KR100709376B1 - Searching apparatus for fast vector quantization encoding and method thereof - Google Patents

Abstract

Disclosed are a search apparatus and a method for fast vector quantum coding. The preprocessing unit generates the code words of the code table for each layer in the order of ascending order of L ₂ -norm. Minimum value determining unit input vector and L ₂ of the candidate vectors present in the VLC table-by calculating the average closest L ₂ - determining the initial optimal codeword having an average, and an initial minimum value of the distance eel input vector and the initial optimum code Decide on In the code table performed by the calculator, the L ₂ -average of the selected candidate vectors is performed according to a predetermined search order. The control unit compares the calculation result of the operation unit with the initial minimum value and outputs a control signal for removing the candidate vector from the code table for each layer in the first and second removal loop units according to whether the vertical and horizontal exclusion conditions are satisfied. The decision unit determines the optimum matching codeword according to the optimal condition. By creating a code table in which the size of the codewords is arranged in ascending order for each layer through the preprocessing process, and eliminating the codewords for the same layer and the lower layer in the process of determining the best match codeword Therefore, the computation time and the time required for searching for the best match code can be shortened.

Code Book, L2-Means, Match, Vector, Quantization, Preprocessing

Description

Searching apparatus for fast vector quantization encoding and method

을 도시한 도면,1 is an image vector of size 2 ^L × 2 ^L represented by L ₂ -norm pyramid data structure.

Drawing,

FIG. 2 is a block diagram illustrating an embodiment of an L ₂ -norm pyramid based search apparatus for fast vector quantum encoding according to the present invention; FIG.

3A to 3C are flowcharts of an embodiment of an L ₂ -norm pyramid based search method for fast vector quantum encoding according to the present invention;

FIG. 4 is a diagram illustrating a result of sorting codewords of a code book and a table for each layer in ascending order of L ₂ -norm values, and FIG.

5 is a diagram illustrating a search order for efficiently searching for candidate vectors in a code table.

The present invention relates to a search apparatus and a method for fast vector quantum coding. More particularly, the present invention relates to a fast vector quantum coding capable of searching for an optimal approximation codeword at high speed in a vector quantum coding based on an L ₂ -norm pyramid. It relates to a L ₂ -norm pyramid based navigation apparatus and method.

로의 사상 함수 Q로 정의될 수 있다. 즉, Q: R^K→

로 정의된다. 여기서

={

; i=1, 2, …, N}는 부호책이라 불리우는 복원 벡터들의 집합이며, N은 부호책의 크기를 의미한다.Vector quantization (VQ) shows good performance in terms of bit rate-distortion, and thus can effectively compress the signal. Therefore, it has been applied to many coding fields such as voice and image coding for a long time. VQ coding of images is a finite vector set of ^K -dimensional Euclidean space R ^K to R ^K.

It can be defined as the mapping function Q of. That is, Q: R ^K →

Is defined as here

= {

; i = 1, 2,... , N} is a set of reconstruction vectors called a code book, and N is the size of the code book.

=(x₁, x₂, …, x_k)는 부호책의 모든 부호어(codword)들과 비교된 후, Q(

)=

=(

,

, …,

)의 관계에 따라 가장 근사한 부호어로 양자화된다. Conventional vector quantization coding divides an input image into several non-overlapping blocks (hereinafter, "vectors") prior to encoding. Next, each K-dimensional input vector

= (x ₁ , x ₂ ,…, x _k ) is compared with all the codewords of the code book, and then Q (

) =

= (

,

,… ,

Is quantized as the closest codeword.

와

사이의 거리는 다음 식에 의해 구한다. When using the squared Euclicean distance as the distance scale

Wow

The distance between them is obtained by the following equation.

In addition, the mapping function Q is defined as the mapping from x to yi that satisfies the following condition.

Therefore, compression is performed by transmitting the index values of the optimum matching codewords obtained by the above Equations 1 and 2 to the decoder instead of the vector values. At this time, the decoder has the same code book as the encoder, and decoding is simply performed through a table searching process using the transmitted index value. The simplicity of such decoding is the biggest advantage of VQ.

Meanwhile, in order to find an optimal match codeword in the conventional vector quantization coding process, the VQ coders generally use a full search algorithm (FSA). The FSA requires a huge amount of computation because it is a method of manually calculating and comparing Euclidean distances between all the codewords in the input vector and the code book. For example, when the average number of operations per vector is used as a measure of coding complexity, the computational complexity of the FSA is KN. Therefore, the amount of coding operations increases with the vector dimension and the code book size, which makes it difficult to apply VQ to a real-time application system.

In order to solve the above problem, several fast searching techniques have been proposed to ensure the same coding performance as FSA. One of them is the Fast Search Algorithms for Vector Quantization of Images Using Multiple Triangles, published in the IEEE Image Processing Division (March 2000). Inequalities and Wavelet Transform) "(prior art 1). In the multi-triangular inequality-based search method proposed in the prior art 1, since the intersection of the search areas is determined as the final search area, the search area itself can be reduced, thereby reducing the calculation amount.

The other is "Fast VQ Encoding by an Efficient Kick-Out Condition," published in the Journal of Circuits and Systems for IEEE Video Technology (February 2000). 2). Prior art 2 proposes a new distance measure that has the same performance with less computation as compared to the Euclidean distance measure, and uses the Cauchy-Schwarz inequality to find codewords that minimize the proposed trading measure. Unreliable candidate codewords are eliminated in advance by the equation derived based on.

However, both prior art 1 and prior art 2 do not show satisfactory improvement in processing speed. In case of the prior art 1, unnecessary inequalities are used to remove unnecessary operations in advance, but there is a limit of speed improvement because the correlation between the inequalities is small. On the other hand, in the case of the prior art 2, there is a limit in speed improvement because it removes the unreliable operation through a single comparison process.

An object of the present invention is to provide an apparatus and method for searching for an optimal approximation codeword at high speed based on the L ₂ -norm pyramid of code words in order to improve the VQ coding speed.

