KR900009121B1 - Special data processing system - Google Patents

Abstract

No content.

Description

Method for classifying feature data at high speed and feature data processing apparatus for realizing the method

1A and 1B are diagrams for explaining a conventional method of classification processing, in which FIG. 1A is a diagram showing a multi-member slice method, and FIG. 1B is a diagram showing a clustering method.

2 is a block diagram showing the construction of a first embodiment of an image data processing apparatus according to the present invention;

FIG. 3 illustrates in detail the storage area of the main memory shown in FIG. 2 for storing sorted image data.

4A and 4B are flowcharts for explaining the operation of the first embodiment of the image data processing apparatus according to the present invention.

5 is a diagram for explaining a process for determining a category from a reference density vector for a reference pixel.

FIG. 6 is a diagram showing space allocation data indicating the state of belonging to the category of each pixel of the input image data. FIG.

Fig. 7 is a diagram showing a state of belonging to a category of a pixel.

FIG. 8 is a diagram showing space allocation data and classification data in the belonging state illustrated in FIG.

9A to 9C are diagrams for explaining the classification of the clustering scheme according to the present invention.

10 is a diagram for explaining the operation of the second embodiment;

Fig. 11 is a block diagram showing the construction of the third embodiment.

12 is a flowchart for explaining the operation of the third embodiment.

* Explanation of symbols for main parts of the drawings

12: main memory 14: R memory

15: G memory 16: B memory

17: linear combination processor 18: space allocation data memory

19: classification processor 20: classification data memory

31: reference concentration vector storage region

The present invention relates to partitioning feature data, such as color image data, into n categories in a clustering manner, and more particularly, to a method for quickly classifying feature data and a feature data processing apparatus for realizing the method.

Classification of color image data input through an A / D converter in feature data, such as an ITV camera or a drum scanner, into a predetermined category with respect to color is a technology required in various fields. When the input color image data is reflected in a three-dimensional feature space defined by RGB, the image can be classified into a category determined by its distribution state. As a method of performing such classification, the multi-slice slice method shown in FIG. 1A and the clustering method shown in FIG. 1B have been known. 1A and 1B show the state of classification in the secondary space for convenience.

In the multiple-way slice method shown in FIG. 1A, it is common to classify color combinations by preparing a simple look up table for each component of the input image data and referring to this table. According to this method, high-speed processing is possible, but each component of the color image data cannot be handled independently. Therefore, there is a problem that the classification error depends on the input condition of the image data.

In contrast, in the clustering method illustrated in FIG. 1B, the image data displayed in the feature space can be classified by reducing the classification error. In particular, according to the method using the K averaging algorithm, better clustering results can be obtained.

In the K averaging algorithm, the image data is classified by calculating the distance to find out which density vector of each pixel constituting the color image data is closest to the n reference density vectors. However, this K-average algorithm has a problem in that high-speed processing becomes difficult because the numerical calculation amount is enormous.

(k=0∼n-1)에서 2개의 기준 특징 벡터

를 선택하는 각 조합에 있어서 조합의 2개의 기준 특징 벡터중의 어느 것이 기준 특징 벡터에 가까운가를 나타내는 소속비트 데이타 L_ij를 대상 특징 데이타를 구성하는 모든 특징 요소에 대하여 상기 조합의 전부에 걸쳐서 구하므로써, 각 특징 요소에 대한 공간 할당 데이타를 작성하는 단계를 구비하는데, 여기에서 상기 공간 할당 데이타는 각 특징 요소에 있어서의 상기 조합 전부에 대한 소속 비트 데이타로 구성되며, 대응하는 공간 할당 데이타에 의거하여, 상기 미리 지정되고 있는 n개의 기준 특징 요소에 각각 대응하는 어떤 카테고리에 특징 요소가 할당되는가를 표시하는 분류 데이타를 작성하는 단계를 구비한다.The present invention has been studied in view of the above facts, and the first object of the present invention is a feature consisting of a plurality of feature elements that can be represented by feature vectors and represented by U bits in a P (P is a positive integer) dimensional feature space. A method of classifying data at high speed is provided. The method includes a reference feature vector group corresponding to each of n reference feature elements that are predefined.

2 reference feature vectors at (k = 0 to n-1)

For each combination to select, the belonging bit data L _ij indicating which of the two reference feature vectors of the combination is close to the reference feature vector is obtained for all the feature elements constituting the target feature data over all of the combinations. Creating space allocation data for each feature element, wherein the space allocation data consists of the belonging bit data for all of the combinations in each feature element, based on the corresponding space allocation data. And generating classification data indicating which category corresponds to each of the n reference feature elements specified in advance.

(k=0∼n-1)으로부터 변수 i와 j(i＜j)에 의하여 지정되는 기준 농도 벡터중의 어느 것이 기준 특징 벡터에 가까운가를 구하고, 구해진 결과에서 변수 i와 j의 분류 데이타를 갱신하고, 변수 j가 n-1로 되었을때 분류 데이타에 의하여 당해 화소가 할당되는 카테고리를 결정하는 단계 및, 각 특징 요소에 대하여 상기 초기 상태로 세트하는 단계와 상기 카테고리를 결정하는 일을 반복하여 실행하는 단계를 포함한다.A second object of the present invention is a method of rapidly classifying feature data composed of a plurality of feature elements, each of which can be represented by a feature vector and represented by U bits, in a P (P is positive integer) dimension feature space directly from a calculation result. In such a method, the initial state setting step of clearing the classification data for designating a category corresponding to the reference concentration vector of the reference feature element and setting the variables i and j to "0" and "1", respectively, is provided. And a reference feature vector group respectively corresponding to n reference feature elements specified in advance.

From (k = 0 to n-1), which of the reference density vectors designated by the variables i and j (i <j) is close to the reference feature vector, and the classification data of the variables i and j are updated from the obtained result. And determining the category to which the pixel is assigned based on the classification data when the variable j becomes n-1, setting the initial state for each feature element, and determining the category. It includes a step.

