JP2002170122A - Method and program for texture analysis - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a texture analyzing method dividing a texture image into domains wherein texture is uniform through the detection of excellent features of a texture identification capability using a density histogram method, while also enabling even color images to be subjected to proper texture analysis. SOLUTION: The distance Dist (A, B) between two textures A and B is defined as the sum of absolute values of differences between densities provided by density histograms. Based on whether or not the distance Dist (A, B) is smaller than a predetermined threshold, a determination is made as to whether the A and B are the same texture, whereby the texture image can be divided into domains. That is, the texture image is divided into appropriately sized blocks in each of which the texture is uniform, so as to produce density histograms. For two sets of blocks, their textures are identified using the density histograms and an assembly of adjacent blocks is extracted as one domain.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a texture analysis method and a program for a complex image in which texture plays an important role, such as an aerial photograph, a satellite image, or a medical image.

[0002]

2. Description of the Related Art A texture is a pattern in which a fine repetitive pattern exists microscopically in an image and which is uniformly felt macroscopically. The texture is defined by this repeating pattern and its arrangement rules. Humans use texture for object identification and identification and depth perception. When analyzing various images such as aerial photographs, satellite images, and medical images, texture analysis is performed to classify regions based on textures.

[0003] A texture has a structural texture whose elements are relatively easy to extract and the arrangement rule can be easily understood as a relation between the elements, and a description based on the element and the arrangement rule is difficult. There are statistical textures whose features cannot be described. The present invention is directed to the latter statistical texture.

[0004] Texture features are often mathematically defined for purposes such as image classification, region segmentation, and object identification. The simplest texture features are density histograms and statistics that can be derived therefrom. For example, a certain amount of texture can be distinguished by statistics such as the average, variance, third moment, and fourth moment of the density histogram. Other texture features include co-occurrence matrices, difference statistics, Fourier features, autoregressive models, and the like. However, these texture features are much more complex to compute than those using a density histogram and are often not practical. For information on texture characteristics and textures in general, see, for example, "Mori Onoe (ed.): Image Processing Handbook: Shokodo (Showa 62)
Year) ”.

[0005]

Conventional texture features using density histograms are not sufficient for texture identification. For example, although the density histograms of two textures have significantly different shapes, their averages and variances are almost the same, so that there is a case where they cannot be distinguished by a conventional method using a density histogram.

Further, in a color image, since the number of colors becomes enormous, it is difficult to directly apply a texture analysis method using a density histogram. For example, the density of a color image is 8b for each component of red, green, and blue.
If it, then the number of colors is 2 to the 24th power, or 1677721
There are six colors. On the other hand, a block of a color image usually contains several hundreds of pixels. At this time, since the number of colors is enormous compared to the number of pixels included in the block, most of the values of the density histogram of each block are 0.
And the meaning of the density histogram is lost. In a multispectral image such as a satellite image, the number of colors is even greater, and this situation becomes more severe.

Therefore, a method of reducing the number of colors to the order of several tens is an important problem in texture analysis of a color image. One way to reduce the number of colors is to ignore the lower bits of the color components. However, in this method, the contrast of the color image is excessively emphasized, so that the generated color image is considerably different from the original color image. Therefore, it is not suitable for texture analysis.

A first object of the present invention is to find a texture feature having a better texture discriminating ability while maintaining the good calculation efficiency of the conventional method using a density histogram, and to form a texture image with a uniform texture. That is, the method is applied to the division into regions. Further, the second aspect of the present invention
An object of the present invention is to provide a texture analysis method and a program capable of performing appropriate texture analysis even for a color image.

[0009]

In order to enhance the texture discriminating ability, it is desirable to use a texture feature that more closely reflects the shape of the density histogram. Compared to a texture feature calculated from a single density histogram, a texture feature quantifying the difference in shape between the two density histograms is considered to be more powerful in terms of texture identification.

The present invention is characterized in that, for two textures, a texture feature defined by the sum of the absolute values of the differences between the densities of the density histograms is used.

