KR20010028051A - An index assignment method for coding based on Tree-Structured Vector Quantization - Google Patents

Abstract

PURPOSE: The index allocating method for coding is provided to equally maintain a quality of an image comparing with a conventional method and improve efficiency of the coding. CONSTITUTION: The tree structure vector quantizer quantizes source vectors successively inputting into corresponding code vectors based on a tree structure vector quantizer code book. An index of the code vector quantized among the code vectors is allocated by using special correlate among the code vectors successively quantized in the tree structure vector quantizer. The allocated index is decided by a shortest path between successive two code vectors by computing in shortest path generator(S). A state machine(SM) is used for sequence of the shortest path. Thereby, performance of communication system such as transmission time and necessary memory capacity can greatly improved.

Description

An index assignment method for coding based on Tree-Structured Vector Quantization}

The present invention relates to an index allocation method for encoding based on tree structure vector quantization (TSVQ), and in particular, by adaptively assigning an index according to a correlation between two source vectors that are successively vector quantized using a TSVQ codebook. In addition, the present invention relates to an index allocation method that greatly improves the compression encoding efficiency while maintaining the same quality of an image.

In general, vector quantization (hereinafter, referred to as 'VQ') technique in contrast to scalar quantization uses source code or K-dimensional region of the source parameters to correspond to a code vector ('representative') using a VQ codebook. Vector '). The vector quantization technique further compresses the data amount of the transmitted source information by finding an index identifying the code vector from the VQ codebook and transmitting it to the channel instead of the quantized code vector. Therefore, the vector quantization technique is very effective for source coding at low bit rates because the data compression rate of source coding is very large.

The Tree-Structured Vector Quantization (TSVQ) technique uses a Tree-Structured codebook among the above-described techniques of vector quantization. FIG. 1 shows an implementation of a codebook used in an encoding system based on Variable Length-TSVQ (VL-TSVQ), which further improves coding efficiency among various TSVQ techniques. The VL-TSVQ technique improves coding efficiency by allocating an index having a different amount of data according to a probability of generating an image having a predetermined characteristic. Specifically, an index having a smaller amount of data is obtained as the occurrence probability of an image having a predetermined characteristic is larger. The smaller the probability of occurrence, the larger the index of data.

Referring to Fig. 1, a codebook of VL-TSVQ modified from the TSVQ technique, and a general signal coding method based on the same, will be described. The VL-TSVQ codebook shown in FIG. 1 consists of two kinds of nodes having one K-dimensional vectors, such as a source vector to be quantized, and branch passes thereof. Two types of nodes are divided into non-terminal nodes indicated by '○' and terminal nodes denoted by '□'. Here, the right and left right and left branch paths of any non-terminal node have binary tag values (0, 1) for distinguishing the branch paths. In the terminal nodes, code vectors in which source vectors are quantized are located. Here, the VL-TSVQ technique is further considered such that the path length from the root node to each terminal node is different according to the frequency of occurrence of the K-dimensional source vector. The sequence of binary values representing the path from the root node to each terminal node is assigned to the index of the code vector located in the terminal node.

The vector quantization based on the VL-TSVQ codebook of FIG. 1 and its index allocation method will be described. First, the VL-TSVQ vector quantizer selects a terminal node in which a code vector most similar to an input K-dimensional source vector is located among the code vectors located in one of the terminal nodes (terminal nodes a to k) shown in FIG. 1. do. Next, the vector quantizer assigns a sequence of binary tag values indicated on a path from a root node to a selected terminal node as an index for identifying a vector quantized code vector. Description For example, it is assumed that two source vectors sequentially input to the vector quantizer are sequentially quantized to the terminal node a and the terminal node b shown in FIG. 1. In this case, binary bit values "00100" are assigned to the index of the code vector located in the terminal node a, and binary bit values "001010" are assigned to the index of the code vector located in the terminal node b. Accordingly, the TSVQ-based coding system compresses or transmits source information to be transmitted by transmitting or storing an index allocated according to a vector quantization of a source vector using a TSVQ codebook.