가 기준 특징 벡터 d_E와 D_j의 어느쪽에 가까운가를 나타내는 소속 비트 데이타 L_ij를 구하므로써, 각 특징 요소에 대한 공간 할당 데이타를 작성하기 위한 공간 할당 데이타 작성 수단을 구비하는데, 상기 공간 할당 데이타는 각 특징요소에 있어서의 상기 조합의 전부에 대한 소속 비트 데이타로 구성되며, 대응하는 공간 할당 데이타에 의거하여 상기 미리 지정되어 있는 n개의 기준 특징 요소에 각각 대응하는 카테고리의 어디에 착목 특징 요소가 할당되는가를 나타내는 분류 데이타를 작성하기 위한 분류 수단을 구비한다.A third object of the present invention is to provide a feature data processing apparatus for classifying feature data consisting of a plurality of feature elements, each of which can be represented by a feature vector and represented by U bits, in a P (P is a positive integer) dimensional feature space. Such an apparatus comprises a first memory means for storing each component of a feature vector corresponding to each of a plurality of feature elements to be classified, and a reference feature vector group d _k corresponding to n reference feature elements specified in advance. a second memory means for storing each component of (k = 0 to n-1) and each combination of selecting two reference feature vectors d _i and d _j from the feature vector group d _k with respect to the planting feature vector x; , Relative to the planting reference concentration vector

Obtaining space allocation data for each feature element by obtaining the belonging bit data L _ij indicating which is close to the reference feature vectors d _E and D _j . It is composed of the bit data belonging to all of the combinations in each feature element, and based on the corresponding space allocation data, where the planting feature element of the category corresponding to each of the predetermined n reference feature elements is assigned. And a classification means for creating classification data indicating.

A fourth object of the present invention is to provide a feature data processing apparatus for rapidly classifying feature data consisting of a plurality of feature elements, each of which can be represented by a feature vector and represented by U bits, in a P (P is a positive integer) dimensional feature space. In the present invention, the apparatus includes a first memory means for storing each component of a feature vector corresponding to each of a plurality of feature elements to be classified, and a reference concentration vector group corresponding to n predetermined reference feature elements. second memory means for storing each component of d _k (k = 0 to n-1) and classification data storage means for storing classification data for specifying a category corresponding to the reference concentration vector of the reference feature element And setting means for clearing the classification data of the classification data storage means and setting an initial state in which the variables i and j are " 0 " and " 1, " With respect to, obtain the second one side gakkaunga of the memory means variables i and j (i <j) of the standard density vectors d _i and d _j is specified by the in, sorted from the obtained result to the variable i and j data And classification data updating means for determining the category to which the pixel is assigned based on the classification data when the variable i becomes n-1.

As described above, according to the present invention, since the calculation amount is significantly reduced as compared with the conventional clustering method, the classification process for the feature data can be speeded up.

The number of categories to be classified at this time can be arbitrarily selected, and since the classification process is performed in the distribution state in the RGB feature space of the color image data as the feature data, a good classification process that does not affect the input condition of the color image data is obtained. In this case, the linear combination processor, histogram processor, data conversion processor, and the like can be realized by a general hardware module which the image data processing apparatus generally has, and thus does not require special hardware.

A first embodiment of an image data processing apparatus according to the present invention will be described below with reference to the accompanying drawings. First, the configuration of the embodiment will be described with reference to FIG.

In FIG. 2, the CPU 11 connected to the system bus 22 controls the entire apparatus and performs necessary calculations. The main memory 12 connected to the system bus 22 includes three storage areas 13-0 to 13- corresponding to R, G, and B components, respectively, for storing calculation results of the histogram processor 21, which will be described later. 2) and has an n number of the reference concentration of the crude vector _{d k (k = 0~n-1} ) based on vector density storage area 31 for storing the R, G, B components of each vector of the reference concentration. Each storage area 13-0 to 13-2 is classified from block 0 to block n-1 corresponding to the number of reference density vectors as shown in FIG. 3 to classify each pixel constituting the color image data. It is divided into n blocks of. The number of entries in each block is 2 ⁸ , which handles color image data in which RGB components of density data are represented in 8 bits.

The R memory 14, the G memory 15, and the B memory 16 are connected to the system bus 22. Each of the memories 14 to 16 has R (red) component, G (green) component and B (for example) of density data of each pixel with respect to color image data as a feature element input through an A / D converter (not shown). Each blue component is stored. Each component of one pixel is represented by 8 bits as described above, and is stored at the same address of the memory 14 to 16. Therefore, each R, G, and B component of the pixel can be read out simultaneously. In this way, each pixel can be vector-marked by the color component of each density data. Each memory 14-16 is connected to the linear coupling processor 17 via the lines 23-25, and is connected to the histogram processor 21 via the line 30, respectively.

와, 메인 메모리(12)의 기준 농도 벡터 격납 영역(31)에 격납되어 있는 n개의 기준 농도 벡터의 조중 선택된 두 개의 기준 농도 벡터

와의 사이에서 선형 결합 연산을 행한다. 선형 결합 연산 결과는 소속 비트 데이타로서, 라인(26)을 통하여 공간 할당 데이타 메모리(18)에 출력된다.The linearly coupled processor 17, connected to the system bus 22 via line 23, is the density vector of each pixel simultaneously deciphered in the memory 14-16.

And two reference concentration vectors selected from among the n reference concentration vectors stored in the reference concentration vector storage region 31 of the main memory 12.

Perform a linear combination operation between and. The result of the linear combination operation is the belonging bit data and is output to the space allocation data memory 18 via the line 26.