와 최근접한 L₂-평균을 갖는 초기최적부호어를 결정하여 최소값으로 설정하는 최소값설정부; 상기 최소값을 저장하는 제2저장부; 일정한 검색순서에 따라 상기 부호테이블로부터 후보벡터

를 선정하여 상기 입력벡터

와 [수학식 7]에 따른 거리값을 계산하는 연산부; 상기 후보벡터

가 위치하는 계층 내지 최하위계층에 대한 각각의 상기 부호테이블로부터 일정한 제거조건에 해당하는 상기 부호어를 검색대상에서 제거하는 동작을 반복적으로 수행하는 제1제거루프부; 상기 제2저장부에 저장되어 있는 상기 최소값을 입력벡터

에 대한 최적정합부호어로 결정하여 출력하는 결정부; 및 상기 거리값이 [수학식 8]의 조건을 만족하면 상기 제1제거루프부로 제거동작을 수행하도록 하는 제어신호를 출력하고, 상기 최상위계층의 상기 부호테이블에서 모든 상기 부호어가 제거되면 상기 결정부로 최 적정합부호어의 출력동작을 수행하도록 하는 제어신호를 출력하는 제어부;를 포함한다.L ₂ -norm of the pyramid based on the navigation device for the high-speed technical problem both vector according to the present invention, for achieving the coding, the codeword L ₂ - for each layer to lift car ordered by the average (norm) A preprocessor for generating a code table for the apparatus; A first storage unit which stores the code table generated by the preprocessor; An input vector from the code table for the highest layer

A minimum value setting unit for determining an initial optimum codeword having a nearest L ₂ -average value and setting the minimum optimal value to a minimum value; A second storage unit for storing the minimum value; Candidate vectors from the code table in a predetermined search order

Select the input vector

And a calculation unit for calculating a distance value according to [Equation 7]; The candidate vector

A first removal loop unit repeatedly performing an operation of repeatedly removing the codeword corresponding to a predetermined removal condition from the target table from each of the code tables for the hierarchical layer or the lowest hierarchical layer in which a code is located; Input the minimum value stored in the second storage unit

A decision unit for determining and outputting an optimal matching coder for the output; And outputting a control signal for performing a removal operation to the first removal loop unit if the distance value satisfies the expression (8), and if all the codewords are removed from the code table of the uppermost layer. And a control unit for outputting a control signal for performing an output operation of the most suitable matching codeword.

가 존재하는 상기 부호테이블에 대한 하위계층의 부호테이블로부터 후보벡터

를 선정하여 상기 연산부에서 계산된 상기 [수학식 7]에 의한 계산값을 상기 거리값으로 대치하는 동작, 및 상기 후보벡터

를 각각의 상기 부호테이블로부터 제거하는 동작을 최하위계층에 도달할 때까지 반복적으로 수행하는 제2제거루프부;를 더 포함하며, 상기 제어부는, 상기 거리값이 [수학식 8]의 조건을 만족하지 않으면, 상기 제2제거루프부로 갱신, 대치, 및 제거동작을 수행하도록 하는 제어신호를 출력하며, 상기 제2제거루프부의 동작수행중에 상기 [수학식 8]의 조건을 만족하는 계층에 도달하면 상기 제2제거루프부로 동작수행을 중지하도록 하는 제어신호 및 상기 제1제거루프부로 제거동작을 수행하도록 하는 제어신호를 출력한다.Preferably, taking the distance value as the minimum value and updating the minimum value, the candidate vector

A candidate vector from a code table of a lower layer with respect to the code table

Selecting and substituting the calculated value according to [Equation 7] calculated by the calculating unit with the distance value, and the candidate vector.

And a second removal loop unit which repeatedly performs an operation of removing from each of the code tables until the lowest layer is reached. The control unit further comprises: the distance value satisfies the condition of Equation (8). Otherwise, a control signal is output to the second removal loop unit to perform an update, replacement, and removal operation, and when a layer that satisfies the condition of Equation 8 is reached during the operation of the second removal loop unit. A control signal outputs a control signal for stopping the operation of the second removal loop unit and a control signal for performing the removal operation of the first removal loop unit.

∥≥∥

∥이며, 상기 제거조건을 만족하면, 최하위계층에 도달할 때까지 각각의 계층에 대한 상기 부호테이블로부터 상기 후보벡터

보다 큰 L₂-평균(norm)을 같는 부호어들을 검색대상에서 제거하고, 상기 제거조건을 만족하지 않으면, 최하위계층에 도달할 때까지 각각의 계층에 대 한 상기 부호테이블로부터 상기 후보벡터

보다 작은 L₂-평균(norm)을 같는 부호어들을 검색대상에서 제거한다.Preferably, the L ₂ -norm pyramid based search apparatus for fast vector quantum coding according to the present invention, wherein the removal condition is

∥≥∥

?, If the elimination condition is satisfied, the candidate vector from the code table for each layer until the lowest layer is reached.

If the codewords having a larger L ₂ -norm is equal to the target to be removed, and the elimination condition is not satisfied, the candidate vector from the code table for each layer is reached until the lowest layer is reached.

Codewords with the same L ₂ -norm are removed from the search.

의 크기를 계산하여 상기 연산부로 제공하는 크기계산부를 갖는다.Preferably, the L ₂ -norm pyramid based search apparatus for fast vector quantum coding according to the present invention includes the input vector.

It has a size calculation unit for calculating the size of and provide the calculation unit.

Preferably, the preprocessor records the same codeword in the same table in the code table for each layer.

Preferably, the search order searches alternately before and after the initial optimal code.

와 최근접한 L₂-평균을 갖는 부호어이다.Preferably, the initial optimal code is an input vector retrieved through a binary search for the code table

Is the codeword with the nearest L ₂ -average.

와 상기 초기최적부호어와의 거리는 [수학식 7]에 의해 계산된다.Preferably, the input vector

The distance between the initial optimal code and the equation is calculated by Equation 7.

와 최근접한 L₂-평균을 갖는 초기최적부호어를 결정하고 상기 입력벡터

와 상기 초기최적부호어와의 거리를 최소값으로 설정하는 단계; (c) 일정한 검색순서에 기초하여 상기 최상위계층의 상기 부호테이블에서 후보벡터

를 선택하는 단계; (d) 선택된 상기 후보벡터

에 대한 [수학식 7]에 의해 계산된 거리값이 상기 최소값에 대해 [수학식 8]을 만족하는지 여부를 판단하는 단계; (e) 상기 (d)단계에서 상기 거리값이 상기 최소값에 대해 상기 [수학식 8]을 만족하면 모든 하위계층에 대한 상기 부호테이블로부터 일정한 제거조건에 해당하는 상기 부호어를 검색대상에서 제거하는 단계; 및 (f) 상기 최하위계층에 제거되지 않은 상기 부호어가 존재하면 상기 (c)단계로 진행하고, 상기 최하위계층에 제거되지 않은 상기 부호어가 존재하지 않으면 현재의 상기 최소값을 최적정합부호어로 결정하는 단계;를 포함한다.On the other hand, L ₂ -norm of the pyramid-based search method for high-speed vector encoding both according to the invention, the (a) codeword L ₂ - sorted in ascending order of the average difference (norm) of the code table for each layer Generating; (b) input vector

