KR970003105B1 - Video signal coding apparatus urilizing adaptive pattern quantization and quantization pattern selecting method - Google Patents

Abstract

The image coding device using adaptive pattern quantization comprises a pattern quantization means for M x N difference signal between expected frames and a coder for coding the pattern quantized image. The pattern quantization means includes: a quantization pattern code book(100) having a plurality of N x N quantization patterns constituted by an one-dimensional quantizer(Qk); a quantization pattern selector(60) for selecting a quantization pattern in which an error value is lowest; and a quantizer(50) for performing pattern quantization for a delayed difference signal supplied from a delayer(90) which delays the input signal(30) for a predetermined time period, on the basis of the selected optimal quantization pattern suppled from the quantization pattern selector(60).

Description

Image Coding Device Using Adaptive Pattern Quantization and Quantization Pattern Selection Method

1 is a block diagram of an image encoding apparatus using adaptive pattern quantization according to a preferred embodiment of the present invention.

2 is a flowchart illustrating a process of selecting arbitrary optimal quantization patterns targeted from a plurality of quantization patterns according to a preferred embodiment of the present invention in frame prediction block coding.

* Explanation of symbols for main parts of the drawings

10: difference image encoder 20: motion compensator

30: motion estimator 40: frame memory

50: quantizer 60: quantization pattern selector

70: difference image decoder 80: variable length encoder

90: delay unit 100: quantization pattern codebook

110: subtractor 120: adder

[Purpose of invention]

[Technical Field to which the Invention belongs and Prior Art in the Field]

The present invention relates to image coding and techniques, and more particularly, to an image encoding apparatus suitable for using an adaptive pattern quantization technique having a plurality of quantizers in frame prediction block encoding, and a quantization pattern selection method therein.

Recently, various techniques for effectively compressing data in order to effectively transmit digital video signals through conventional communication channels have been proposed. Compressor methods commonly used among them are, for example, transform coding schemes for reducing intra-frame correlation such as Discrete Cosine Trasform (hereinafter referred to as DCT) and temporal interframes such as motion compensation. It is a differential image coding method that reduces correlation. The high-definition television (HDTV) system currently being developed uses both the above-described DCT and video encoding.

In this case, motion compensation refers to a signal of a previous frame (or field) in a motion vector, that is, a pixel or a block of pixels of a current frame in a motion image signal by estimating the degree of motion of an object in a video signal processing by a predetermined algorithm. The amount of movement in these directions relative to the previous frame is expressed by the vector amount of the pixel unit and moved by the motion vector.

In addition, the differential image encoding method is an image compression method for encoding by using the above-described motion compensation, and compares the previous frame and the current frame to estimate how much the image of the current frame has moved compared to the previous image, and estimates the estimated motion vector. Motion compensation is performed in the previous frame as described above. The motion-compensated previous frame is motion-compensated similarly to the current frame by the motion vector, and subtracts the motion-compensated frame from the current frame to form a difference signal. This difference signal is transform-coded as a way to reduce intraframe correlation.

In general, when encoding using a quantization pattern in frame prediction block encoding, coding efficiency may vary according to a method of selecting a scalar quantizer constituting each quantization pattern. There is an urgent need for a method of selecting an easy and efficient quantizer and configuring an optimal quantization pattern.

[Technical problem to be achieved]

Accordingly, an object of the present invention is an adaptive pattern in frame prediction block coding, which can improve compression coding efficiency by using a quantization pattern codebook including a plurality of different quantization patterns composed of combinations of primary quantizers. An image encoding apparatus using quantization is provided.

Another object of the present invention is to provide an optimal quantization pattern codebook suitable for constructing a plurality of different quantization patterns composed of combinations of necessary one-dimensional quantizers when performing vector quantization in frame prediction block coding. To provide a method for selecting a quantization pattern.