에 의해 결정되는 위치에 격납한다. 따라서, n개의 기준 농도 벡터의 조중 임의로 선택된

(다만, i＜j)의 등거리 평면에서 RGB 공간을 2분할한 경우에, 해당 화소의 RGB 특징 공간에 있어서의 농도 벡터

가

축 공간에 속하는가를 나타내는 소속 비트 데이타를 L_ij라 하면, 개개의 공간 할당 데이타는 이 L_ij를 d_i, d_j의 모든 조합(다만, i＜j)에 대하여 일정한 순서로 결합하므로써 생성되는 것이다.The space allocation data memory 18 connected to the system bus 22 via line 33 transfers the belonging bit data from the linear combination processor 17 to two reference density vectors of space allocation data for that pixel.

It is stored in the position determined by. Thus, a randomly selected set of n reference concentration vectors

(However, when the RGB space is divided into two in the equidistant plane of i <j), the density vector in the RGB feature space of the pixel.

end

If the belonging bit data indicating whether belonging to the axis space is L _ij , the individual space allocation data is generated by combining the L _ij in all the combinations of d _i and d _j (but i <j) in a certain order. .

The classification processor 19 connected to the system bus 22 via the line 34 sequentially extracts the space allocation data for each pixel from the space allocation data memory 18 via the line 27, and extracts the space allocation. Based on the data, classification data k indicating where the pixel belongs to n categories d _k (k = 0, 1, ... n-1) is generated. The generated classification data k is stored in the classification data memory 20 via the line 28. The memory 20 is connected to the system bus 22 via a line 35.

The histogram processor 21, which is connected to the system bus 22 by line 36, is a component of the density data of each pixel decoded in the memory 14-16 according to the classification data stored in the memory 20. Perform a histogram operation on the. The histogram calculation result is stored in the area corresponding to the storage areas 13-0 to 13-2 of the main memory 12 via the line 36.

Next, the operation of the embodiment having the configuration shown in FIG. 2 will be described with reference to the flowcharts shown in FIGS. 4A and 4B. The R, G, and B components of each reference concentration vector in which the set d _k (k = 0 to n-1) of the n reference concentration vectors are preset by a pointing device such as a mouse are stored in the reference concentration vector storage region 31. It shall be done. The reference concentration vector may be also by means other than a man machine interface such as a pointing device, for example, a program. At this time, a process of classifying each pixel of the input color image data into n categories corresponding to the number of reference density vectors will be described.

(k=0∼n-1)중에서 임의로 선택된 2개의 기준 농도 벡터

(다만 i＜j)에 의해 정의되는 등거리 평면에 의하여 분할되는 RGB 특징 공간의 어디에 착목 화소의 농도 벡터

가 존재하는가를 즉

의 어느 벡터에 가까운가를 결정하기 위한 제어를 행한다. 이 때문에 공간 분할명령이 선형 결합 프로세서(17)에 출력된다. 여기에서, 벡터

에 관한 등거리 평면에 나타내는 벡터를

로 하면, 벡터

는 다음식(1)을 만복시킨다.The CPU 11 sets the set of n reference density vectors in accordance with the classification instruction in step S2.

2 reference concentration vectors arbitrarily selected from (k = 0 to n-1)

Where the density vector of the implanted pixel is in the RGB feature space divided by the equidistant plane defined by i <j

That exists

Control to determine which vector is close to is performed. For this reason, the spatial division instruction is output to the linear combination processor 17. Here, vector

A vector representing the equidistant plane of

Let's say, vector

Undo the following equation (1):

가 존재하는 부분 공간을 결정하기 위하여 공간 분할 명령에 따라 메모리(14 내지 16)의 동일 어드레스가 리이드 액세스(read acess)되고, 당해 화소의 농도 벡터

=(X_R, X_G, X_B)가 해독된다. 해독된 농도 벡터

는 라인(23 내지 25)를 통하여 선형 결합 프로세서(17)에 출력된다. 또, 두개의 기준 농도 벡터 d_i=(r_i, g_i, b_i)와 d_j(r_j, g_j, b_j)(다만 i＜j)가 메인 메모리(12)의 기준 농도 벡터 격납 영역(31)에서 해독되고, 계수(r_i-r_j), (g_i-g_j) 및 (b_i-b_j)와, 바이어스 (｜d_i｜³-｜d_j｜²)가 계산된다. 그 계산 결과는 라인(32)을 통하여 선형 결합 프로세서(17)에 공급된다.Concentration vector

The same address of the memory 14 to 16 is read-accessed in accordance with the space division command to determine the subspace in which is present, and the density vector of the pixel

= (X _R , X _G , X _B ) is decoded. Decoded concentration vector

Is output to the linear combination processor 17 via lines 23-25. In addition, two reference concentration vectors d _i = (r _i , g _i , b _i ) and d _j (r _j , g _j , b _j ) (where i <j) are stored in the reference density vector of the main memory 12. Decoded in the area 31, the coefficients (r _i -r _j ), (g _i -g _j ) and (b _i -b _j ) and the bias (| d _i | ^3- | d _j | ² ) are calculated do. The calculation result is supplied to the linear combination processor 17 via line 32.

The linear combination processor 17 obtains the value d _ij of the expression (2) using the expression (1).

dij = (r _i -r _j ) · X _R + (g _i -g _j ) · x _G + (b _i -b _j ) · X- (| d _i | ^2- | d _j | ² ). … … … … (2)

는

측의 공간에 속하는 것으로 판정하고, 논리 "1"의 소속 비트 데이타 L_ij를 생성한다. 연산 결과가 부이면, 농도 벡터

는

측 공간에 속한다고 판정되고, 논리 "0"의 소속 비트 데이타 L_ij가 생성된다.When the linear combination operation result d _ij of the formula (2) is obtained, the linear combination processor 17 checks whether the operation result is positive, "0" or negative. If the result of the operation is positive or "0", the density vector of the pixel

Is

It determines with belonging to the space of the side, and produces | generates belonging bit data L _ij of logic "1". If the result of the operation is negative, the concentration vector

Is

It is determined to belong to the side space, and the belonging bit data L _ij of logic "0" is generated.