FIGS. 1 (a) and 1 (b) show the texture A
2 and B show the density histograms. The horizontal axis of the density histogram indicates the pixel density, and the vertical axis indicates the frequency. Fig. 1 (c)
Is obtained by superimposing the two density histograms, and the black portion indicates the difference. The area of this black portion corresponds to the texture feature of the present invention.

More precisely, two textures A and B
In contrast, let the respective density histograms be f _A (k) and f _B (k). Here, k indicates a density having a value of 0 to N. The distance Dist (A, B) between the textures A and B is defined by the following equation (1).

[0013]

(Equation 1)

This definition can be applied to a multispectral image if the density k is extended to a multidimensional vector.

The distance Dist (A, B) helps to identify the texture. For example, if Dist (A, B) <T holds for a certain threshold T, it can be determined that A and B are the same texture. If this expression does not hold, it can be determined that A and B are different textures.

The texture image can be divided into regions by using the above-described texture identification method. The distance Dist is considered to define the "distance" between two elements in the texture feature space. In fact, it is easy to prove that this satisfies the so-called distance axiom. Therefore, by measuring the distance between the elements, a texture cluster is detected in the feature space, and the elements belonging to the cluster are inversely mapped in the image space to generate a (generally plural) uniform texture region. can do.

Specifically, the area can be divided by the following procedure. It is assumed that the input texture image is a sufficiently large image. (1) A density image is generated by dividing a texture image into blocks of an appropriate size in which the texture is uniform. (2) For two sets of blocks, the texture is identified using the density histogram, and the number of the texture group is assigned to each block. (3) A group of blocks having the same group number and adjacent to each other is extracted as one region.

For black and white images, good results are obtained using the above-described texture analysis method. However, as described in the section of the problem to be solved by the invention, when analyzing a color image, it is necessary to reduce the number of colors. The present invention is characterized in that vector quantization is applied as a method for effectively reducing the number of colors. In general, vector quantization is a mapping from a set of vectors to a finite set of vectors. In the present invention, the density of each pixel of a color image is regarded as one vector, and the color component of the density is made to correspond to the vector component. Knowledge on vector quantization is described in, for example, "Gersho A. and Gray RM, Saburo Tazaki et al. Translation: Vector Quantization and Information Compression: Corona (1998)".

A vector quantizer Q of dimension k and size N
Is a mapping from a k-dimensional Euclidean space R to a finite set C containing N output vectors named code vectors. That is, this is expressed as Q: R → C. Here, C = {y1, y2,..., YN}, and there is a relationship of yi∈R for each i∈I = {1,2,.
The set C is called a codebook and has a size N.

For all N points yi, vector quantizer Q divides R into N regions or cells Ri. i
The second cell is defined by Ri = {x∈R: Q (x) = yi}. The encoder is completely defined by dividing R into cells of R1, R2,..., RN, and the decoder is fully defined by the codebook C = {y1, y2,.

First, consider the optimization of an encoder for a fixed decoder. Let d (x, y) be a non-negative function called a strain measure in space R. For a given codebook, the optimal division is a division that satisfies the following nearest neighbor condition. Ri = {x: d (x, yi) ≤ d (x, yj); all j} ie Q (x) = yi; where d (x, yi) ≤ d (x, yj) all j

Thus, given a decoder, the encoder maps with minimal distortion or with the nearest neighbor.

Conversely, consider the optimization of the decoder for a fixed encoder. In other words, consider the optimality of the codebook for a given partition. This leads to a centroid condition for defining a code vector associated with each divided region. Set R
Is defined as follows:

First, given a point X in R, a code vector y averaged over the entire probability distribution of X is defined as y which minimizes distortion between the point X and R.

[0025]

(Equation 2)

If the strain measure d is a square error strain measure (the square of the Euclidean distance), then cent (R) = E [X | X∈R].

The finite set R = {xi: i = 1, 2,…, || R ||}
In the case of, the definition of centroid in the above equation is still applicable. Here, || R || is the number of elements in the set R. Therefore, if the distortion measure d is a square error distortion measure and each point xi in R has equal probability, the centroid is the arithmetic mean. That is,

[0028]

(Equation 3)

Is as follows.

For a given division {Ri; i = 1,..., N}, the optimal codebook is a codebook satisfying the following centroid condition.

