering a current threshold. The method of claim 1, And a sixth step of lowering the bit rate by using an entropy encoder to generate a bit stream using the entropy encoder.