In step S4, the linear combination processor 17 transfers the generated bit data L _ij through the line 26 to the sixth bit of the space allocation data for the pixel x of the space allocation data memory 18 via the line 26. Fill in as shown in the figure. Further, e is represented by the following equation (3), where (i, j) of d _ij is (0, 1) (0, 2). (0, n) (1, 2) (1, 3)... Calculate the number of rows in parallel with (n-1, n).

Perfume to j to i-1 Perfume to j to i-1

= ni-i (i + 1) / 2 + (j-i-1)

= i (n-i-1) 2 + j-i-1

Where n is the reference concentration vector number.

로 하여

을 채용한 경우에는 제0비트에 소속 비트 데이타 L₁이 세트되고 ,

을 채용한 경우에는 제1비트에 소속 비트 데이타 L₂가 세트된다.Thus, as shown in FIG.

By

Is adopted, the bit data L ₁ belonging to the 0th bit is set.

Is adopted, the bit data L ₂ belonging to the first bit is set.

의 모든 조합(다만 i＜j)에 대하여 종료되었는지의 여부가 조사된다.When the operation of the steps S2 and S4 ends, the calculation process and the determination process are performed in step S6.

It is checked whether or not all combinations of (but i <j) are finished.

의 조합(다만 i＜k)은, n(n-1)/2와 같다. 따라서, 스텝(S6)에서 "아니오"로 판정되면, 선형 결합 프로세서(17)는 스텝들(S2, S4)의 동작을 반복하게 된다. 이로 인해 선형 결합 프로세서(17)가 제작동되기 때문에 공간 분할 명령이 재차 출력된다.any set of n reference concentration vectors

Is a combination of n (n-1) / 2. Thus, if determined "NO" in step S6, the linear combination processor 17 repeats the operations of steps S2 and S4. As a result, since the linear combination processor 17 is manufactured, the spatial division instruction is output again.

If "Yes" is determined in step S6, the program proceeds to step S7. In step S7, all the pixels stored in the memories 14 to 16 are examined whether or not the arithmetic processing and determination processing have been completed. If no, steps S2 to S6 are repeated. If YES, step S8 is executed. When it is determined in step S7 that " YES &quot;, n (n-1) / 2 belonging bit data L _ij is written in the space allocation data for each pixel of the space allocation data memory 18. do.

In step S8, a data conversion instruction is output to the classification processor 19 to determine the category to which each pixel belongs to the classification processor 19. FIG. According to this instruction, the classification processor 19 decodes each space allocation data through the line 27 in the space allocation data memory 18, and based on the decoded space allocation data, the pixel belongs to the n categories s. Generate classification data s to indicate whether there is. The generated classification data is stored in the classification data memory 20 through the line 28. The above processing is executed for all the space allocation data stored in the space allocation data memory 18.

의 비트상태는 제8도에 도시한 바와 같이 2³=8과 같고, 해당 화소가 카테고리 0, 카테고리 1, 카테고리 2를 지정하는 기준 농도 벡터

의 어느 쪽의 공간에 속하는 가를 나타낸다. 클래스(class) "0"에 속하는 데이타는 L₁ㆍL₂=1이고, 클래스 "1"에 속하는 데이타는

ㆍL₁₂=1이고, 클래스 "2"에 속하는 데이타는

=1이다. 그러나 이와 같이 하면, (0, 1, 0)과 (1, 0, 1)의 경우는 이해하기가 곤란하게 된다. 알기 어려운 화소는 버려도 좋다. 본 실시예에서는 L₁에서 차례로 데이타를 판단하고, (1, 0, 1)은 클래스 "0"에, (0, 1, 0)은 클래스 "1"에 소속시키고 있다. 이상의 설명으로부터 명백히 알 수 있는 바와 같이, 분류프로세서(19)에 의한 스텝(S8)의 동작이 모두 종료한 시점에서 분류데이타 메모리(20)에는 칼라 이미지 데이타의 분류결과가 화소 단위로 격납되어 있는 셈이 된다.Here, the case of classifying data generation by the classification processor 19 will be described by taking, for example, the case where n = 3, i.e., when the RGB feature space is divided into categories 0 to 2 as shown in FIG. . In the case of n = 3, the number of belonging bit data in the space allocation data is 3 · 2/2 = 3, and is composed of bit data of L ₁ , L ₂ , and L ₁₂ in order from the 0th bit. remind

The bit state of is equal to 2 ³ = 8 as shown in FIG. 8, and the reference density vector for which the corresponding pixel specifies category 0, category 1, and category 2.

Indicates which space in. Data belonging to class "0" is L ₁ ㆍ L ₂ = 1, and data belonging to class "1"

L ₁₂ = 1 and data belonging to the class "2"

= 1. However, in this case, it becomes difficult to understand the case of (0, 1, 0) and (1, 0, 1). Pixels that are difficult to understand may be discarded. In this embodiment, data is judged in turn in L ₁ , where (1, 0, 1) belongs to class "0", and (0, 1, 0) belongs to class "1". As is apparent from the above description, the classification data memory 20 stores the classification result of the color image data in pixel units at the time when the operation of step S8 by the classification processor 19 is finished. Becomes

(다만 i＜j)의 조합에 대한 등거리 평면에서, RGB 특징 공간상의 각 화소를 분류하고, 각 화소가 기준 농도 벡터

의 어느 쪽의 부분 공간에 속하는 가를 검출하는 동작을

의 모든 조합, 즉 n(n-1)/2 방식에 대하여 실행한다. 또한 이 검출동작이 종래에 있어서의 모든 기준 농도 벡터까지의 거리를 계산하는 것과는 달리, 상기 (2)식에 보이는 선형 결합 연산의 결과에 따라 한번에 행할 수 있으므로, 카테고리 분류를 고속으로 행할 수 있다.In the present embodiment, two reference concentration vectors arbitrarily selected among the set of n reference concentration vectors d _k (k = 0 ^N n-1)

In the equidistant plane for the combination of i <j), each pixel in the RGB feature space is classified, and each pixel is a reference density vector.