Yi = cent (Ri)

An iterative codebook improvement algorithm that converges to a codebook that simultaneously satisfies the nearest neighbor condition and the centroid condition, which are necessary conditions for optimality, is as follows. (1) Start with the initial codebook C1. Set m = 1. (2) For the given codebook Cm, find the optimal partition using the nearest neighbor condition. (3) For the given division, find the optimal codebook Cm + 1 using the centroid condition. (4) Calculate the average strain for Cm + 1. If the change in strain is sufficiently small, the process ends. Otherwise, m + 1
→ As m, go to (2).

[0033]

Embodiments of the present invention will be described below with reference to the drawings. Here, as one embodiment of the present invention, a region division of a texture image using a texture feature will be described.

FIG. 2 is a block diagram of the hardware system according to the first embodiment of the present invention. This hardware system includes an input device, an output device, a processing device, and a storage device. As an input device, a keyboard 20
1 is coupled to the system. The keyboard 201 is used for inputting parameters, starting commands, and the like. As an output device, a display 206 is coupled to the system. The display 206 is used to display texture images, texture group classification results, region division results, and the like. The processing device 202 includes a module 203 for generating a density histogram, a module 204 for identifying and classifying a texture, and a module 20 for dividing into regions.
5 is included. The storage device 207 stores the texture image 208
And the density histogram 209 generated by the module 203
And the classification result 2 of the texture generated by the module 204
10 and the result 21 of the area division generated by the module 205
1 and the like are stored.

Next, main symbols used in the first embodiment will be described. FIG. 3 shows an example in which the image img is divided into blocks. Each block (each rectangle in FIG. 3) has a size of bSize * bSize in units of pixels. Here, "*" indicates a product. The blocks are numbered sequentially from left to right, starting at 0, from top to bottom. In FIG. 3, i = 0 to 15. Horizontal coordinate X of block i in block units
And the vertical coordinate Y have the following relationship.

X = i% Col Y = i / Col

Here, "%" indicates the remainder of the division, "/" indicates the quotient of the division, and Col indicates the number of columns in the block. If Row is the number of blocks, the number of blocks included in the image
Becomes Size = Col * Row. f [i] [k] indicates the number of pixels of density k in the density histogram in block i. As defined in Equation 1 above, Dist (i, j) indicates the distance between block i and block j. When the value of group [i] is a positive integer, it indicates the number of the texture group to which block i belongs. When the value is 0, the texture group to which the block i belongs is undecided, or the block i is separated from the other blocks by a predetermined distance or more and cannot be grouped, and becomes a single area alone. To indicate that The texture group is for classifying two or more blocks within a distance smaller than a predetermined threshold into one group.

FIG. 4 shows the density histogram f of the block i.
5 is a flowchart illustrating an algorithm for generating [i] [k]. First, in steps 401 to 404, 0 is substituted for f [i] [k] (k is 0 to N) to perform initialization.
Note that (expression) ++ is equivalent to (expression) = (expression) +1. Thus, k ++ is equivalent to k = k + 1. Next, step 4
In steps 05 and 406, the lower limit x0 and the upper limit x1 of the horizontal axis of the block i are set in pixel units. Similarly, steps 407 and 408
Then, the lower limit y0 and the upper limit y1 on the vertical axis of the block i are set in pixel units. Finally, in steps 409 to 415, the density img [y] [x] is obtained for all the pixels (x, y) in the block i, and the corresponding density histogram f [i] [img [y ] [x]] is incremented by one. As described above, the density histogram f [i] [k] of the block i is generated.

FIG. 5 shows the distance between block i and block j.
It is a flowchart showing the algorithm which calculates Dist (i, j). In step 501, the variable v is set to 0. From step 502 to step 505, the absolute value of f [i] [k] -f [j] [k] is added to the variable v while changing the variable k from 0 to N. (Equation 1) + = (Equation 2) is (Equation 1) = (Equation 1) +
This is equivalent to (Equation 2). Abs is a function that returns an absolute value. From the above, the distance Dist between block i and block j
(i, j) is required.

FIG. 6 is a flowchart showing an algorithm for dividing blocks into texture groups. In this algorithm, the group of the texture is undecided, and a set of blocks i and j satisfying i <j is extracted, and the identity of the texture is determined.