However, in the existing TSVQ-based coding system, whenever the quantized consecutive source vectors are coded by using the above-described TSVQ codebook and an index is assigned to each quantized code vector, the index is always set to the tree structure. By expressing the binary bits corresponding to the path from the root node to the terminal node, the compression encoding efficiency cannot be improved any more. In particular, the coding system based on the VL-TSVQ, although the coding efficiency is improved compared to the general TSVQ techniques, the coding efficiency can not be further improved for the reasons described above.

An object of the present invention for solving the above-described problems, in the above-described index assignment method for encoding based on tree structure vector quantization, the correlation between the source vectors to be continuously quantized vector or the code vectors according to the vector quantization By using the present invention, it is possible to provide an optimal index allocation method which significantly improves the encoding efficiency while maintaining the same image quality as compared to the conventional method.

1 is a view for explaining an existing index allocation method in a coding system based on a variable length-tree structure vector quantization (VL-TSVQ) codebook;

2 is a diagram briefly showing a TSVQ-based encoding system to which an index assignment method is applied according to an embodiment of the present invention;

3 is a view for explaining in detail the index allocation method and the system according to the present invention in the TSVQ-based encoding system of FIG.

4 is a view illustrating an example of obtaining a shortest path between two code vectors on a TSVQ codebook with reference to FIG. 3;

5 is a view showing an algorithm of a state machine (SM) for index allocation of the present invention with respect to FIG.

FIG. 6 is a view showing a list of finite state machines for each branch path on the obtained shortest path and an example of the description with respect to FIG. 5;

FIG. 7 is a diagram illustrating an index of a predetermined terminal node allocated according to the present invention with reference to an index according to a conventional scheme with respect to FIGS. 2 to 6.

8 is a diagram showing coding performance of an index allocation method of the present invention as compared to an existing index allocation method in a TSVQ-based encoding system;

9A and 9B are graphs showing coding gains according to a scheme of the present invention with respect to a conventional scheme, for various images having different correlation factors, and FIG. 9A is a graph when the block size K = 4. 9b is a graph of block size K = 16.

The index allocation method of the present invention for achieving the above object, in the index allocation method for encoding based on the tree structure vector quantization (TSVQ) codebook, (1) the source vectors to be continuously input, the tree structure vector Quantizing the corresponding codevectors based on the quantization codebook; And (2) assigning an index of a later quantized codevector among the codevectors using spatial correlation between the continuously quantized codevectors on the tree structure vector quantization codebook.

The technical idea of the present invention is based on the fact that there is a considerably high correlation between source vectors of a K-dimensional image to be generally quantized continuously. In this case, there is a considerably high spatial correlation between code vectors that are continuously quantized using the tree structure vector quantization codebook. Therefore, the present invention, in the tree structure vector quantization codebook, a value representing the relative position between two consecutively quantized codevectors, i.e., the shortest path between the terminal nodes where the codevectors are located, as an index of the quantized codevector. This greatly improves the compression encoding efficiency while maintaining the image quality.

In addition, it is apparent to those skilled in the art that the above-mentioned index assignment method is applied to a coding system based on variable length-tree structure vector quantization (VL-TSVQ), and that the coding efficiency will be further improved. The technical scope of the invention is not limited.

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

2 briefly shows a TSVQ-based encoding system to which an index allocation method according to an embodiment of the present invention is applied. As the Tree Structured Vector Quantizer (Q) shown in FIG. 2, a K-dimensional source vector Y _{i of an} image to be quantized is input. This source vector is obtained by dividing an image to be source coded into predetermined blocks having K-pixels, and then increasing the intensity of pixels in each block in the K-dimensional Euclidean space R ^K. It is obtained by expressing. The set of source vectors is represented by {Y _i , 1 ≦ _i ≦ T}, where T represents the number of source vectors.

Fig tree structure vector quantizer shown in 2 (Tree Structured Vector Quantizer; Q ) is a source vector of the input image (Y _i) a tree structure of the vector quantization (TSVQ) code in a particular terminal node of the codebook vector (Q ( Y _i ); quantize with X ^T (l _j , n _j )). The set of code vectors according to the quantization process of each source vector is represented by Equation 1 below.

Q (Yi) ∈ {XT (lj, nj), 1≤j≤N}, 1≤i≤T

Here, N represents the number of terminal nodes, and T represents the number of source vectors.