To detect which subspace of

For all combinations of n, n (n-1) / 2. Unlike the calculation of the distances to all the reference density vectors in the related art, the detection operation can be performed at once according to the result of the linear combination operation shown in Equation (2), so that the category classification can be performed at high speed.

However, when the classification processor 19 ends the operation of step S8, it is notified to the CPU 11 via the system bus 22.

The CPU 11 starts the histogram calculation processor 21 in the case of the system bus 22. As a result, the histogram processor 21 performs the histogram calculation displayed in step S10.

That is, the histogram processor 21 first accesses the R memory 14 and the classification data memory 20 simultaneously in units of pixels, and the classification data from the classification data memory 20 and the R component from the R memory 14 are simultaneously accessed. A histogram calculation is performed for the data connected to the concentration of. Thereby, the histogram h _rk (m) which shows the number of pixels which the density | concentration of R component becomes m (m = 0, 1 ..., 255) among the pixel group which belongs to category k (k = 0, 1 ..., n-1). ) Is obtained. The histogram h _rk (m) is transmitted to the main memory 12 via the system bus 22, and the k-th category of one area 130-2 of the three storage areas 13-0 to 13-2 is provided. It is stored in the mth entry of. The histogram processor 21 finishes the histogram operation for the R memory 14 and the classification data memory 20, and then performs the histogram operation for the G memory 15 and the B memory 16 for the R memory 14. The same as). Thereby, histograms h _gk (m) and h _bk (m) for the G and B components of the pixel can also be obtained.

를 구하기 위하여, 메인 메모리(12)의 3개의 격납영역(13-0 내지 13-2)에 격납되어 있는 h_rk(m), h_gk(m), h_bk(m)를 사용하여 다음 식을 계산한다.When all the histogram calculations of the histogram processor 21 are complete | finished, CPU11 will mean the average density vector for every category K.

In order to obtain the following equations, h _rk (m), h _gk (m), and h _bk (m) stored in the three storage areas 13-0 to 13-2 of the main memory 12 are Calculate

Here, n represents the number of pixels belonging to category k.

를 구하면, 스텝(S16)에서 이

가 미리 설정되어 있는 오차의 범위내에서

와 같은가 여부를 모든 카테고리 k에 걸쳐서 체크한다. 만약, 하나라도 같지 않은 것이 있으면, CPU(11)는

를 스텝(S12)에서 구한

로 바꿔놓고, 선형 결합 프로세서(17)를 재기동시킨다. 그리고, 상기의 동작이 반복되고 모든 k에 대하여

가

의 미리 설정되어 있는 오차 범위 이내가 되면, 영역분할 처리는 종료하게 된다.The CPU 11 performs the average concentration vector for each category k by the processing of step S12.

If is found, the step (S16)

Is within the range of preset error

Is checked across all categories k. If any one is not the same, the CPU 11

Obtained at step S12

Replace with and restart the linear coupling processor 17. And the above operation is repeated and for all k

end

When it is within the preset error range of, the area division processing is terminated.

The above state is shown over FIGS. 9A-9C. In Fig. 9A, two reference density vectors denoted by the symbol + are arbitrarily given to designate two regions, and a state obtained by dividing an image into two in the equidistant plane of the two reference density vectors is shown. FIG. 9B shows a state where the image is divided twice again in the equidistant plane of each average density vector of the two regions divided into two in the equidistant plane shown in FIG. 9A. In FIG. 9C, each average density vector of two regions divided by two in the equidistant plane shown in FIG. 9B (shown by the symbol ▲) and the image are further divided in two equidistant planes of the two vectors. The state of the final cluster in which each average concentration vector of the two regions obtained is matched is shown.

Next, a second embodiment will be described. The configuration is as shown in FIG. 2, but the method of creating the classification data by the classification processor 19 is different from that of the first embodiment.

의 조에 대한 소속 비트 데이타 L₁이 조사된다. 그 결과에 따라 L₁이 "1"이면 횡측상에서 "0"이 선택되고 L₁이 "0"이면 횡측상에서 "1"이 선택된다. L₁이 "1"이면 다음에 L₂가 조사된다. L₁이 "0"이면 다음에 L₁₂가 조사되고, 횡측 "1"과 종축 L₁₂의 교점의 난에 표시되는 클래스가 된다.In the second embodiment, the classification is performed using the data in FIG. In this example, when the reference concentration data is five, and L _ij is " 1 " in the upper portion, a class is shown when L _ij is " 0 " in the lower portion. First, the reference concentration vector

The belonging bit data L _{1 for the set of} is examined. As a result, when L ₁ is "1", "0" is selected on the lateral side, and when L ₁ is "0", "1" is selected on the lateral side. If L ₁ is "1", then L ₂ is irradiated. If L ₁ is "0", L ₁₂ is irradiated next, and it becomes the class displayed in the column of the intersection of the horizontal "1" and the longitudinal axis L ₁₂ .

At this time, if L ₁₂ is "0", the class is "2", and then L ₂₃ is examined. In the sorting processor 19, a look-up table according to the above considerations is prepared, and the space allocation data is sorted at high speed. The operation of the other second embodiment is the same as that of the first embodiment.

Next, a third embodiment will be described.

The configuration of the third embodiment is shown in FIG.

The main memory 12 'is the same as the main memory 12, but further includes an i register 42 and a j register 43. A linear combination processor (17 ') is a linear combination processor 17 and gateuna, the value d to not necessarily to write the belonging-bit data after the space allocation data obtained, _ij that value d _ij is whether Mrs. Jeong In-ji CPU (11') Notify. The classification data in the classification data memory 20 'is ganged by the CPU 11' in accordance with the calculation result by the linear combination processor 17 '. Therefore, as compared with the first embodiment, the space allocation data memory 8 and the classification processor 19 are omitted.