Determine an initial optimal code with L ₂ -means closest to

Setting a distance between the initial optimal code and a minimum value; (c) candidate vectors in the code table of the highest layer based on a predetermined search order;

Selecting a; (d) the candidate vector selected

Determining whether the distance value calculated by Equation 7 for Equation 7 satisfies Equation 8 with respect to the minimum value; (e) in step (d), if the distance value satisfies Equation 8 with respect to the minimum value, removing the codeword corresponding to a predetermined elimination condition from the code table for all lower layers from the search target. step; And (f) if the codeword not removed in the lowest layer exists, proceeding to step (c), and if the codeword not removed in the lowest layer does not exist, determining the current minimum value as the best match code. It includes;

가 속하는 계층이 최하위계층인가를 판단하는 단계;를 포함하며, 상기 (d2)단계에서 최하위계층으로 판단되면 상기 (f)단계로 진행한다.If it is determined in step (d) that the distance value does not satisfy the expression (8) with respect to the minimum value, (d1) taking the distance value as the minimum value and updating the minimum value; And (d2) the candidate vector.

And determining whether the layer to which the layer belongs is the lowest layer. If it is determined as the lowest layer in step (d2), the process proceeds to step (f).

를 선택하고 상기 (d)단계로 진행한다.If it is determined in the step (d2) that it is not the lowest layer, the candidate vector in the code table for the next lower layer

Select and proceed to step (d).

∥≥∥

∥이며, 상기 제거조건을 만족하면, 최하위계층에 도달할 때까지 각각의 계층에 대한 상기 부호테이블로부터 상기 후보벡터

보다 큰 L₂-평균(norm)을 같는 부호어들을 검색대상에서 제거하고, 상기 제거조건을 만족하지 않으면, 상기 최하위계층에 도달할 때까지 각각의 계층에 대한 상기 부호테이블로부터 상기 후보벡터

보다 작은 L₂-평균(norm)을 같는 부호어들을 검색대상에서 제거한다.Preferably, the removal condition is

∥≥∥

?, If the elimination condition is satisfied, the candidate vector from the code table for each layer until the lowest layer is reached.

If the codewords having the same L ₂ -norm are larger than the search target, and the elimination condition is not satisfied, the candidate vector from the code table for each layer is reached until the lowest layer is reached.

Codewords with the same L ₂ -norm are removed from the search.

Preferably, in step (a), the same codeword is recorded at the same position in the code table for each layer.

와 최근접한 L₂-평균을 갖는 부호어를 초기최적부호어로 선택하며, 입력벡터

와 초기최적부호어와의 거리는 [수학식 7]에 의해 계산된다.Preferably, the input vector searched through the binary search for the code table in the step (b)

Codeword with the nearest L ₂ -average as the initial optimal code and selects the input vector.

And the distance between the initial optimal code and Equation 7 are calculated by Equation 7.

Preferably, in the step (c), the search order alternately searches the front and rear of the initial optimal codeword.

Through the preprocessing process, the code table is arranged in ascending order of codewords for each layer, and the same and lower layers are simultaneously searched in the process of determining the best match codeword. It can shorten the time required for.

을 도시한 것이다. 도 1에서 2^L×2^L 크기의 영상벡터

의 피라미드는 {

, …,

,

,

, …,

}의 행렬 집합으로 정 의된다. 여기서

은 2 ^{l -1}×2 ^{l -1}의 크기를 가지며

의 해상도를 1/4로 줄인 영상에 해당한다. 결국

는 한 화소에 해당한다. In the L ₂ -norm pyramid data structure, an image is divided into layers, and each layer image has a lower resolution than the original image. 1 is an image vector of size 2 ^L × 2 ^L represented by L ₂ -norm pyramid data structure.

It is shown. 1, 2 ^L × 2 ^L size image vector

The pyramid of {

,… ,

,

,

,… ,

} Is defined as a set of matrices. here

^Has a size of 2 ^{l -1} × 2 ^{l -1}

Corresponds to the image whose resolution is reduced to 1/4. finally

Corresponds to one pixel.

The pyramid data structure is constructed in such a way that each pixel of the current layer is obtained by performing an operation on adjacent 2x2 pixel blocks of the upper layer. That is, one pixel of a layer l ^{-1 u l -1 (m, n} ) is the 2 × 2 adjacent pixels corresponding in the layer ^{l u l (2m-1,} 2n-1), u l (2m-1, 2n ), u ^l (2m, 2n-1), u ^l (2m, 2n). This process is expressed as the following equation.

u ^{l -1} (m, n) = O (u ^l (2m-1, 2n-1), u ^l (2m-1, 2n), u ^l (2m, 2n-1), u ^l (2m, 2n ))

In Equation 3, O () denotes an operation function, and there are many kinds of image pyramids according to the function O. The simplest pyramid structure is the L ₁ -norm pyramid structure, which is obtained by taking the average of adjacent 2 × 2 pixels. In addition, an L ₂ -norm pyramid data structure can be obtained, and the present invention implements a fast search method using the L ₂ -norm pyramid.

In the L ₂ -norm pyramid data structure, the lowest layer is the original image itself, and the highest layer corresponds to the L ₂ -norm of the entire image. The L ₂ -norm pixel u ^k-1 (m, n) in the layer l ^-1 is obtained by the following equation.

Hereinafter, an L ₂ -norm pyramid based fast search method for fast vector quantum coding according to the present invention will be described in detail.

First, a measurer is defined in the L ₂ -norm pyramid based fast search method for fast vector quantum coding according to the present invention.

의 최적정합부호어

을 찾기 위해서는 부호책 내의 모든 부호어들과 입력벡터

를 비교하여야 한다. 이러한 과정은

=Q(

)로 표현될 수 있다. 대부분의 벡터 양자화 기반의 부호기가 제곱 유클리디안 거리를 양자화를 위한 측정자로 사용하기 때문에 본 발명에서도 제곱 유클리디안 거리 즉, L₂-norm 거리를 기본적인 측정자로 한다. 이 때, 상기 [수학식 2]는 다음과 같이 재정리된다.Input vector

Best Match Code

To find all the codewords and input vectors in a codebook,

Should be compared. This process

= Q (

Can be expressed as Since most vector quantization-based coders use the square Euclidean distance as a measurer for quantization, in the present invention, the square Euclidean distance, that is, L ₂ -norm distance, is used as a basic measurer. At this time, Equation 2 is rearranged as follows.

The ultimate goal of vector quantization is to find a codeword that minimizes Equation 5 above. However, a time-saving way to achieve the same result is to find a codeword that satisfies the following conditions.                     

,

)는 아래와 같다.Where d '(

,

) Is shown below.