In accordance with the present invention, a pattern of MxN difference signals between a current frame and a prediction frame obtained through motion estimation and compensation using a previous frame reconstructed after the current frame and the pattern quantization is obtained. A video encoding apparatus comprising means for quantizing and means for variable length encoding the pattern quantized video, wherein the pattern quantization means comprises: a QD {Q ⁰ } {{Q ¹ }} {Q ² }} … Q k ^-1 (where Q ⁰ is a one-dimensional quantizer with a minimum representative value and Q ^N-1 is a one-dimensional quantizer with a maximum representative value) (Q ^k A quantization pattern codebook comprising a plurality of different M × N quantization patterns composed of combinations of the following; A delay unit for delaying the M × N difference signal for a predetermined time; When the M × N difference signal is quantized among a plurality of quantization patterns included in the quantization pattern codebook, a pattern having the smallest error value is selected as an optimal quantization pattern for quantization, and an optimal quantization pattern is defined in the quantization pattern codebook. A quantization pattern selector for selecting a pattern having a different number of quantization bits and having the same quantization noise as the optimum quantization pattern and providing the variable length encoding means with an exponent number of the selected optimal quantization pattern; And a quantizer for pattern quantizing a delayed difference signal provided from the delay unit based on the selected optimal quantization pattern provided from the quantization pattern selector. .

According to another aspect of the present invention, there is provided a quantization system for vector quantizing an arbitrary image signal using a quantization pattern codebook having a plurality of M × N quantization patterns composed of combinations of one-dimensional quantizers. In the method for selecting the plurality of M × N quantization patterns used in the above-mentioned method, the one-dimensional quantizer includes: {Q ⁰ } ⊂ {Q ¹ } ⊂ {Q ² } ⊂. {Q ^N-1 }, where Q ⁰ is a one-dimensional quantizer having a minimum representative value and Q ^N-1 is a one-dimensional quantizer having a maximum representative value, and 1 in the quantization pattern codebook. A first step of selecting each one-dimensional quantizer Q ^k when the set of representative values of the dimensional quantizer Q ^k is represented by the following equation; k = 0,1... N-1 (where N is the number of one-dimensional quantizers and Z _k is the number of representative values of Q ^k .) Bits required for encoding for each representative value (y ^k ) of the selected one-dimensional quantizer. Allocating a number B (y ^k ), respectively; By selecting each one-dimensional quantizer that generates the minimum noise for each element in each of the plurality of training vectors obtained by dividing an arbitrary image into M × N blocks, a plurality of M × N quantization patterns respectively corresponding to the respective training vectors are selected. Generating a third step; A fourth step of assigning each of the generated plurality of quantization patterns to each initial cluster in a cluster group; A fifth step (S18), for each of the clusters in the cluster group, calculating the total quantization bit increase rate occurring when merging the quantization patterns belonging to any two clusters with all possible two cluster combinations; Merge two quantization patterns belonging to two clusters of which the calculated total quantization bit number increase rate is the minimum among each of the clusters to generate one new quantization pattern, and convert the generated one new quantization pattern into one new cluster. A sixth step of merging each of the two quantization patterns with a one-dimensional quantizer having a larger representative value among two corresponding one-dimensional quantizers in the two quantization patterns; Removing each cluster of the merged two quantization patterns from the cluster group to belong to the one new cluster; It is checked whether the number of residual clusters included in the cluster group matches the predetermined number of target quantization patterns, and until the number of residual clusters included in the cluster group reaches the predetermined target quantization pattern number. An eighth step of repeatedly performing the fifth to seventh steps; And if it is determined that the number of remaining clusters included in the cluster group matches the predetermined target quantization pattern number as a result of the check in the eighth step, for configuring the quantization pattern codebook, each of the residual clusters belonging to each of the residual clusters is included. A quantization pattern selection method comprising a ninth step of extracting respective quantization patterns is provided.

[Configuration and Function of Invention]

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

1 is a block diagram of an image encoding apparatus using adaptive pattern quantization according to a preferred embodiment of the present invention. As shown in the drawing, the video encoding apparatus of the present invention includes a differential image encoder 10, a differential image decoder 70, an adder 120, a frame memory device 40, a motion estimator 30, and a motion compensator. 20, a subtractor 110, and a variable length encoder 80.