First, steps 601 to 604
Group variable while changing the variable i from 0 to Size-1.
Initialize the value of [i] to 0. Next, in step 605, the value of a variable gnum indicating the number of the texture group is initialized to 1. Steps 606, 607, and 620 form a loop for the variable i. At step 606, the value of the variable i is set to 0. In step 607, it is determined whether the variable i is smaller than Size-1. If it is not smaller, the process exits the loop for the variable i and ends the processing. If smaller, block i
The process proceeds to step 608 in order to perform grouping on.

In step 608, it is determined whether group [i] is 0. If it is not 0, it means that the texture group number group [i] to which the block i belongs has been determined, and the process proceeds to step 620. If it is 0, it means that the group to which the block i belongs has not been determined yet, and the process proceeds to step 609 to determine the group. In step 609, the value of the variable flag is set to 0.

Steps 610, 611, and 617 form a loop for variable j. At step 610, the value of the variable j is set to i + 1. At step 611, it is determined whether the variable j is smaller than Size. If it is smaller, it is determined whether group [j] is 0. If the value is not 0, it means that the group number has already been determined for the block j, and the process proceeds to step 617. In step 617, the value of the variable j is increased by one, and the process returns to step 611, which is the head of the loop related to the variable j. Step 61
If group [j] is 0 in step 2, the distance Dist (i, j) between block i and block j is calculated in step 613, and the threshold T
Determine if smaller than If not, step 61
Go to 7. If it is smaller, block i and block j can be in the same group.
At 16, the value of the flag is set to 1, and gnum is assigned to group [i] and group [j].
Is assigned. As a result, the same group number gnum is set for the block i and the block j, and the blocks are grouped. The flag is a flag indicating that writing has been made to group [i] and group [j]. After step 616, proceed to step 617.

If it is determined in step 611 that the variable j is not smaller than the size, the process exits the loop for the variable j, and in step 618, it is determined whether or not flag is 1 If 1, block i and some other blocks are group numbers g
Since it is grouped by num, that is, the group number gnum is used, the value of gnum is increased by one in step 619, and the process proceeds to step 620. Flag flag
If not 1, proceed directly to step 620. Step 6
At 20, the value of the variable i is increased by one, and the process returns to the step 607 which is the head of the loop for the variable i.

FIG. 7 is a flowchart showing an algorithm for dividing an image into regions having a uniform texture. Here, processing of searching for adjacent blocks in the same group and classifying them as one connected area is performed.

First, in step 701, a variable rnum representing a number (area number) for specifying an area is initialized to 1.
Next, in step 702, a variable i representing a block number
Is initially set to 0. In step 703, the group number group [i] of the block i is equal to -1 and the variable i is equal to Size
Determine if smaller than Here, the value -1 indicates that the block i has already been processed (Step 707 described later,
716). If the decision result in the step 703 is true, the process proceeds to a step 704, in which the value of the variable i is increased by one, and the process returns to the step 703. If the decision result in the step 703 is not true, in a step 705, the variable i is set to Size
Is determined to be equal to If they are equal, all the blocks have been scanned, and the process ends.

If not equal in step 705, it is determined in step 706 whether group [i] is equal to zero. Here, the value 0 indicates that the block i is a single area. If they are equal, array
Register with egion. More specifically, in order to record that block i has been processed, in step 707, group
Substitute [-1] for -1. Next, in step 708, region [i]
Substitute rnum for The value of region [i] is the number of the region to which block i belongs. Finally, at step 709, rnum
Is incremented by one, and the process returns to step 703.