∥²는 모든 부호어들에 대해 공통항이기 때문에 제거해도 무방하다. 따라서, {∥

∥| 1≤i≤N }이 전처리 과정을 통해 미리 알려져 있다면 상기 [수학식 6]을 계산하는 것이 상기 [수학식 1]을 계산하는 것보다 빠르다. 따라서 본 발명에 따른 고속 탐색 방법은 상기 [수학식 1]의 유클리디안 거리를 최소화하는 것이 아니라, 상기 [수학식 6]의 d' 거리를 최소화하는 데 초점을 맞춘다.On the other hand, ∥in Equation 6 above

Since ² is a common term for all codewords, it can be removed. Thus, {∥

∥ | If 1≤i≤N} is known in advance through the pretreatment process, calculating [Equation 6] is faster than calculating [Equation 1]. Therefore, the fast search method according to the present invention focuses on minimizing the d 'distance of Equation 6, not minimizing the Euclidean distance of Equation 1.

Assuming that some of the codewords in the code book have been examined so far, the minimum distance to date is defined as d _min . The following equation can be derived by the above Equation 6 and the Kosch-Schwartz inequality.

가 아래의 조건을 만족하면, 항상 d'(

,

)≥d_min 이므로

를 후보에서 제거할 수 있다. 즉,

는 현재까지의 최적근사부호어보다 입력벡터

에 더 근사할 수 없기 때문에 더 이상의 거리계산없이 제거될 수 있다. So codeword

Is always satisfied by d '(

,

) ≥d _min

Can be removed from the candidate. In other words,

Is the input vector rather than the best approximation

Since it is no closer to, it can be removed without further distance calculation.

∥| 1≤i≤N }이 전처리 과정을 통해 미리 계산되어 있으므로 입력벡터

에 대한 상기 조건의 계산은 매우 간단하다.Also {∥

∥ | 1≤i≤N} is precomputed through preprocessing, so the input vector

The calculation of the above conditions is very simple.

와 부호어

의 계층 l에서의 벡터들을 각각

과

로 표현하기로 한다. 이 때,

=∥

∥이다. 그러면 코쉬-쉬바르쯔 부등식으로부터 다음 부등식을 유도할 수 있다.Now, Equation 7 is extended to the general case. The vector dimension K is 2L × 2L, and the input vector is

And codeword

Each of the vectors in layer l of

and

It is expressed as. At this time,

= ∥

It is. Then we can derive the next inequality from the Kosch-Schwartz inequality.

은

의 k번째 원소를 가리키며, 계층 l+1의 바로 인접한 2×2 화소들의 L₂-norm에 해당한다. 상기 [수학식 9]로부터 상기 [수학식 7]은 다음과 같이 일반화될 수 있다.In Equation 9 above

silver

Points to the k-th element of, and corresponds to L ₂ -norm of immediately adjacent 2 × 2 pixels of the layer l + 1. Equation 7 from Equation 9 may be generalized as follows.

이며,

이다.In Equation 10,

Is,

to be.

The process of demonstrating [Equation 10] is as follows.

,

의 내적은,Two vectors in layer l

,

Of the inner product,

이다.

to be.

이므로 계층 l-1에서 두 벡터의 내적은,On the other hand, from a one-dimensional perspective

Since the dot product of the two vectors in layer l -1,

to be. Therefore, from the dot product and the Kosch-Schwartz inequality of the two vectors at layer l and layer l- 1,

을 의미한다. 따라서 이를 일반화하면 다음과 같다.In other words,

Means. Therefore, generalizing it is as follows.

2 is a block diagram showing the configuration of an embodiment of the L ₂ -norm pyramid based fast search apparatus 20 for fast vector quantum coding according to the present invention.

Referring to FIG. 2, the L ₂ -average pyramid-based search apparatus 20 for fast vector quantum encoding according to the present invention includes a preprocessor 21, a code table storage 22, and a minimum value setter 24. And a minimum value storage section 25, a calculation section 23, a first removal loop section 28, a second removal loop section 29, a determination section 27, and a control section 26.

The preprocessor 21 generates code tables for each layer by sorting the code words in ascending order of L ₂ -norm. The code table for each layer generated by the preprocessor 21 is stored in the code table storage 22. On the other hand, the preprocessor 21 records the same codeword in the same position in the code table for each layer.

와 최근접한 L₂-평균을 갖는 초기최적부호어를 결정하여 최소값으로 설정한다. 최소값설정부(24)에서 설정된 최소값은 최소값저장부(25)에 저장된다. 최소값설정부(24)에서 결정된 초기최적부호어는 부호테이블에 대한 이진탐색을 통해 검색된 부호어이며, 입력벡터

와 초기최적부호어와의 거리는 [수학식 7]에 의해 계산된다.The minimum value setting section 24 inputs the input vector from the sign table for the highest layer.

Determine the initial optimal code with L ₂ -average and set it as the minimum value. The minimum value set by the minimum value setting unit 24 is stored in the minimum value storage unit 25. The initial optimal codeword determined by the minimum value setting unit 24 is a codeword searched through binary search for a code table, and an input vector.

And the distance between the initial optimal code and Equation 7 are calculated by Equation 7.

를 선정하여 입력벡터

와 [수학식 7]에 따른 거리값을 계산한다. 연산부(23)는 초기최적부호어의 전후방을 번갈아가며 검색하여 부호테이블로부터 후보벡터

를 선정한다.The calculating section 23 uses the candidate table from the code table in a predetermined search order.

Select the input vector

Calculate the distance value according to [Equation 7]. The calculation unit 23 alternately searches the front and rear of the initial optimal codeword to search for candidate vectors from the code table.

Select.

가 위치하는 계층 내지 최하위계층에 대한 각각의 부호테이블로부터 일정한 제거조건에 해당하는 부호어를 검색대상에서 제거하는 동작을 반복적으로 수행한다. 제거조건은 ∥

∥≥∥

∥로 설정되며, 제1제거루프부(28)는 연산부(23)에서 계산한 거리값이 제거조건을 만족하면, 최하위계층에 도달할 때까지 각각의 계층에 대한 부호테이블로부터 후보벡터

보다 큰 L₂-평균(norm)을 같는 부호어들을 검색대상에서 제거하고, 제거조건을 만족하지 않으면, 최하위계층에 도달할 때까지 각각의 계층에 대한 부호테이블로부터 상기 후보벡터

보다 작은 L₂-평균(norm)을 같는 부호어들을 검색대상에서 제거한다.The first removal loop unit 28 is a candidate vector

An operation of repeatedly removing a code word corresponding to a predetermined elimination condition from a target to be searched is performed from each code table for the hierarchical layer or the lowest hierarchical layer. Removal condition

∥≥∥

If the distance value calculated by the calculation unit 23 satisfies the elimination condition, the first elimination loop unit 28 is a candidate vector from the code table for each layer until it reaches the lowest layer.