In addition, the difference image encoder 10 forming the biggest technical gist of the present invention includes a quantization pattern codebook 100, a quantization pattern selector 60, a delay unit 90, and a quantizer 50.

Referring to FIG. 1, the subtractor 110 subtracts an input frame signal and a predicted current frame signal provided from a motion compensator 20 to be described later, and subtracts the difference image (error signal) onto a line 130. The difference image generated on the line 130 is provided to the quantization pattern selector 60 and the delay unit 90, respectively.

On the other hand, the quantization pattern selector 60 is selected from a plurality of M × N quantization patterns (for example, 4 × 4 quantization patterns composed of one-dimensional quantizers) selected according to the present invention and included in the quantization pattern codebook 100. When the M × N input signal (difference image) on the line 130 is quantized, an optimal quantization pattern having the least error is selected, and the selected optimal pattern is provided to the quantizer 50 described below for quantization. Here, a detailed method of selecting quantization patterns constituting the quantization pattern codebook according to the present invention will be described later in detail.

In this case, when the optimal quantization pattern is not in the quantization pattern codebook 100, the quantization pattern selector 60 selects the pattern as the optimal quantization pattern if the quantization noise is the same even though the number of quantization bits is different.

In addition, the quantization pattern selector 60 transmits the exponent number for each selected quantization pattern to the variable length encoder 80 described later. In addition, the quantization pattern codebook 100 includes a plurality of different quantization patterns composed of combinations of one-dimensional quantizers.

On the other hand, the delay unit 90 forming the encoding path of the difference image delays the line 130 input signal (difference image) for a predetermined time and then outputs the result to the quantizer 50. In the quantizer 50, the quantization pattern described above. The delayed differential image signal provided from the delay unit 90 is quantized using the selected quantization pattern provided from the selector 60, and thus the quantized image signal is transferred to the differential image decoder 70 and the variable length encoder 80. Provided at the same time.

On the other hand, the difference image decoder 70 is a reverse process of the signal processing performed by the difference image encoder 10 described above, and is performed for motion estimation and motion compensation used to reduce spatial correlation between frames. The difference image quantized through 50 is restored to the original signal before being quantized and provided to the next adder 120.

Next, the adder 120 adds the reconstructed signal (difference image) provided by the difference image decoder 70 and the motion compensated prediction signal provided from the motion compensator 20 to be described later to reconstruct the previous frame. A signal is generated, and the reconstructed previous frame signal is stored in the frame memory 40. Therefore, the previous frame signal for every frame to be encoded is continuously updated through this process, and the reconstructed previous frame signal is updated to the motion estimator 30 and the motion compensator 20 for motion estimation and compensation. Will be provided.

Therefore, the motion estimator 30 estimates the motion using a technique such as a pixel cyclic algorithm or a block matching algorithm based on the input current frame signal and the reconstructed previous frame signal provided from the frame memory 40, that is, the current frame signal. Each pixel or MxN pixel block (e.g., 8x9 blocks) in the frame is searched within the reconstructed previous frame and then estimated by motion vectors between each pixel or MxN pixelblock to the motion compensator 20. to provide. Here, the motion vector is a displacement between each pixel or M × N block of the current frame and each pixel or M × N candidate block predicted in a predetermined search range of the previous frame.

In addition, although the detailed illustration in FIG. 1 is omitted, the estimated motion vectors are provided to the variable length encoder 80 described later. Accordingly, the quantized signal provided by the quantizer 50 as well as the motion vectors provided by the motion estimator 30 in the variable length encoder 80 will be variable length encoded and transmitted to the next coder. .

Meanwhile, the motion compensator 20 generates a motion compensated prediction frame signal based on the estimated motion vectors and the reconstructed previous frame signal provided from the frame memory 40 as described above, that is, by the estimated motion vector. The reconstructed previous frame signal is moved to generate a predictive frame signal, and the generated predictive frame signal is provided to the above-described subtractor 110 and adder 120, respectively.

Therefore, the difference image between the current frame and the prediction frame is continuously generated from the subtractor 110 through the processing path as described above, and the difference images are efficiently pattern quantized by the difference image encoder 10 according to the present invention. do.