In step 706, if group [i] is not equal to 0, the "connected component" process in step 710
An area including the block i is extracted and extracted as an array w1.
Although the details of the “connected component” process will be described later, the block i and the block of the same group adjacent thereto are taken out as an area into the array w1. Specifically, the array w1 [j] takes a binary value of 0 or 1, and when 0, indicates that the block j does not belong to the area including the block i. , Respectively. After step 710, processing for registering the extracted region in the array region is performed. First, in step 711, the value of the variable j is set to 0. Next, step 7
At 12, it is determined whether the variable j is smaller than Size. If it is smaller, it is determined in step 715 whether w1 [j] is equal to one.
If they are equal, it means that the block j is included in the area including the block i. Therefore, in order to record that the block j has been processed,
Substitute [j] with -1. Next, in step 717, rnum is substituted into region [j] to indicate that the block j is a block included in the region with the region number rnum. Next, the process proceeds to step 713. Also, in step 715,
If not equal, proceed directly to step 713. In step 713, the value of the variable j is increased by one,
Return to 12. In step 712, if the variable j is not smaller than Size, the process proceeds to step 714, the value of rnum is increased by one, and the process returns to step 703.

FIG. 8 shows the details of the "connected component" processing in step 710 of FIG. That is, it is a flowchart showing an algorithm for extracting a region including a block i. First, in steps 801 to 805, 0 is assigned to the elements of the arrays w1 and w2 to initialize them. Next, in step 806, the value of w1 [i] is set to 1. Further, at step 807, 1 is substituted for a variable cnt1. Variable cnt1
Indicates the number of elements of the array w1 having the value 1. Next, in step 808, the value of group [i] is substituted for a variable gnum indicating a group number.

From step 809 to step 821, for a block in which the value of the element of the array w1 is 1, the elements of the array w2 corresponding to a total of five blocks (upper, lower, left and right) when viewed as a two-dimensional array with itself. Set the value to 1. That is, a region obtained by expanding the region of the array w1 by one block around the region is generated in the array w2. For this purpose, first, in step 809, the value of the variable j is set to 0. next,
At step 810, it is determined whether the variable j is smaller than Size.
If not, go to step 822. If small
In step 812, it is determined whether w1 [j] is 1. If not 1, the process proceeds to step 811. If it is 1, the process proceeds to step 813, and the value of w2 [j] is set to 1. Step 814
When X (the definition of which will be described later) is not 0, the value of the left element w2 [j−1] is set to 1 in step 815. In step 816, if X is not Col-1, then in step 817,
Set the value of the right element w2 [j + 1] to 1. At step 818,
If Y (the definition of which will be described later) is not 0, step 81
In step 9, the value of the above element w2 [j-Col] is set to 1. Step 8
If Y is not Row-1 in step 20, the value of the lower element w2 [j + Col] is set to 1 in step 821. Here, X and Y are defined by j% Col and j / Col, respectively, and represent the abscissa and ordinate of block j. After the above processing, the flow advances to step 811 to increase the value of the variable j by one, and
Return to 10.

In steps 822 to 828, the block i is extracted from the inflated area in the array w2.
Remove elements that have a different texture from the texture of the. First, in step 822, the variable cnt2 is set to 0. At step 823, the variable j is set to 0. At step 824, it is determined whether the variable j is smaller than Size. If so, in step 826, block j has the same texture as block i (ie, the same group number),
Also, it is determined whether w2 [j] is 1. If true, step 827 increments the value of variable cnt2 by one. If not true, step 828 sets w2 [j] to zero. That is, block j is deleted from the area. In either case,
Proceeding to step 825, increment the value of the variable j by one, and return to step 824. In step 824, if the variable j is not smaller than Size, the process proceeds to step 829. Where cn
t2 is the number of elements of the array w2 having the value 1. At step 829, it is determined whether or not cnt1 is equal to cnt2, and if they are equal, the processing is terminated. If not, step 32
Go to 0.

In steps 830 to 834, array w2 is copied to array w1. In step 830, cn
Substitute the value of cnt2 for t1. In step 831, the variable j is set to 0
Set to. In step 832, it is determined whether the variable j is smaller than Size. If smaller, w1 [j] is set to w in step 834.
2 Substitute the value of [j]. Next, in step 833, the variable j
Is incremented by one, and the process returns to step 832. Step 8
32, if the variable j is not smaller than Size, step 8
Go to 09 and inflate the area again.

FIG. 9 shows an aerial photograph image which is a processing target example of the system of the first embodiment. FIG.
2 shows an example in which the image of FIG. 1 is analyzed by the system of the first embodiment. The portion surrounded by the rectangle is the recognized area.