In FIG. 2, a transformer (H) extracts position information l _j , n _j of a terminal node where each code vector of Equation 1 is located, that is, quantized in a TSVQ codebook. This relationship is expressed as in Equation 2 below.

Where l denotes the node depth on the tree structure, and n denotes the corresponding terminal node number at the node depth.

The location information (l, n) of the terminal node is used as input data x _i of the differential index generator (DIG). This is expressed as Equation 3 below.

Therefore, when a finite sequence "x ₁ x ₂ ... x _T " is supplied to the differential index generator (DIG), the differential index generator is used to connect the terminal nodes where the quantized code vectors are sequentially located on the tree structure vector quantization codebook. Using the spatial correlation of, one finite sequence "y ₁ y ₂ ... y _T " is generated. Here, by varying the length of the differential index y _i output from the differential index generator, it is possible to enable variable-length coding.

FIG. 3 is a detailed view showing the configuration and operation of a differential index generator (DIG) in the TSVQ-based encoding system of FIG. 2, and specifically illustrates a preferred differential index allocation method and system according to the present invention. Referring to FIG. 3, the differential index generator (DIG) considers high correlation between adjacent image blocks, and previous input data (corresponding to terminal node position information) x _i-1 is a reference of the current input data x _i . Used as a reference. The main function of such a differential index generator (DIG) is represented by the following equation (4).

Referring to FIG. 3 and Equation 4, if two terminal node position information x _i-1 and x _i continuously input are identical to each other, that is, two adjacent blocks of an image are TSVQ codebooks (or VL-TSVQ codebooks). When successively quantized to the same terminal node, the differential index generator (DIG) determines and outputs the index y _i of the code vector located in the current terminal node as a 1-bit vector "v ₁ = 0".

On the other hand, if they do not coincide with each other, the Shortest Path Generator (S) of the differential index generator (DIG) determines the shortest path from the last terminal node position (x _i-1 ) to the current terminal node position (x _i ). Search and express primarily as the sequence "I _i = (d ₁ b ₁ ) (d ₂ b ₂ ) ... (d _{L (i)} b _{L (i)} ) = z ₁ z ₂ … z _{L (i)} " . This is summarized in Equation 5 below.

Here, L (i) represents the length of one sequence I _i corresponding to the shortest path between two consecutively quantized terminal nodes, and the sequence I _i consists of pairs of (d _k , b _k ). . Here, the shortest path described above is divided into individual branch paths between two nodes existing on the shortest path, and d _k represents a direction from the current node to the next node on the individual branch path. In a tree structure, if the next node is a lower node than the current node (that is, the direction of the individual branch path is downward), then "d _k = + 1", and if the next node is higher than the current node (that is, the direction of the individual branch path). Up) "d _k = -1". And, b _k represents a binary label (binary label) of the individual branch paths connecting two nodes (the left branch path is set to bit value "0" and the right branch path is set to bit value "1").

FIG. 4 shows an example in which the shortest path generator S shown in FIG. 3 obtains the shortest path between terminal nodes of two vector quantized code vectors on a TSVQ codebook. Referring to FIG. 4, UE node b from UE node a in which two consecutively quantized code vectors, X ^T (l _i-1 , n _i-1 ) and X ^T (l _i , n _i ), are respectively located. The shortest path up to is one sequence "I = (d ₁ b ₁ ) (d ₂ b ₂ ) (d ₃ b ₃ ) = (-1,0) (+ 1,1) (+ 1,0)" It is expressed as In this case, there are three individual branch paths that constitute the shortest path, and the length of the sequence I _i is determined by the number of individual branch paths. Therefore, the shorter the shortest path between two terminal nodes, i.e., the spatial correlation between source vectors to be continuously vector quantized using the TSVQ codebook or between consecutive vector quantized code vectors, thus increasing the vector quantization. The amount of data at the index assigned to the code vector can be reduced.

Meanwhile, the state mashine shown in FIG. 3 is further reduced by further reducing the data amount of the sequence I _i of the shortest path generated from the shortest path generator S and at the same time further considering the accurate recovery capability of the receiver. It is used for the sequence I _i of this shortest path. Each output sequence I _i of the shortest path generator S shown in FIG. 3 is input to the state machine SM.