Next, a process of selecting an optimal quantization pattern according to the present invention when vector quantizing a difference image in the image encoding apparatus having the above-described configuration will be described in more detail.

Referring to FIG. 1, assuming that an input signal (difference image) on a line 130 is composed of 4 × 4 blocks as an example, the quantization pattern codebook 100 has a plurality of forms in which one-dimensional quantizers are arranged in 4 × 4. It consists of quantization patterns of. That is, assuming that there are four types of one-dimensional quantizers, Q ⁰ , Q ¹ , Q ^2, and Q ^3, the following table shows a preferable arrangement of the quantization pattern.

TABLE 1

As described above, the quantization pattern codebook 100 is a combination of the one-dimensional quantizer as described above when the input signal (difference image) on the line 130 provided from the subtractor 110 of FIG. In general, such a codebook is composed of hundreds to thousands of different quantization patterns. In addition, the quantization patterns prepared in the quantization pattern codebook 100 are described according to the present invention described below.

Accordingly, the quantization pattern selector 60 of FIG. 1 selects the pattern having the minimum error when the input signal on the line 130 is quantized among the plurality of quantization patterns in the codebook 100, that is, with respect to the input difference image signal. An optimal quantization pattern in the pattern codebook 100 is selected. However, since the input difference image signal is not input only training vectors used in constructing the codebook, when there is no optimal quantization pattern, the most similar pattern among the patterns in the pattern codebook 100 may be selected as the optimal pattern. . In this case, when the configuration requirements of the one-dimensional quantizer Q ^k , such as the following equation, are satisfied, the quantization error may be minimized.

[Equation 1]

{Q ⁰ } ⊂ {Q ¹ } ⊂ {Q ² } ⊂… Q {Q ^N-1 }

In other words, if a pattern having a low quantizer index k is required, but such a pattern does not exist, the one having a high quantizer index k is replaced. For example, suppose that the pattern desired by the user is a pattern like A in the following table. If there is no such pattern, a similar pattern like B is selected as an optimal pattern.

TABLE 2

In the above table, the quantization noise nq of the A pattern and the B pattern is the same. That is, the quantization noise can be regarded as an optimal quantization pattern in both patterns, and the number of quantization bits is increased but the actual number of quantization bits can be effectively used in such a case. However, if the quantizer configuration requirements as described above are not satisfied, the quantization noise (nq) of the patterns A and B is not the same, so at this time, the optimal quantization pattern cannot be selected despite the quantization bit number.

In addition, the quantization pattern selector 60 uses a variable length encoder to determine the index number (number indicating the number of patterns) of the selected quantization pattern so that the selected quantization pattern can be recognized by the decoding apparatus at the receiving side as described above. Forward to the receiving end via 80. Accordingly, the quantizer 50 quantizes the difference signal input through the line 130 using the selected quantization pattern as described above and delayed for a predetermined time by the delay unit 90.

In addition, if the 1-dimensional quantizer is expressed as Q ^k (x) = y ^k , this equation shows the value of x among the representative values of the Q ^k quantizer when quantizing the difference signal and one element x into the Q ^k quantizer. This means that it is represented by the representative value y ^k which is the closest to, that is, the error is minimized. Accordingly, when quantizing the signal 130 input to the quantizer 50, each element of the input signal on the line 130 is quantized by a quantizer corresponding to each element of the pattern.

For example, suppose that the first quantization pattern is quantized by the first quantization pattern and the input difference signal is configured as shown in the following table.

TABLE 3

At this time, the difference signal is quantized using the quantization pattern selected by the quantizer 50 is shown in the following table.

TABLE 4

Then, the quantized values as described above are variable length coded through the variable length encoder 80 and then sequentially transmitted to a coding unit not shown in FIG. 1.