The operation will be described below with reference to FIG. Here, the number n of reference concentration vectors is five.

If a sort command is inputted first, the sort data of the sort data memory 20 'is cleared in step S22. Further, the value i of the i register 32 of the main memory 12 'is set to "0" and the value j of the j register 33 is set to "1". A linear combine command is output to the linear combine processor 17 'via line 32.

의 데이타가 R, G 및 B 메모리(14, 15 및 16)에서 라인(23, 24 및 25)를 통하여 해독된다. 또한 값 i와 j에 의하여 지정되는 기준 농도 벡터

가 메인 메모리(12')의 기준 농도 벡터 격납영역(31)에서 해독된다. 대상 화소의 농도 벡터 x와 기준 농도 벡터

로부터, 제1의 실시예의 스텝(S2)과 같이하여 값 d_ij를 계산하기 위한 선형 결합연산이 실행된다.In response to the command, in step S24, by the linear combination processor 17 ', the density vector for the first pixel to be classified.

Is decoded via lines 23, 24 and 25 in R, G and B memories 14, 15 and 16. Also the reference concentration vector specified by the values i and j

Is read from the reference concentration vector storage area 31 of the main memory 12 '. Density vector x and reference density vector of the target pixel

From this, a linear combined operation for calculating the value d _ij is performed as in step S2 of the first embodiment.

The result of the operation determines whether d _ij is positive or negative. It is assumed here that d _ij is positive. The determination result is notified to the CPU 11 'via the line 23.

In step S26, it is checked whether the value j is equal to n-1. If YES, step S36 is executed. If no, step S28 is executed. Since the reference density vector number is 5, it is determined as "no", and then step S28 is executed. In step S28, it is checked whether the value d _ij in the linear combination processor 17 'is positive or negative. If the value d _ij is positive, step S30 is executed; if negative, step S32 is executed. In this case, since the value d _ij is positive, step S30 is executed and the value of the j register 43 is decreased by one. Therefore, the value i becomes "0" and the value j becomes "2".

Thereafter, step S24 is repeated in the same manner.

This time, it is assumed that d _ij is negative. Then, it is determined as "no" in step S28 following step S26, and step S32 is performed. In step S32, the value of the j register 43 is set in the i register 42, and the value of the j register 43 is increased by one. That is, the value i is "2" and the value j is "3". The classification data for this pixel in the classification data memory 20 'is updated by the CPU 11' via the line 33 according to the value i. Therefore, the classification data is updated to "2".

After that, in step S34, it is determined whether or not the value of the classification data is equal to n-1. If YES, the classification processing for the pixel is finished, and step S38 is executed. In the present case, the classification data is "2" and "no" is determined.

Subsequently, step S24 is executed in accordance with the values i and j. If this result value d _ij is assumed, steps S26, S28 and S30 are executed. The values i and j become "2" and "4", respectively. Thereafter, step S24 is executed. As a result, d _ij becomes denial.

Since the value j is "4", it is determined as "Yes" in step S26, and then step S36 is executed. If no, step S28 is executed. In step S36, if the value d _ij is positive, step S38 is executed. If the value d _ij is negative, the classification data is updated according to the value j. That is, the value of the classification data is "4". Thereafter, step S38 is executed.

In step S38, it is checked whether or not the classification process for all the pixels constituting the image data to be processed is completed. If no, the program returns to step S22 to process the density vector for the next pixel. If YES, the process after step S10 in FIG. 4B is executed. After executing step S14 of FIG. 4B, the program returns to step S22 again.

As described above, the pixel is classified into four values of d ₁ , d ₂ , d ₂₃ , and d ₂₄ , that is, four calculations.

In the first or second embodiment, since all combinations of i and j are considered, the number of reference concentration vectors is 5, so 5 (5-1) / 2 = 10 calculations were required. Therefore, classification of concentration data can be realized at a higher speed. In this embodiment, the space allocation data memory and the classification processor can be omitted.

In addition, in the above embodiment, the division of the color image has been described, but the present invention can be applied to clustering processing for multi-spector information such as a remote sensing image, a characteristic x-ray image, and the present invention. As will be described with respect to the data, the invention can be applied throughout the feature vector as described above.

Claims (26)

A method for rapidly classifying feature data consisting of a plurality of feature elements each represented by a feature vector in a P (P is a positive integer) dimensional feature space and characterized by u bits, the method comprising n predetermined numbers Reference feature vector group corresponding to each reference feature element

two reference feature vectors from (k = 0 to n-1)

For each combination to select, the belonging bit data L _ij indicating which of the two reference feature vectors of each combination is close to the reference feature vector is obtained over all of the combinations for all the feature elements constituting the target feature data. And generating space allocation data for each feature element, wherein the space allocation data consists of the belonging bit data for all of the combinations in each feature element, based on the corresponding space allocation data. And a step of generating classification data indicating where each feature element of the category corresponding to each of the n reference feature elements specified in advance is assigned. The method according to claim 1, wherein the step of creating space allocation data comprises: performing a first process for _obtaining belonging bit data L _ij over all of the combinations for one feature element, and all the feature elements constituting the target feature data; Performing the first process with respect to the method. 2. The method of claim 1, wherein the step of creating space allocation data comprises: performing a second process of obtaining the belonging bit data L _ij over all the feature elements for one of the combinations, and for the all of the combinations; Performing the process of step 2; 2. The belonging bit data for a combination of one feature element according to claim 1, wherein

Is determined based on a value of d _ij , wherein d _i and d _j are two reference feature vectors of a combination selected from the n reference feature vector groups d _k (k = 0 to n-1),

Is a feature vector representing a feature element. 5. A method according to claim 4, wherein the relationship of two reference concentration vectors i &lt; j selected as a combination is satisfied. The reference feature vector of claim 4, wherein when the value of Equation (1) is positive or " 0 "

If the value of the above equation (1) is negative, the feature element is a reference feature vector.