According to the first embodiment, good results can be obtained for a black-and-white image. However, as described in the section of the problem to be solved by the invention, in a color image, since the number of colors becomes enormous, it is difficult to directly apply the above-described texture analysis method using the density histogram. Therefore, in the second embodiment described below, the number of colors is reduced by using a vector quantization technique.

FIG. 11 is a block diagram of a hardware system according to the second embodiment of the present invention. This hardware system includes an input device, an output device, a processing device, and a storage device. Keyboard 1 as an input device
101 is coupled to the system. Keyboard 1101
Is used for inputting parameters and starting commands. As an output device, a display 1107 is coupled to the system. The display 1107 is used for displaying a texture image, a texture group classification result, a region division result, and the like. The processing device 1102 includes a module 1103 for compressing (reducing) the number of colors, a module 1104 for generating a density histogram, a module 1105 for identifying and classifying textures, and a module 1106 for dividing into regions. The storage device 1108 is
Texture image 1109, density histogram 1110 generated by module 1104, classification result 1111 of texture generated by module 1105, and module 11
06, the result 1112 of the area division generated and the like are stored.

FIG. 12 is a flowchart showing details of the color compression module 1103 of FIG. Init of step 1201 generates an initial codebook y. Details of the init process will be described with reference to FIG. Init of step 1202
R sets an initial value NULL to the pointer array R [i]. The neighbor of step 1203 divides the set of pixels into cells Ri based on the codebook y according to the nearest neighbor condition. Cell Ri
Are inserted into the linear list. The pointer R [i] points to the head of the linear list. Step 12
The centroid of 04 generates a codebook y [i] from the cell Ri according to the centroid condition. Distorti of step 1205
on calculates the distortion of the codebook y. The result is the variable D1
Is assigned to

Steps 1206 to 1212 are the repetitive part of the codebook improvement algorithm. Step 1206 releases the linear list pointed by the pointer R [i]. Steps 1207 to 1209 improve the codebook y by the same processing as steps 1202 to 1204. In step 1210, the previous strain D
Assign 1 to the variable D0. In step 1211, the distortion of the newly generated codebook y is calculated. The result is the variable D
Assigned to 1. In step 1212, it is determined whether or not the rate of decrease in distortion (D0-D1) / D0 exceeds a certain threshold.
If so, the process returns to step 1206 and the process is repeated. If not, the value of each pixel of the color image is quantized in step 1213 and the pointer is
Release the linear list pointed to by R [i].

FIG. 13 is a flowchart showing details of init for generating the initial codebook y in FIG. The variable i is an index in the codebook of the color that is currently registered in the codebook y. First, in step 1301, a variable i is initialized to 0. Next, in step 1302, it is determined whether the variable i is smaller than the quantizer size nCell (nCell represents the number of cells, that is, the number of colors resulting from the quantization). If not, the process ends. If smaller, in steps 1303 to 1307,
A pixel img [x] of the color image is randomly selected by a random number x, and its value is substituted into an array variable clr indicating a color. Here, the variable j is an index representing a color component (red, green, blue) of the color. Note that img [x] is the image to be processed img
Represents one pixel in, and img [x] [j] indicates that pixel img [x]
Represents the value of the color component j.

Next, from step 1308 to step 13
At 16, it is checked whether the color clr has already been registered in the codebook y. The result is indicated by the value of flag. If the flag is 0, it indicates that it has not been registered, and if the flag is 1, it indicates that it has been registered. Specifically, the value of the variable k is set to 0
, While incrementing until k <i (steps 1309, 1310, 1315).
If clr [j] = y [k] [j], the flag is set to 1 (steps 1311 to 1316). Note that y [k]
[j] is an array indicating colors registered in the codebook y, k represents an index in the codebook, and j represents a color component.

If the flag is 0, the process proceeds from step 1317 to 1
Proceeding to step 318, steps 1318 to 1321
Registers the color clr in the codebook y, increments the value of the variable i by one in step 1322, returns to step 1302, and repeats the processing. If the flag is 1, the process returns to step 1303 without registering the color clr and repeats the process.