Therefore, the receiving side decoding apparatus receives the quantization pattern codebook 100 having the exponent number of the quantization pattern, where the quantization pattern codebook included in the receiving side decoding apparatus is the same as the quantization pattern codebook included in the transmitting side encoding apparatus. A quantization pattern corresponding to the index number is selected, and the quantized value transmitted using the selected quantization pattern is restored to the original signal before quantization. In this case, the decoding process of the differential image decoder 70 of FIG. 1 is performed in the same manner as the differential image encoding 10 described above.

Next, a method for selecting a quantization pattern included in the pattern codebook employed in the video encoding apparatus having the above-described configuration for the bettor quantization according to the present invention will be described in detail with reference to FIG.

2 is a flowchart illustrating a process of selecting an arbitrary optimal quantization pattern prepared for vector quantization in a plurality of quantization patterns according to a preferred embodiment of the present invention.

2 Referring to FIG., The step (S10) in a set ({Q ^k}) of the representative value in the one-dimensional quantized (Q ^k) one-dimensional quantizer Q ^k, in selecting the representative values of the selected following formula do.

[Formula 2]

Where k is 0, 1,... Is N-1, N is the number of one-dimensional quantizers Q ^k , and Z _k is the number of representative values of Q ^k . At this time, the configuration requirements of the quantizer Q ^k are as follows.

[Equation 3]

{Q ⁰ } ⊂ {Q ¹ } ⊂ {Q ² } ⊂… Q {Q ^N-1 }

If the configuration conditions such as the above equation are satisfied, the quantization pattern codebook 100 can be optimally configured.

Next, in step S12, the number of bits required for encoding the respective representative values Y ^k selected in step S10 described above. Assign each one. The reason for assigning the number of bits at the time of encoding to each representative value in this step S12 is for the calculation of the total quantization bit number increase rate required for determining the quantization pattern merge in step S18 described later.

Meanwhile, in step S14, a plurality of training vectors obtained by dividing an arbitrary image input for selecting a quantization pattern into M × N blocks are selected. For example, assuming that the input vector is a training vector having a size of 4 × 4, one-dimensional quantizer is selected for each of 16 elements, and each M × N is selected by selecting one-dimensional quantizer for each element. A number of M × N quantization patterns respectively corresponding to the training vectors is generated.

In this case, since the difference signal input for encoding is a general TV signal, and the type of a general TV image is infinite, a sample of a portion of the infinite signal is statistically extracted to generate a quantization pattern codebook 100, which participates in constructing the codebook. It is limited to the assumption that all of the remaining infinite values are not dependent on the characteristics of the extracted samples. Therefore, the quantization pattern codebook 100 composed of a merge algorithm using the sampled signal satisfies an optimal condition for a training vector but does not satisfy an optimal condition for all signals. However, since all other signals are assumed to conform to the characteristics of the extracted samples, the quantization noise (nq) is calculated by the following equation when selecting the one-dimensional quantizer.

[Equation 4]

Nq = d {W ₁ (1), Q ^k (W ₁ (1)}

In other words,

Same as

In the above equation, d is a distortion measurer, W (1) represents a training vector, Q ^k (W ₁ (1) represents a representative value when W ₁ (1) is quantized to Q ^k , and the distortion value). Measured d represents the distortion value between W ₁ (1) and Q ^k (W ₁ (1), where L is the total number of training vectors and S is the number of blocks. For each element W ₁ (1), choose a one-dimensional quantizer that generates the minimum quantization noise, where the quantizer index (k) is the best if there are multiple quantizers that produce the minimum quantization noise. A small quantizer is selected, thus generating an initial quantization pattern P ₁ (1) corresponding to each element W ₁ (1) with the selected quantizer index k.

In this case, the present invention introduces a cluster concept to calculate the total number of bits continuously increasing through the quantization pattern merge described below, where each cluster represents all training vectors (or quantization patterns) merged into the cluster.

Therefore, in step S16, each of the plurality of quantization patterns generated corresponding to each training vector belongs to each initial cluster C ^{0 in the} cluster group as in the following equation. That is, each training vector (each generated quantization pattern) belongs to one corresponding cluster in the cluster group, respectively.

[Equation 5]

℃ (1) = W ₁ (1)

Where 1 = 0, 1, 2... Is L-1, and L is the total number of training vectors.