And the belonging bit data L _ij determined to be close to is stored at a bit position determined according to a combination of two reference feature vectors. The method of claim 3, wherein the two reference feature vectors

Wow

The belonging bit data L _ij created for the combination of is the reference feature vector of the divided feature space.

Logic halttaeneun the feature vectors present in the side "1", and a reference feature vector d _j logic "0" halttaeneun the feature vectors present in the side of the divided feature space, based on vector density d ₀ and a standard density vector

When the belonging bit data L _ot for (t = 1 ^N n-1) is a logic “1”, a classification indicating that the zero category falls among the 0 categories from 0 to n-1 corresponding to the reference concentration vector. Create data, and the belonging bit data L _ts for the reference concentration vector d _t (t = 0 ^N s-1) and the reference concentration vector d _s are logical "0", and the reference concentration vector d _s and the reference concentration vector d _t When the belonging bit data L _st for (t = s + 1 ^N n-1) is logical " 1 &quot;, the classification data indicating that the bit data L _st belongs to the s category among the n categories corresponding to the reference concentration vector is generated, When the belonging bit data L _tn-1 for the vector d _n and the reference concentration vector d _t (t = 0 ^N n-2) is logical "0", the n-1 category of the n categories corresponding to the reference concentration vector is applied. And generating classification data indicating that it belongs. The method of claim 3, wherein the classification data with reference to the look-up table is determined from the space allocation data, the lookup table category "s" to belong to examine the belonging-bit data L _st belongs bit data L _st "to 1 When belonging to the category "s" is determined to belong to category "s", and then the belonging bit data L _sj (j = t + 1) is determined to belong to the category "t" when belonging bit data L _st is "0". And _{classifying the} next bit data L _ij (j = t + 1) by repetitively checking until j becomes n-1. The method according to claim 1, wherein an average value for each component of the feature vector in each category is determined from the generated classification data and each component of the feature vector of each feature element, and the average value of each component of the determined feature vector is a corresponding criterion. Determining whether each component of the feature vector is within a predetermined error, and when the average feature vector having an average value of each component is determined as a reference feature vector when not within a predetermined error range, And repeating the space allocation data creation, the classification data creation, the average feature vector creation, and the determination process until each component is within a predetermined error. 10. The method of claim 9, wherein the determining of the mean feature vector comprises performing a histogram operation for each category corresponding to the reference feature vector for the generated classification data and each component of the feature vector of each feature element, A _gk = (1 / N _k ) in the histogram operation

obtaining each component of the average feature vector using the equation h _gk (m) · m, where A _gk is the q (q = 1 to P) lateral direction of the average feature vector in category k. N _k denotes the number of feature vectors assigned to category k, m denotes the feature amount of each feature vector element, and h _gk (m) denotes q (q = 1-P) representing a histogram of lateral elements. A method of rapidly classifying feature data consisting of a plurality of feature elements each represented by a feature vector and represented by a u-bit in a P (P is a positive integer) dimensional feature space, corresponding to a reference density vector of a reference feature element A reference feature vector group corresponding to n reference feature elements, each of which is set in an initial state by clearing the classification data for specifying a category to be specified and setting the variables i and j to "0" and "1", respectively.

From (k = 0 to n-1), which of the reference density vectors designated by the variables i and j (i <j) is close to the reference feature vector, the classification data is updated with the variables i and j from the obtained result. Determining the category to which the corresponding pixel is assigned by the classification data when the variable j is n-1, and setting each feature element to the initial state, and repeatedly performing the determining of the category. High speed data classification method characterized in that. 12. The method of claim 11, wherein the determining of the category further comprises determining that the category to which the pixel is assigned is determined when the value of the updated classification data is n-1. 12. The method of claim 11, wherein determining the category comprises two reference concentration vectors, wherein one feature element is designated by variables i and j.

When x is the feature vector of the feature element to determine which is close to,

The value of d _ij is d _ij

If it is 0, it is determined to be close to the reference concentration vector d _i , and if d _ij <0, it is determined

Determining that it is close to. The method of claim 13, wherein the determining of the category comprises: d _ij

When 0, the variable i and the classification data are not updated, but the variable j is updated by 1, and when d _ij <0, the classification data is updated with the current variable j value, the variable i is updated with the variable j value, and the variable j is 1 Updating by increments. 12. The method of claim 11, wherein determining the average value of each component of the feature vector in each category from the generated classification data and each component of the feature vector of each feature element, and the average value of each component of the determined feature vector corresponds. Determining whether each component of the reference feature vector is within a predetermined error, and using the average feature vector having an average value of each of the components when not within a predetermined error as the reference feature vector, And setting the initial state until each component is within a predetermined error, determining the category, creating the average feature vector, and determining whether the component is within the error range. Characterized in that the method. 6. The method of claim 5, wherein the determining of the average feature vector comprises performing a histogram operation for each category on the generated classification data and each component of the feature vector of each feature element, and A _{gk in the} result of the histogram operation. = (1 / N _k )

obtaining each component of the average feature vector using the equation h _gk (m) · m, where A is the element in the q (q = 1 to P) axis direction of the average feature vector in category k. Where N _k represents the number of feature vectors assigned to category k, m represents the feature quantities of elements of each feature vector, and h _gk (m) represents q (q = 1) of the average feature vectors in category k. ~ P) A histogram of the elements in the axial direction. A feature data processing apparatus for classifying feature data consisting of a plurality of feature elements represented by u bits each represented by a feature vector in a P (P is a positive integer) dimensional feature space, each of a plurality of feature elements to be classified. Of the first memory means 12 for storing each component of the feature vector corresponding to and the reference feature vector group d _k (k = 0 to n-1) corresponding to the n reference feature elements specified in advance. For each combination of selecting the two reference feature vectors d _i and d _j from the feature vector group d _k with respect to the second feature means 14, 15, 16 for storing each component and the planting feature vector x, By obtaining the belonging bit data L _ij indicating whether the planting reference concentration vector x is close to the reference specific vector d _i and d _j , the space allocation data creating means 18 for creating the space allocation data for each feature element is obtained. Comparing to the above, the space allocation data is composed of belonging bit data L for all the combinations in each feature element, and each category corresponds to the n reference feature elements specified in advance based on the corresponding space allocation data. And sorting means (19) for creating sorting data indicating where the picking feature elements are to be assigned. 18. The method according to claim 17, wherein the belonging bit data (1)