The state machine SM, based on the algorithm of a finite-state machine, has one finite sequence "I _i = z ₁ z ₂ ... z _{L (i)} " corresponding to the shortest path found and its shortest. The binary bit sequence "yi = v ₁ v ₂ ... v _{L (i)} " is obtained based on the state values w _k related to the state of the path, and is assigned as the index of the current vector quantized code vector. This is summarized in Equation 6 below.

Then, the value indicating the next state value w _{k + 1} is, each of the path of the current of each of paths of the state machine (SM) having the state value w _k of the current private branch path constituting the shortest path to the above-described It is determined by (d _k , b _k ) and the current state value w _k , which is expressed by Equation 7 below.

In other words, the current state value w _k of the state machine SM is a value representing the state value w _k-1 of the previous individual branch path among the individual branch paths constituting the shortest path described above and the individual branch path ( d _k-1 , b _k-1 ). Here, when the state value w _k of the current individual branch path is set to the binary tag value b _k-1 of the previous individual branch paths (d _k-1 , b _k-1 ), the above equation (7) is It is summarized by Equation 8.

Here, the state value of the first individual branch path constituting the shortest path described above is set in advance to "w ₁ = 1".

FIG. 5 is a diagram of the state machine SM shown in FIG. 3 and shows a specific method for performing the functions of Equations 6 and 7. FIG. And the state machine SM is a deterministic finite-state machine, in which all states are clearly obtained and there is no uncertain state transition. Referring to FIG. 5, when the state machine receives a sequence I _i primarily expressing the shortest path described above, a pair of (direction, bit values) of each individual branch path constituting the sequence (that is, (d _k , b _k )) and the current state value (w _k ), assigns one binary bit value (0, or 1) (i.e., v _k ) for each pair, and the next state value (w _{k + 1} ) Determine. This operation is performed with respect to each pair of the input sequence Ii, so that the sequence y _i = v ₁ v ₂ ... Of one binary bit value corresponding to one shortest path. Generates _{L (i)} and assigns it to the index of the current vector quantized codevector. The state table which is the basis of the operation of the functions of the state machine SM shown in FIG. 5 is shown in Table 1 below.

Current state value (w _k ) Output bit value (v _k ) / next state value (w _{k + 1} ) z _k = (-1,0) (+1,0) (-1,1) (+1,1) 0 0/0 0/0 0/1 1/1 One 1/0 0/0 1/1 1/1

With reference to Table 1 and FIG. 5 described above, for each individual pair (z _k = (b _k , d _k )) forming the sequence I _i input to the state machine, all possible next pairs (zk +) 1) and its output bit values v _{k + 1} are shown in FIG. 6. In particular, the state of the third line and the state of the fourth line in the list shown in FIG. 6 were respectively shown as tree structures on the right and left sides based on the list. In summary, the number of possible next inputs z _{k + 1} for any particular input z ^* _k is two. Then, the next possible output bit values v _{k + 1} for each of the two z _{k + 1} have different bit values (0 or 1).

For example, as shown in FIG. 7 to more clearly understand the index allocation method according to the present invention with reference to FIGS. 2 to 6. In the illustrative example of FIG. 7, when two source vectors are sequentially vector quantized to the terminal node a and the terminal node b shown in FIG. 1 by the TSVQ codebook, the index of the code vector located in the vector quantized terminal node b will be described later. Is shown in contrast to the existing scheme and the present invention. According to the conventional index assignment scheme, the index of the terminal node b is allocated as a sequence " 001010 " of binary bit values representing a path from the root node of the TSVQ codebook to the terminal node b. On the other hand, according to the index allocation method of the present invention, the index of the terminal node b is a sequence of binary bit values representing the shortest path from the terminal node a to the current terminal node b, which is the result of vector quantization immediately before, in the TSVQ codebook. 110 ". Therefore, it can be seen that the compression encoding efficiency is greatly improved by significantly reducing the data amount of the index according to the present invention rather than the index of the code vector according to the conventional scheme.