Next, in step 18, for each of the clusters in the cluster group, the total quantization bit increase rate that occurs when merging the quantization pattern belonging to any two clusters is computed for each of the two possible cluster combinations, respectively. That is, two random quantization patterns are selected from the existing quantization patterns, that is, two ^t quantization patterns P ^t (n) and P ^t (m) (n ≠ m, n <m) are selected to be P ^t (n, m Calculate the total quantized bit growth rate (IB ^t (n, m)) when merged with) for all possible combinations (i.e., all possible cluster combinations) as shown in the following equation.

[Equation 6]

In the above formula, BC ^t (n, m) represents the total number of quantization bits, Is a quantization pattern representing elements belonging to one cluster.

Meanwhile, in step S20, two quantization patterns belonging to two clusters of which the calculated total quantization bit increase rate is minimum among each cluster in the cluster group are merged, that is, all total quantization bit increase rate values of each quantization pattern pair are merged. A pair of quantization patterns, i.e., quantization pattern pairs of (n, m) and Is merged into the n side, that is, P ^{t = 1} (n) as shown in the following equation, and the m side pattern is erased (if there is a condition of n &lt;

[Formula 7]

Where i = 0, 1,... , S-1, and the remaining merged other yangization patterns remain as follows.

Equation 8

Further, in step S20, one new quantization pattern is generated through the above-described process, and the generated one new quantization pattern belongs to one new cluster.

In this case, a merge between two quantization patterns belonging to two clusters is respectively merged into a one-dimensional quantizer having a larger representative value among two corresponding one-dimensional quantizers in the two quantization patterns.

Then, in step S22, one new quantization pattern is generated through the merge between the quantization pattern pairs in step S20, and the generated one new quantization pattern belongs to one new cluster. In this case, two merged quantization patterns, that is, two clusters to which the merged two quantization patterns belong, are removed from the cluster group. Instead, one new cluster belonging to the newly generated quantization pattern is generated and included in the cluster group.

Next, in step S24, the two clusters merged through the above-described steps S18 to S22 are removed from the cluster group, and instead, the remaining clusters currently remaining in the cluster group including one new cluster, that is, It is checked whether the number of residual clusters to which one quantization pattern (or at least one merged quantization pattern) belongs corresponds to a predetermined number of predetermined quantization patterns targeted for the construction of the preferred quantization pattern codebook.

As a result of the check in the step S24, it is determined that the number of remaining clusters (ie, the number of merged and remaining quantization patterns) remaining in the current cluster group does not match the predetermined number of target quantization patterns for pattern selection. That is, if it is determined that the number of remaining clusters is larger than the predetermined number of target quantization patterns, the process returns to the above-described step S18 and repeats the subsequent steps S18-S24, and each of these steps (S18-S24). Iterative execution of) continues until the number of remaining clusters remaining in the cluster group becomes the predetermined target both and the number of patterns.

On the other hand, as a result of the check in step S24, it is determined that the number of remaining clusters (ie, the number of merged and remaining quantization patterns) remaining in the current cluster group matches the predetermined number of preset quantization patterns to be selected for the pattern selection. In operation S26, quantization patterns having a target number finally selected for optimal vector quantization among a plurality of training vectors are extracted.

Accordingly, by adopting the optimal quantization patterns selected through the process as described above according to the present invention in the image encoding apparatus having a vector quantization technique, the M input from the delay unit 90 in the quantizer 50 of FIG. The quantized signal is output to the next difference image encoder 70 and the variable length encoder 80 by effectively vector quantizing the delayed difference images in × N block units.

[Effects of the Invention]

As described above, the present invention uses a quantization pattern codebook using a merge algorithm and an optimal quantizer pattern selection technique to select an easier and more efficient quantization pattern and to configure an optimal quantization pattern in frame prediction block encoding. There is an effect that the coding efficiency can be improved.