If the codewords having the same L ₂ -norm are removed from the search target, and the elimination condition is not satisfied, the candidate vector from the code table for each layer is reached until the lowest layer is reached.

Codewords with the same L ₂ -norm are removed from the search.

가 존재하는 부호테이블에 대한 하위계층의 부호테이블로부터 후보벡터

를 선정하여 연산부(23)에서 계산된 [수학식 7]에 의한 계산값을 거리값으로 대치하는 동작 및 후보벡터

를 각각의 부호테이블로부터 제거하는 동작을 최하위계층에 도달할 때까지 반복적으로 수행한다.The second removal loop unit 29 takes the distance value as the minimum value and updates the minimum value. The candidate vector

A candidate vector from a code table of a lower layer with respect to a code table

To replace the calculated value by [Equation 7] calculated by the calculation unit 23 by the distance value and the candidate vector

The operation to remove from each code table is repeatedly performed until the lowest layer is reached.

의 크기를 계산하여 연산부(23)로 제공하는 크기계산부(도면에는 도시되어 있지 않음)를 구비하는 것이 바람직하다. 또한 상기 크기계산부가 후보벡터

의 크기를 계산하여 연산부(23)에 제공하도록 할 수도 있다.On the other hand, the L ₂ -average pyramid based search apparatus 20 for fast vector quantum coding according to the present invention is input vector

It is preferable to have a size calculation unit (not shown in the figure) for calculating the size of and providing it to the calculation unit 23. In addition, the size calculating unit is a candidate vector

Calculate the size of may be provided to the calculation unit 23.

에 대한 최적정합부호어로 결정하여 출력한다.The determination unit 27 inputs the minimum value stored in the minimum value storage unit 25 according to the control signal of the control unit 26.

Determine and output the best matching code for.

The control unit outputs a control signal for performing the removal operation to the first removal loop unit 28 when the distance value calculated by the calculation unit 23 satisfies the condition of [Equation 8], and in the code table of the uppermost layer. When all codewords are removed, the control unit 27 outputs a control signal for performing an output operation of the best match coder to the decision unit 27.

On the other hand, if the distance value calculated by the calculation unit 23 does not satisfy the condition of [Equation 8], the control unit 26 updates the minimum value to the second removal loop unit 29, replaces the distance value, and replaces the codeword. Outputs a control signal to perform the removal operation, and if the layer that satisfies the condition of Equation 8 is reached during the operation of the second removal loop unit 29, the operation is performed by the second removal loop unit 29. The control signal for stopping and the control signal for performing the removal operation to the first removal loop unit 28 are output.

3A to 3C are flowcharts of an embodiment of a L ₂ -norm pyramid based fast search method for fast vector quantum coding according to the present invention.

계층에 대한 부호어들을 L₂-norm 올림차 순으로 정렬하여 최상위 계층인

계층에 대한 부호테이블을 생성한다(S200). 다음으로

에 기반하여

내지

계층에 대한 부호어들을 L₂-norm 올림차 순으로 정렬하여 각 계층에 대한 부호테이블을 생성한다(S202). 상기 S200 및 S202단계에서 생성된 부호테이블들은 부호테이블저장부(22)에 저장된다. 2 and 3A to 3C, the preprocessor 21 sorts the L ₂ -norm values of all code words in the code book in ascending order before encoding the input signal to generate a code table for each layer. Create The preprocessor 21 is a space vector

Sort the codewords for the layer in L ₂ -norm ascending order

A code table for the layer is generated (S200). to the next

Based on

To

The codewords for the layers are arranged in order of L ₂ -norm ascending order (S202). The code tables generated in steps S200 and S202 are stored in the code table storage unit 22.

계층을 포함하여 각 계층의 부호어들은 ∥

∥≤∥

∥≤…≤∥

∥ 순으로 나열된다. 이러한 전처리단계는 입력벡터

가 입력되기 전에 이루어지므로 이를 오프라인 전처리라 한다. 도 4는 부호책 및 각 계층에 대한 테이블의 부호어들을 L₂-norm 값들을 올림차 순으로 정렬한 결과를 나타내는 도면이다. 도 3에서 부호책은 원 영상을 나타내는

계층에 대한 것이며,

내지

계층은 각각의 하위 계층으로부터 2×2 픽셀 값을 순차적으로 연산하여 얻어진다. According to the result of the alignment process performed in the preprocessor 21

The codewords of each layer, including the layer

∥≤∥

∥≤… ≤∥

Listed in order. This preprocessing step is an input vector.

This is called offline preprocessing, because it is done before is entered. FIG. 4 is a diagram illustrating a result of sorting codewords of a code book and a table for each layer in ascending order of L ₂ -norm values. In Fig. 3, the code book represents the original image.

For tiers,

To

The layer is obtained by sequentially calculating 2x2 pixel values from each lower layer.

가 입력되면, 입력벡터

에 대해 ∥

∥ 및

내지

에 대한 계산을 수행한다(S204). 입력벡터

에 대한 크기계산은 연산부(23)에서 수행되며, 별도의 연산수단(도면에는 도시되어 있지 않음)을 통해 수행되어 계산된 크기값이 연산부(23)로 공급될 수도 있다. 최소값설정부(24)는 입력벡터

에 가장 가까운 L₂-norm 을 갖는 초기최적부호어

을 찾는다(S206). 최소값설정부(24)는 초기최적부호어

이 결정되면 입력벡터

와의 거리를 계산하여 초기최소값 d_min=d'(

,

)로 설정한다(S208). 초기최소값은 최소값저장부(25)에 저장되며 이후 초기최소값보다 작은 거리값이 존재하면 해당 거리값으로 대치된다. 초기최적부호어를 결정하는 과정에서, 계산량을 줄이기 위해

가 올림차 순으로 정렬되어 있는 점을 이용한다. 따라서 이진 탐색을 통해 초기최적부호어를 찾을 수 있으며, 이 경우 초기최적부호어를 log₂N 단계만에 찾을 수 있다. vector

If is input, input vector

About

∥ and

To

Perform the calculation for (S204). Input vector

The size calculation for is performed in the calculator 23, and the calculated size value may be supplied to the calculator 23 through a separate calculator (not shown). The minimum value setting section 24 is an input vector

Initial optimal codeword with L ₂ -norm nearest

To find (S206). Minimum value setting section 24 is the initial optimal code

Once this is determined, the input vector

Calculate the distance from and the initial minimum value d _min = d '(

,

(S208). The initial minimum value is stored in the minimum value storage unit 25, and if there is a distance value smaller than the initial minimum value, it is replaced with the corresponding distance value. In the process of determining the initial optimal code,

Use points that are arranged in ascending order. Therefore, the initial optimal codeword can be found through binary search. In this case, the initial optimal codeword can be found in log ₂ N steps.