Next, it is examined whether the receiver can correctly restore the original code vector from the index of the code vector according to the method of the present invention.

First, successive source vectors are vector quantized to terminal nodes in the TSVQ codebook, and the position index I _i = (l _j , n _j ) of the terminal nodes is shown in FIG. 2 and FIG. 3. Consecutive codevectors, except for the first codevectors, are assigned as differential indexes. Here, between two consecutive vector quantized code vectors, the specific terminal node information of the previous vector quantized code vector is x _p = (l _p , n _p ) and the current vector quantized code vector is located on the tree structure codebook. It is apparent to those skilled in the art based on the description of the preferred embodiment of the present invention that, if one of the terminal nodes is selected as one of the different terminal nodes, different differential indexes are assigned to each of the different terminal nodes. This is summarized in Equation 9 below.

Second, among the different differential indexes of each of the immediately different terminal nodes represented by the shortest path from the common (or fixed) previous terminal node, as described above, the bits of any particular differential index are different. It cannot be a prefix of bits in the differential index. This is because any particular differential index is a prefix of the other indexes, because the particular differential index becomes a parent node of the other indexes and cannot be a terminal node. This is obvious to those skilled in the art from the structure of the tree structure vector quantization codebook shown in FIG. 1 and the foregoing descriptions.

Therefore, based on the first and second descriptions of the foregoing, any differential index x _i allocated according to the present invention is the terminal node position information x _p of the code vector reconstructed just before the decoder side and any current difference. From the index y _i , it may be uniquely restored to a code vector located at a specific terminal node. And, the specific algorithm of the restoration process is made through the reverse process of the above-described differential index allocation method of the present invention, which will be apparent to those skilled in the art will not be described in detail.

Now, the coding performance of the TSVQ-based coding system to which the differential index allocation method according to the present invention is applied will be described with reference to computer simulation results comparing the corresponding conventional methods.

In this simulation, 40 magnetic images (MR) images (512 × 512 size) were used as training images. Each pixel of the MR images is represented by 12 bits. The distortion of the compressed image is measured by the peak-to-noise ratio (PSNR). The test image is a 512 × 512 MR image different from the above-described training image.

8 shows coding performance of the index allocation method of the present invention as compared to the existing index allocation method in a TSVQ-based encoding system. In Fig. 8, the horizontal axis represents PSNR and the vertical axis represents the amount of bits bpp required per pixel, respectively. In addition, "VLTSVQ + CI" and "VLTSVQ + DI" quantize MR images using the VL-TSVQ technique, and thus select terminal nodes selected according to the conventional index (CI) and the present invention. The result of encoding each of the differential index (DI) is shown. In addition, "PTSVQ + CI" and "PTSVQ + DI" quantize MR images using PTSVQ (Pruned-TSVQ) techniques known in the art, and thus select terminal nodes selected according to the conventional index (CI) and the proposal. The result of each encoding with the differential index DI shown is shown. In addition, "VLTSVQ + Hilbert + DI" is to further improve the correlation characteristics between successive source vectors, and further applies the Hilbert Scan technique known in the art to the proposed "VLTSVQ + DI". Shows the result of the encoding. According to FIG. 8, the coding system employing the proposed DI has a much better coding performance in the entire area of the PSNR than the coding system employing the conventional CI. And, this feature shows that the total amount of bits required for "VLTSVQ + DI" takes much less than that of "PTSVQ + CI". This means that the differential index (DI) technique is more efficient at reducing the amount of coded bits than the 'segmentation and prediction technique'. The proposed coding system (DI) reduces the coded bit rate without reducing the quality of the image at all.

Next, most of the computational load of the coding scheme proposed in the present invention is in the quantization step of the source vector and the differential index generation (DIG) step. According to the simulation by the inventors of the present invention, the computational complexity of the proposed coding system using the differential index (DI) technique is estimated to be approximately three times that of the conventional coding system.

In addition, in order to measure the relationship between the performance of the differential index technique and the correlation factor of the source vectors (Y _i 's), several MR images having different correlation factors were tested, and the results are shown in FIGS. 9A and 9B. It is shown in Figure 9b.