Device determined by the value of d _ij . 19. The apparatus of claim 18, wherein the belonging bit data has a first logical value when the value of Equation (1) is positive or " 0 &quot;, and a second logical value when the value of Equation (1) is negative, The sorting means includes a range from 0 to n-1 corresponding to the reference concentration vector when the belonging bit data L _ij for the reference concentration vector d ₀ and the reference concentration vector d _t (t = 1 n-1) is the first logical value. Create classification data indicating that the category belongs to the 0th category among the n categories, and the belonging bit data L _ts for the reference concentration vector d _t (t = 0 to s-1) and the reference concentration vector d _s is the second logical value. And the belonging bit data L _st for the reference concentration vector d _s and the reference concentration vector d _s (t = s + 1 to n-1) is the first logical value, the s of the n categories corresponding to the reference concentration vector. creating a classification indicating that data belonging to the category, for the standard density vector d _n-1 and the reference density vector _{d t (t = 0~n-2} ) Sector-bit data L _tn-1 is the device, characterized in that to create the classification data indicating that n of the categories corresponding to the logical value when the reference vector concentration of 2 n-1 belonging to the category. 19. The apparatus of claim 18, wherein the belonging bit data has a first logical value when the value of Equation (1) is positive or " 0 &quot;, and a second logical value when the value of Equation (1) is negative, the classification means are investigated belongs bit data L _st to belong to the category "s", and the Sector bit data L _st is the judges that belong to a logical value when category "s" 1, belonging to the following bit data L _ij (i = t + 1) is examined to determine that the belonging bit data L _st belongs to category "t" when the second logical value, and then belong to belonging bit data L _ij (j = t + 1). And a lock-up table configured to determine the class by iteratively examining j of the bit data until n-1. 18. The apparatus according to claim 17, further comprising: space allocation data creating means for creating the space allocation data for each feature vector of all the feature elements constituting the feature data, first drive means for operating the classification means, and created classification Average value calculation means for calculating the average value of each component of the feature vector in each category from all components of the data and the feature vector of each feature element, and the average value of each component of the determined feature vector is the corresponding reference feature. Error judging means for determining whether or not within a predetermined error in each component of the vector and an average feature vector having an average value of each calculated component when it is not within a predetermined error range as the reference feature vector. Stored in the memory means, until each component of all the mean vectors is within a predetermined error, Device characterized in that a second drive means for driving the drive means, said average value calculation means and the error determining means of the first group as well. A feature data processing apparatus for rapidly classifying feature data composed of a plurality of feature elements, each of which can be represented by a feature vector and represented by u bits, in a P (P is a positive integer) dimensional feature space, wherein the plurality of features to be classified Reference density vector group d _k (k = 0 n-1) corresponding to the n reference feature elements in which the first memory means 12 'for storing each component of the feature vector corresponding to each of the elements is predefined. Second memory means (14, 15, 16) for storing each component of &lt; RTI ID = 0.0 &gt;,&lt; / RTI &gt; and classification data storage means 20 'for storing classification data for specifying a category corresponding to the reference concentration vector of the reference characteristic element. ), Initial value setting means (42, 43) for clearing the classification data of the classification data storage means, and setting the variables i and j to initial states of "0" and "1", and the tree feature vector Is the second memory Find which of the two reference concentration vectors d _i and d _j specified by the variables i and j (i <j) in the means, update the classification data of the variables i and j from the result, and and classification data updating means (11 ') for determining the category to which the pixel is assigned by the classification data when n-1. 23. The apparatus of claim 22, wherein the classification data updating means further includes first determining means for determining that a category to which a corresponding feature element is assigned is determined when the value of the updated classification data is equal to n-1. Device. 24. The method of claim 23, wherein the classification data updating means selects a feature vector of the feature element to determine which one is close to the two reference concentration vectors d _i and d _j designated by the variables i and j. hereinafter to as x (1) formula that _{is, d ij = (d i -d} j) and _{{x- (d i + d j} ) / 2} values d _ij is d _ij of

And means for determining that it is close to the reference concentration vector d _i when 0, and determining that it is close to the reference concentration vector d _j when d _ij <0. 25. The method of claim 24, wherein the classification data updating means is d _ij

When 0, the variable i and the classification data are not updated but the variable j is incremented by 1, when d _ij <0, the classification data is updated with the current variable j value, the variable i is updated with the variable j value and the variable j is Means for updating to increase by one. 24. The apparatus according to claim 23, further comprising: first driving means for driving the setting means and the classification data updating means for each feature element constituting the feature data, and from the generated classification data and each component of the feature vector of each feature element. An average value calculating means for calculating an average value of each component of the feature vector in each category, and for determining whether the average value of each component of the calculated feature vector is within a predetermined error in each component of the corresponding reference feature vector The second determining means and an average feature vector having an average value of each component determined when not within a predetermined error are stored in the second memory means as reference feature vectors and each component of all the average feature vectors is within a predetermined error. Driving the first driving means, the average value calculating means and the second determining means until And a second drive means for the device.

Images