={

|

∈

,

≠

}이라 할 때, 제어부(26)는 벡터집합

에 하나 이상의 원소가 존재하는가를 확인한다(S210). 벡터집합

이 공집합이면 제어부(26)는 d_min에 대응하는 부호어를 최적정합부호어

으로 선택하도록 하는 제어신호를 결정부(27)로 출력하며 결정부는 최소값저장부(25)에 저장되어 있는 현재의 최소값을 최적정합부호어로 결정하여 출력한다(S212). 벡터집합

이 공집합이 아닌 경우에는 벡터집합

의 원소로 규정되는 부호테이블 내의 부호어들을 도 4에 도시된 바와 같이

에서 출발하여 전후방을 번갈아가며 탐색된다. Vector set of codewords in the code table for the layer to search

= {

|

∈

,

≠

}, The control unit 26 is a vector set.

It is checked whether at least one element exists in S210. Vector set

If this empty set, the control unit 26 selects the codeword corresponding to d _min as the best match codeword.

The control unit outputs a control signal for selection to the determination unit 27, and the determination unit determines and outputs the current minimum value stored in the minimum value storage unit 25 as an optimum matching coder (S212). Vector set

Vector set if not empty

As shown in FIG. 4, codewords in a code table defined by elements of

Searches alternately back and forth starting at.

이

보다

에 더 근사한지를 알기 위해, L₂-norm 피라미드의 최상위 계층의

에 대해 d^'0(

,

)을 계산한다(S214). 제어부(26)는 d^'0(

,

)의 계산결과를 상기 S208단계에서 초기최소값으로 설정한 d_min과 비교한다(S216). The calculating unit 23 selects candidate vectors selected by the search order.

this

see

To see if it's more approximate to the top level of the L ₂ -norm pyramid,

For d ^'0 (

,

) Is calculated (S214). The control unit 26 is d ^'0 (

,

) Is compared with d _min set as the initial minimum value in step S208 (S216).

,

)≥d_min이면, 제어부(26)는 제1제거루프부(28)로 제거동작을 수행하도록 하는 제어신호를 출력한다. 제1제거루프(28)는 후보벡터의 크기 ∥

∥와 입력벡터의 크기 ∥

∥를 비교한다(S220). 제1제거루프부(28)는 후보벡터의 크기 ∥

∥이 입력벡터의 크기 ∥

∥보다 크면 최상위계층에서

보다 큰 모든 후보벡터

를 제거하고(S222), 작으면 최상위계층에서

보다 작은 모든 후보벡터

를 제거한다(S224). The comparison result in step S216 is d ^'0 (

,

If ≥ d _min , the control unit 26 outputs a control signal for performing the removal operation to the first removal loop unit 28. The first removal loop 28 is the size of the candidate vector ∥

∥ size of input vector

∥ compare (S220). The first removal loop portion 28 is the size of the candidate vector ∥

∥ size of this input vector

If it is bigger than ∥

All candidates greater than

Remove (S222), if the smallest at the top layer

All candidate vectors smaller than

Remove (S224).

,

)≥d_min이면 아래의 [수학식 11]에 의해 d^'L(

,

)≥d_min이며, 따라서 모든 하위계층에서

은 최적정합부호어가 될 수 없음을 알 수 있다. When the removal operation for the top layer is completed, the flow proceeds to step A to perform the removal operation for each lower layer. d ^'0 (

,

) If d _min, then d ^'L (

,

) ≥d _min , so in all sublayers

It can be seen that is not an optimal matching codeword.

d'(,)≡d^'L(, )≥d^'L-1(, )≥Λ≥d^'0(, )

d '(,) ≡d ^{' L} (,) ≥d ^{' L-1} (,) ≥Λ≥d ^{' 0} (,)

In [Equation 11],

이며,

Is,

이다.

to be.

∥와 ∥

∥를 비교하여(S226). ∥

∥이 ∥

∥보다 크면 하위계층에서

보다 큰 모든 후보벡터

를 제거하고(S228), 작으면 하위계층에서

보다 작은 모든 후보벡터

를 제거한다(S230). According to [Equation 11], the first removal loop portion 28 for the continuous lower layer of the highest layer ∥

∥ and

∥ in comparison (S226). ∥

This

Is greater than ∥

All candidate vectors greater than

Is removed (S228), and if it is small,

All candidate vectors smaller than

Remove (S230).

에 대한 제거과정이 수행되었는가를 확인하고(S232) 최하위계층에 도달하지 않은 것으로 확인되면 다음 하위계층에 대한 부호테이블로 이동하여(S233) 해당 하위계층에 대해 상기 S226단계 내지 상기 S232단계를 반복적으로 수행한다. 만약 최하위계층에 도달한 것으로 확인되면 상기 S210단계로 진행하여 도 5에 도시된 탐색순서에 따라 다음 후보벡터에 대해 상기 S212단계 내지 상기 S232단계가 반복적으로 수행된다.The first removal loop unit 28 is a candidate vector up to the lowest hierarchy.

If it is confirmed that the removal process has been performed (S232) and it is determined that the lowest layer has not been reached, the process moves to the code table for the next lower layer (S233) and repeats steps S226 to S232 for the corresponding lower layer. To perform. If it is determined that the lowest layer is reached, the process proceeds to step S210 and steps S212 to S232 are repeatedly performed for the next candidate vector according to the search procedure shown in FIG. 5.

,

)<d_min이면, 제어부(26)는 제2제거루프부(29)로 갱신, 대치, 및 제거동작을 수행하도록 하는 제어신호를 출력한다. 제2제거루프부(29)는 최소값저장부(25)에 저장되어 있는 최소값을 d^'0(

,

)으로 갱신한 후 다음의 하위계층에 대해 d^'1(

,

)을 계산하여(S234) 계산결과를 최소값 d_min과 비교한다(S236). d^'1(

,

)의 계산결과가 d_min보다 작으면 상위계층에 대한 부호테이블에서 후보벡터

을 제거한 후 최소값저장부(25)에 저장되어 있는 최소값 d_min을 d^'1(

,

)로 변경한다(S238). 제2제거루프부(29)는 연속되는 하위계층에 대한 d^'l(

,

)의 계산결과가 d_min보다 작은 경우에 최소값의 갱신, 거리값의 대치 및 부호어의 제거 동작을 최하위계층에 도달할 때까지 반복적으로 수행한다. If d ^'0 (

,

If) <d _min , the control unit 26 outputs a control signal to the second removal loop unit 29 to perform update, replace, and remove operations. The second removal loop unit 29 returns the minimum value stored in the minimum value storage unit 25 to d' ⁰ (

,

) And then update d ^'1 (

,

) Is calculated (S234) and the calculated result is compared with the minimum value d _min (S236). d ^'1 (

,

) Is less than d _min , the candidate vector in the sign table for the higher layer

After removing the minimum value d _min stored in the minimum value storage section 25, d ^'1 (

,

(S238). The second removal loop portion 29 is d ^'l (

,

If the result of the calculation is less than d _min , the minimum value update, the distance value substitution, and the codeword removal operation are repeatedly performed until the lowest layer is reached.