FIG. 14 shows a set of pixels in FIG.
9 is a flowchart showing details of a neighbor to be divided into the neighbors. In step 1401, a pixel index x is initialized to 0. In step 1402, it is determined whether the index x is smaller than the size of the image. If not small,
The process ends. If it is smaller, in steps 1403 to 1411, a value of i that minimizes the square error distortion measure dist between the pixel img [x] and the color y [i] registered in the codebook is obtained. Assign to the variable min_i. In addition,
In steps 1403 and 1407, dist (A, B) represents the distance between ABs. More specifically, the difference between A and B for each color component is calculated, and the sum of the squares of the differences is calculated.

Next, from step 1412 to step 14
At 15, a record including the value of the index x is inserted at the head of the linear list pointed to by the pointer R [min_i]. This record is index x and pointer n to the next record
This is a structure composed of ext. Finally, step 1
In step 416, increment the value of the index x by one,
Returning to step 402, the process is repeated. As described above, the cells can be divided so that each pixel of the image to be processed is included in any of the cells.

FIG. 15 is a flowchart showing details of the centroid for obtaining the centroid of the cell Ri in FIG. In step 1501, a variable i is initialized to 0. In step 1502, it is determined whether the variable i is smaller than the number of cells nCell. If not, the process ends. If it is smaller, in steps 1503 to 1506, the array su
Initialize each element of m to 0. The variable j is an index corresponding to the color component of the color, and nColor is the number of the color component. In steps 1507 to 1515, the cell Ri
Is obtained, and the sum of the pixels as a vector is substituted into an array sum. Step 1508 indicates that the pointer R [i] to the first record in the list of the cell Ri is set to p. Step 151
3 is the pixel at index x of the record pointed to by pointer p
This is a process of adding the value of the color component j of img [p → x] to sum [j]. Step 1515 is a process of setting the pointer next to the record next to the record indicated by the pointer p to p. In steps 1516 to 1519, the code book
The value obtained by dividing sum by k is substituted into the sign vector y [i] of y.
As a result, the average value of each color component is obtained, and the codebook y
The color registered in is rewritten. Finally, in step 1520, the value of the variable i is increased by one, and
Returning to 502, the process is repeated.

FIG. 16 is a flowchart showing details of distortion for calculating the distortion of the codebook y in FIG. First, in step 1601, the variable sum is initialized to 0. This flowchart includes a triple loop.
Steps 1602 to 1612 constitute a loop for the index i of the cell Ri. This loop is a loop for incrementing i from 0 to the number of cells nCell, in which a pointer R [i] indicating the first record of each cell is set to p (step 1604),
The processing is passed to the next loop (from step 1605). Steps 1604 to 1611 constitute a loop for tracing the linear list of pixels included in the cell Ri. That is, the next record pointer next of the record pointed by the pointer p is newly set to the pointer p, and the processing is looped until the last pixel of the linear list. Steps 1606 to 1610
Constitutes a loop for the index j of the color component of the color. That is, for each color component of the pixel to be processed, in step 1608, the j component img [p-> x] [j] of the pixel included in the cell Ri and the cell R registered in the codebook
The difference v of the j component y [i] [j] of the i color is calculated, and step 16 is performed.
At 09, the square of v is added to the variable sum. Finally, the variable
Returns sum value as distortion value.

FIG. 17 is a flowchart showing details of quantization for quantizing the pixels of the image in FIG.
This flowchart includes the same triple loop as in FIG. In step 1707, the cell R registered in the codebook is stored in the j component img [p-> x] [j] of the pixel included in the cell Ri.
The value of the j component y [i] [j] of the i color is substituted. As a result, the value of the color component j of all the pixels included in the cell Ri becomes y
It is rewritten as [i] [j], which means that quantization has been performed.

After quantization, in the same manner as in the first embodiment, (1) a density histogram is generated for each block of a predetermined size using the quantized pixel values. (2) For the two blocks, obtain the sum of the absolute values of the differences in the respective densities of the density histograms, determine the distance between the two blocks, and (3) determine whether the calculated distance is smaller than a predetermined threshold. In the above, the identity of the texture of the two blocks is determined, (4) based on the determination result of the identity, the texture of each block of the input image is classified into several groups, (5) the same Blocks that are classified into texture groups and that are adjacent can be extracted as one connected area. This allows
Even if the number of colors is huge, reduce the number of colors,
Texture analysis using the density histogram as in the first embodiment is possible.