9A and 9B, the vertical gain (G;) represents the difference between the total amount of bits required for each of the conventional index allocation technique (CI) and the proposed differential index allocation technique (DI), and the total bits of the CI. Divide by quantity and express it as a percentage. The horizontal axis C _Y represents a correlation factor of an image (particularly, an image of a medical field; in general, a medical image has a high correlation between adjacent blocks). And the simulations shown in Figs. 9A and 9B were performed for each block size of 2x2 (K = 4) and 4x4 (K = 16).

9A and 9B, it can be seen that as the correlation factor G _Y of the horizontal axis increases, the coding gain G of the vertical axis increases proportionally. This means that the proposed DI technique is more efficient in proportion to the correlation between successive blocks. 9A and 9B, since the correlation factor G _Y decreases as the block size K increases, the coding gain of the proposed DI decreases as the block size increases.

On the other hand, it is apparent to those skilled in the art that the index allocation proposed in the present invention and the encoding applied thereto may be implemented in software and hardware. In addition, it is also apparent to those skilled in the art that a number of modified embodiments which are not mentioned from the above-described preferred embodiments can be derived based on the technical idea of the present invention.

As described above, the present invention, in the encoding based on the tree structure vector quantization (TSVQ), by assigning the index by using the correlation between the source vectors to be continuously vector quantized or between the consecutive vector quantized code vectors, Compared with the existing index allocation method and its coding system, it provides significantly improved coding efficiency. More specifically, the present invention assigns the shortest paths between terminal nodes in which consecutive vector quantized codevectors are located as indexes, thereby significantly maintaining the encoded data amount while maintaining the same image quality compared to the conventional method. Enable compression Therefore, when the index allocation method and the encoding method proposed in the present invention are employed in a communication system (for example, image information transmitting and receiving device), the performance (transmission time and required memory capacity) of the communication system can be significantly improved. have. In particular, when the present invention is applied to signal source encoding of an image having a large correlation between adjacent image blocks such as a medical image, the encoding performance can be further improved.

Claims (9)

In the index allocation method for coding based on the tree structure vector quantization (TSVQ) codebook, (1) quantizing successively input source vectors into corresponding code vectors based on the tree structure vector quantization codebook; And (2) assigning an index of a later quantized codevector among the codevectors using spatial correlation between the continuously quantized codevectors on the tree structure vector quantization codebook. The method of claim 1, wherein the index assignment method is applied to encoding based on a variable length-tree structure vector quantization (VL-TSVQ) codebook. 2. The method of claim 1, wherein the index assigned in step (2) is determined by a value representing a shortest path between two consecutive code vectors, located on a tree structure vector quantization codebook. The method of claim 1, wherein the (2) step is (2a) searching for the shortest path in the tree structure vector quantization codebook from the position of a previously quantized first code vector to the position of a second quantized code vector among the two consecutively quantized code vectors. step; And (2b) representing the searched shortest path as a sequence of binary values by using a predetermined characteristic of the tree structure vector quantization codebook and assigning the index to the index. The method of claim 4, wherein in the step (2a), the shortest path is divided into individual branch paths between two nodes located on the shortest path. The individual branch paths are pairs of direction values indicating directions of individual branch paths from the current node to the next node and binary values (0, 1) indicating tags of the individual branch paths (direction values of individual branch paths, individual Binary value of the branch path) And said shortest path to be searched is represented as a sequence of said pairs for each individual branch path constituting said shortest path. 6. The method of claim 5, wherein the direction value of the individual branch paths is a value that distinguishes whether a next node is a lower node or a higher node than a current node in a tree structure. The method of claim 4, wherein step (2b) A sequence of pairs of individual branch paths (the direction values of the individual branch paths, the binary values of the individual branch paths) forming the shortest path between the two code vectors, and the individual branch paths forming the shortest path on the tree structure vector quantization codebook. And determining a sequence of the binary values based on a state value according to a state of the equation. 8. The method of claim 7, wherein the length of the sequence of binary values corresponds to the number of individual branch paths that make up the shortest path. The method of claim 1, wherein the index assigned in step (2) is: And an index assignment method for reconstructing the original code vector designated by the index using the spatial correlation associated with the tree structure vector quantization codebook and its index.