,

)의 계산결과가 d_min보다 크거나 같으면 제어부(26)는 제2제거루프부(29)의 동작을 정지하도록 하는 제어신호를 출력하는 동시에 제1제거루프부(28)로 해당 계층에 대한 제거동작을 수행하도록 하는 제어신호를 출력한다. 이러한 과정이 도 3c에 에서 A단계, 즉 상기 S226단계로 진행하는 것으로 도시되어 있다. If d ^'l (for any layer before reaching the lowest layer,

,

If the calculated result of) is greater than or equal to d _min , the control unit 26 outputs a control signal for stopping the operation of the second removal loop unit 29 and simultaneously removes the corresponding layer to the first removal loop unit 28. Output a control signal to perform the operation. This process is shown as proceeding to step A, that is, step S226 in Figure 3c.

Hereinafter, a L ₂ -norm pyramid based search method for fast vector quantum coding according to the present invention is compared with the prior art. For comparison, the experimental and learning images used in the experiment are all 512 × 512 brightness images. Each image is divided into 4 × 4 blocks that do not overlap prior to encoding. Thus, each block becomes a 16-dimensional vector. That is, K = 16 and L = 2.

Well-known LBG (Linde, Buzo and Gray) vector quantization techniques are used to design code books. The Lena image is used as the learning image for the design of the code book, and the designed code book is used to encode the four experimental images Peppers, Jet, Baboon, and Lena.

The search method according to the present invention is compared with two recently developed fast techniques. One is the MTIE (Multiple Triangle Inequalities) technique (prior art 1), and the other is "High-Speed VQ Coding with Efficient Elimination Conditions," published in a paper published in the February 2000 issue of Circuits and Systems for IEEE video technology. Fast VQ Encoding by an Efficient Kick-Out Condition (prior art 2).                     

Both the searching method and the prior arts 1 and 2 according to the present invention have the same encoding performance as the global search algorithm. Table 1 shows the CPU execution time per experimental image performed for the searching method and the prior arts 1 and 2 according to the present invention. Table 1 shows the CPU execution time for four experimental images according to the code book size (N). Measurements were made on three Ns: 256, 512, and 1024. By calculating the relative time complexity for the global search algorithm, the search method according to the present invention is compared with the prior arts 1 and 2, and the comparative scale Sc is defined as follows.

Where TFSA and TCOMP represent the CPU execution time of the technique compared with the global search algorithm, respectively. In Table 1, it can be seen that the searching method according to the present invention is faster than the existing techniques for all experimental images regardless of the size of the code book. Even for the worst case Baboon image, the search method according to the present invention is about twice as fast as other fast techniques.

On the other hand, as the codebook size increases, it can be seen that the efficiency of the search method according to the present invention is improved over other techniques. For example, in Pepper and Jet images when N is 1024, the searching method according to the present invention is three times faster than the prior art 2. This tendency is a great advantage of the search method according to the present invention because the amount of computation is generally required much as the size of the code book increases.

Experimental video Coat book size algorithm Prior Art 1 (MTIE) Prior art 2 The present invention Lena N = 256 6.3% 6.4% 3.3% N = 512 4.3% 5.5% 2.1% N = 1024 3.1% 4.5% 1.5% Pepper N = 256 6.7% 7.3% 3.3% N = 512 4.9% 6.5% 2.7% N = 1024 3.9% 6.0% 1.9% Jet N = 256 5.1% 5.1% 2.6% N = 512 3.8% 4.2% 1.7% N = 1024 3.4% 4.0% 1.3% Baboon N = 256 18.7% 17.3% 8.9% N = 512 16.5% 16.2% 7.2% N = 1024 16.1% 16.1% 6.6%

Meanwhile, Table 2 below shows the calculation amount of various fast search techniques in terms of the average number of operations per pixel. This experiment was performed only for the most common codebook size N = 256. Referring to Table 2, it is shown that the searching method according to the present invention is superior to the prior arts 1 and 2. For example, for the Jet image, the computational amount of the search method according to the present invention is 58.7% of the prior art 1, and only 56.5% of the prior art 2.

Experimental video algorithm Average number of operations × ± compare Square root Sum Lena Prior Art 1 (MTIE) 12.6 26.0 5.68 0.34 44.6 Prior art 2 19.8 21.5 4.9 0.063 46.3 The present invention 9.4 12.2 4.99 0.31 26.9 Pepper Prior Art 1 (MTIE) 13.3 27.1 6.07 0.33 46.8 Prior art 2 22.1 23.9 5.28 0.063 51.3 The present invention 9.3 12.3 5.36 0.31 29.3 Jet Prior Art 1 (MTIE) 11.5 23.6 4.83 0.31 40.2 Prior art 2 16.5 17.9 4.20 0.063 38.7 The present invention 7.93 10.2 4.23 0.31 22.7 Baboon Prior Art 1 (MTIE) 43.7 86.8 14.8 0.40 145.7 Prior art 2 58.8 62.8 12.2 0.063 133.9 The present invention 32.0 37.4 13.0 0.31 82.7 FSA 256 510 16 - 782

As can be seen from Table 1 and Table 2, the search method according to the present invention can not only significantly improve the speed of VQ coding without degrading performance, but also has superior performance to existing search techniques. In the above experiments, we did not use PDE (Partial distance elimination), which is a well-known method to reduce the amount of computation when calculating distance. As PDE is used, additional performance improvement can be obtained. Therefore, the above experimental results can be regarded as upper execution time or upper computation amount.

By creating a code table in which the size of the codewords is arranged in ascending order for each layer through the preprocessing process, refer to the code table corresponding to the search target layer when searching for the candidate vector for each layer and calculating the distance from the input vector. Search time can be shortened. Also, in determining the optimal matching code, the process of eliminating candidate vectors satisfying the exclusion conditions for the same layer and the lower layer may be collectively performed, thereby reducing the amount of computation and searching time.