Claims (3)

Means for pattern quantizing a MxN difference signal between a current frame and a prediction frame obtained through motion estimation and compensation using a previous frame reconstructed and reconstructed after this current frame and pattern quantization, and means for encoding the pattern quantized image. In the video encoding apparatus comprising: the pattern quantization means: the one-dimensional quantizer is {Q ⁰ } ⊂ {Q ¹ } ⊂ {Q ² } Q k ^-1 (where Q ⁰ is a one-dimensional quantizer with a minimum representative value and Q ^N-1 is a one-dimensional quantizer with a maximum representative value) (Q ^k A quantization pattern codebook 100 comprising a plurality of different M × N quantization patterns composed of combinations of a plurality of quantization patterns; A delay unit 90 for delaying the M × N difference signal for a predetermined predetermined time; When the MXN difference signal is quantized among a plurality of quantization patterns included in the sound quantization pattern codebook 100, a pattern having the smallest error value is selected as an optimal quantization pattern for quantization, and is included in the quantization pattern codebook 100. A quantization pattern selector 60 for selecting a pattern having a different number of quantization bits and having the same quantization noise as the optimal quantization pattern and providing an exponent number of the selected optimal quantization pattern to the variable length encoding means; And a quantizer 50 for pattern quantizing the delayed difference signal provided from the delay unit 90 based on the selected optimal quantization pattern provided from the quantization pattern selector 60. Image coding apparatus using quantization. Selecting a plurality of M × N quantization patterns used in a quantization system for vector quantizing arbitrary video signals using a quantization pattern codebook having a plurality of M × N quantization patterns composed of combinations of one-dimensional quantizers. In the method, the one-dimensional quantizer is, {Q ⁰ } ⊂ {Q ¹ } ⊂ {Q ² } ⊂. {Q ^N-1 }, where Q ⁰ is a one-dimensional quantizer having a minimum representative value and Q ^N-1 is a one-dimensional quantizer having a maximum representative value, and 1 in the quantization pattern codebook. Selecting each one-dimensional quantizer (Q ^k ) when the set of representative values of the dimensional quantizer (Q ^k ) is expressed by the following equation (S10); k = 0,1... N-1 Where N is the number of one-dimensional quantizers and Z _k is the number of representatives of Q ^k . Allocating the number of bits (B (y ^k )) necessary for encoding for each representative value (y ^k ) of the selected one-dimensional quantizer (S12); By selecting each one-dimensional quantizer that generates the minimum noise for each element in each of the plurality of training vectors obtained by dividing an arbitrary image into M × N blocks, a plurality of M × N quantization patterns respectively corresponding to the respective training vectors are selected. Generating (S14); Assigning each of the generated plurality of quantization patterns to each initial cluster in a cluster group (S16); Calculating, for each of the clusters in the cluster group, a total quantization bit increase rate that occurs when merging a quantization pattern belonging to any two clusters in all two cluster combinations (S18); Merge two quantization patterns belonging to the number clusters of which the calculated total quantization bit number increase rate is the minimum among each of the clusters to generate one new quantization pattern, and convert the generated one new quantization pattern into one new cluster. Merging each of the two quantization patterns with a one-dimensional quantizer having a larger representative value among two corresponding one-dimensional quantizers in the two quantization patterns (S20); Removing each cluster of the merged two quantization patterns from the cluster group to belong to the one new cluster (S22); It is checked whether the number of remaining clusters included in the cluster group matches the predetermined number of target quantization patterns, and until the number of remaining clusters included in the cluster group reaches the predetermined target quantization pattern number. Repeating steps S18 to S22 (S24); And if it is determined in the step S24 that the number of residual clusters included in the cluster group matches the predetermined number of target quantization patterns, it belongs to each residual cluster for constructing the quantization pattern codebook. And extracting each quantized pattern (S26). The method of claim 2 wherein the total of quantization bits can increase two quantization pattern of the calculation step (S90) of a quantization pattern P ^t (n) and ^{P t (m), (n} ≠ m, n <m) P t (n and quantizing bit increase rate IB ^t (n, m) when merged with &quot; m &quot; Where BC ^t (n, m) represents the total number of quantization bits, Is a quantization pattern representing elements belonging to one cluster.)