[0067]

As described above, according to the present invention,
For two textures, it is determined whether or not the sum of the absolute values of the differences between the densities of the density histograms is smaller than a predetermined threshold. In addition, there is an advantage that the amount of calculation is small. Therefore, efficient area division of the texture image can be realized. Also, even if the analysis target is a color image,
By using vector quantization to reduce the number of colors,
The texture analysis method using the density histogram can be applied to a color image.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating characteristics of texture analysis according to the present invention.

FIG. 2 is a block diagram of a hardware system according to the first embodiment of this invention;

FIG. 3 is a diagram showing how an image is divided into blocks and how block numbers are assigned;

FIG. 4 is a flowchart illustrating an algorithm for generating a density histogram f [i] [k] of a block i.

FIG. 5 is a flowchart illustrating an algorithm for calculating a distance Dist (i, j) between a block i and a block j.

FIG. 6 is a flowchart illustrating an algorithm for dividing blocks into texture groups.

FIG. 7 is a flowchart illustrating an algorithm for dividing an image into regions having uniform textures.

FIG. 8 is a flowchart showing a detailed algorithm of step 710 “connected component” in FIG. 7;

FIG. 9 is a view showing an example of an image of an aerial photograph to be processed by the system according to the embodiment;

FIG. 10 is a diagram showing an example in which the image of FIG. 9 is analyzed by the system of the present embodiment.

FIG. 11 is a block diagram of a hardware system according to a second embodiment of the present invention;

FIG. 12 is a flowchart illustrating an algorithm for compressing the number of colors.

FIG. 13 is a flowchart showing an algorithm for generating an initial codebook y.

FIG. 14 is a flowchart illustrating an algorithm for dividing a set of pixels into cells Ri.

FIG. 15 is a flowchart showing an algorithm for obtaining a centroid of a cell Ri.

FIG. 16 is a flowchart showing an algorithm for calculating the distortion of the codebook y.

FIG. 17 is a flowchart illustrating an algorithm for quantizing pixels of an image.

[Explanation of symbols]

201, 1101 Keyboard, 202, 1102 Processing device, 203, 1104 Module for generating density histogram, 204, 1105 Module for identifying and classifying texture, 205, 1106 Module for dividing into regions, 206, 1107 ... display, 2
07, 1108: Storage device, 1103: Module for compressing the number of colors.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Takashi Onoyama 6-81 Onoecho, Naka-ku, Yokohama-shi, Kanagawa Prefecture In-house Hitachi Software Engineering Co., Ltd. 81 Hitachi Software Engineering Co., Ltd. In-house F-term (reference) 5L096 AA02 AA03 AA06 BA06 BA08 FA01 FA37 FA39 FA41 GA19 JA03

Claims (4)

[Claims] 1. A texture analysis method for analyzing a texture of an input image, comprising the steps of: determining whether a sum of absolute values of differences between respective densities of density histograms of two textures is smaller than a predetermined threshold value; A texture analysis method characterized by determining the identity of a texture. 2. A texture analysis method for analyzing a texture of an input image, comprising: a density histogram generation step of subdividing the input image into blocks of a predetermined size and generating a density histogram for each block; Calculating the sum of the absolute values of the differences between the densities of the density histograms, and defining the distance as the distance between the two blocks; and determining whether the calculated distance is smaller than a predetermined threshold. An identity determination step of determining the identity of the textures of the two blocks; a texture identification step of classifying the texture of each block of the input image into several groups based on the determination result of the identity; Area that is classified as a group and that extracts adjacent blocks as one connected area Texture analysis method characterized by comprising a division step. 3. The texture analysis method according to claim 1, wherein when the input image is a color image, the density of each pixel of the color image is regarded as one vector, and the color component of the density is a vector component. A texture analysis method, wherein the number of colors is reduced by performing vector quantization in correspondence with the above, and the density histogram is generated using the quantized result. 4. A texture analysis program according to a texture analysis method according to claim 1.