Although the present invention has been described in detail through the representative embodiments, those skilled in the art to which the present invention pertains can make various modifications without departing from the scope of the present invention with respect to the embodiments described above. I will understand. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined by the claims below and equivalents thereof.

Claims (17)

A preprocessor for ordering codewords in ascending order of L ₂ -norm to generate a code table for each layer; A first storage unit which stores the code table generated by the preprocessor; An input vector from the code table for the highest layer

A minimum value setting unit for determining an initial optimum codeword having a nearest L ₂ -average value and setting the minimum optimal value to a minimum value; A second storage unit for storing the minimum value; Candidate vectors from the code table in a predetermined search order

Select the input vector

And a calculation unit for calculating a distance value according to [Equation 7]; The candidate vector

A first removal loop unit repeatedly performing an operation of repeatedly removing the codeword corresponding to a predetermined removal condition from the target table from each of the code tables for the hierarchical layer or the lowest hierarchical layer in which a code is located; The minimum value stored in the second storage unit is the input vector.

A decision unit for determining and outputting an optimal matching coder for the output; And If the distance value satisfies the condition of [Equation 8], the control signal outputs a control signal to perform the removal operation to the first removal loop unit. And a control unit for outputting a control signal for performing an output operation of a matching codeword. ₂ . [Equation 7]

[Equation 8]

The method of claim 1, Updating the minimum value by taking the distance value as the minimum value, the candidate vector

A candidate vector from a code table of a lower layer with respect to the code table

Selecting and substituting the calculated value according to [Equation 7] calculated by the calculating unit with the distance value, and the candidate vector.

And a second elimination loop unit which repeatedly performs an operation of eliminating from each of the code tables until the lowest layer is reached. If the distance value does not satisfy the condition of [Equation 8], the control unit may be configured to perform a control signal for performing the minimum value update operation, the distance value replacement operation, and the candidate vector removal operation to the second removal loop unit. And a control signal for stopping the operation of the second elimination loop unit when the layer satisfying the condition of [Equation 8] is reached during the operation of the second eliminating loop unit and the eliminating operation by the first eliminating loop unit. L ₂ -average pyramid-based search apparatus for fast vector quantum coding, characterized in that for outputting a control signal to perform. The method according to claim 1 or 2, The removal condition is

∥≥∥

∥, If the elimination condition is satisfied, the candidate vector from the code table for each layer until the lowest layer is reached.

Remove the codewords with the same L ₂ -norm from the search If the elimination condition is not satisfied, the candidate vector from the code table for each layer until the lowest layer is reached.

Smaller than L ₂ - Average pyramid base of the navigation apparatus-average L ₂ for the high-speed vector encoding both, comprising a step of removing the gatneun codeword (norm) in the search. The method according to claim 1 or 2, The input vector

L ₂ -average pyramid-based search apparatus for fast vector quantum coding, characterized in that it further comprises a size calculation unit for calculating the size of the calculation to provide to the operation unit. The method according to claim 1 or 2, Average pyramid base of the navigation apparatus, wherein the pre-processing unit L ₂ for the high-speed vector encoding both, characterized in that for recording at the same position freezing same references in the code table for each layer. The method according to claim 1 or 2, The search order of the L ₂ -average pyramid for fast vector quantum coding, characterized in that the search order alternately before and after the initial optimal code. The method according to claim 1 or 2, The initial optimal codeword is an input vector searched through binary search for the code table.

And closest L ₂ - Average pyramid based on the navigation device - L ₂ for the high-speed vector, characterized in that both encoded codeword with an average. The method according to claim 1 or 2, The input vector

And a distance between the initial optimal coder and the L ₂ -average pyramid-based search apparatus for fast vector quantum coding according to Equation (7). (a) generating code tables for each layer by sorting code words in ascending order of L ₂ -norm; (b) input vector

Determine an initial optimal code with L ₂ -means closest to

Setting a distance between the initial optimal code and a minimum value; (c) candidate vectors in the code table of the highest layer based on a predetermined search order;

Selecting a; (d) the candidate vector selected

Determining whether the distance value calculated by Equation 7 for Equation 7 satisfies Equation 8 with respect to the minimum value; (e) in step (d), if the distance value satisfies Equation 8 with respect to the minimum value, removing the codeword corresponding to a predetermined elimination condition from the code table for all lower layers from the search target. step; And (f) if the codeword not removed in the lowest layer exists, proceeding to step (c); and if the codeword not removed in the lowest layer does not exist, determining the current minimum value as the best match code; L ₂ -average pyramid-based search method for fast vector quantum coding comprising a. [Equation 7]

[Equation 8]

The method of claim 9, If it is determined in step (d) that the distance value does not satisfy Equation 8 with respect to the minimum value, (d1) taking the distance value as the minimum value and updating the minimum value; And (d2) the candidate vector

Determining whether the layer to which the layer belongs is the lowest layer; If it is determined that the lowest layer in the step (d2), step (f) is characterized in that the L ₂ -average pyramid-based search method for fast vector quantum coding. The method of claim 10, If it is determined in the step (d2) that it is not the lowest layer, the candidate vector in the code table for the next lower layer

L ₂ -average pyramid-based search method for fast vector quantum coding, characterized in that it is selected and proceeds to step (d). The method according to any one of claims 9 to 11, In the step (e), the removal condition is

∥≥∥

∥, If the elimination condition is satisfied, the candidate vector from the code table for each layer until the lowest layer is reached.

Remove the codewords with the same L ₂ -norm from the search If the elimination condition is not satisfied, the candidate vector from the code table for each layer until the lowest layer is reached.

Smaller than L ₂ - Average pyramid-based methods of navigation-average L ₂ for the high-speed vector encoding both, characterized in that for removing from the search target code words having the (norm). The method according to any one of claims 9 to 11, Before the step (b), the input vector

Computing the magnitude of the L ₂ -average pyramid-based search method for fast vector quantum coding, characterized in that it further comprises. The method according to any one of claims 9 to 11, L ₂ -mean pyramid based search method for fast vector quantum coding, characterized in that the same codeword is recorded in the same position in the code table for each layer. The method according to any one of claims 9 to 11, The search order of the L ₂ -average pyramid for the fast vector quantum coding, characterized in that the search alternately before and after the initial optimal code. The method according to any one of claims 9 to 11, The initial optimal codeword is an input vector searched through binary search for the code table.

And closest L ₂ - average search method based pyramid - L ₂ for the high-speed vector, characterized in that both encoded codeword with an average. The method according to any one of claims 9 to 11, The input vector

And a distance between the initial optimal coder and the L ₂ -average pyramid-based search method for fast vector quantum coding according to Equation (7).

Images