JPS629957B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a method for identifying digitized multispectral images of microscopic stained microscopic objects by color. Microscopic objects are dyed so that they can be easily identified by the human eye, and their microscopic images are counted. Attempts have also been made to count each type by classifying them by size. Figure 1 is an example of a microscopic image stained red and blue with Proca staining solution. In stained images of minute objects, the color density is uniform, the center is dark,
A cluster with pale peripheral areas is formed. In Figure 1, the shaded area is the dark colored area 11,
Groups 12, 14, and 15 are dyed darkly in the center and lightly dyed in the periphery, group 13 is dyed lightly overall, and the colors of each group are as shown in Table 1.

[Table] When identifying Figure 1 with the human eye, the groups stained red, those stained blue, and the groups stained pale red are labeled (A), (B), and (C), respectively. The results shown in Figure 4 can be obtained by classifying the results. The groups in FIG. 4 are identified as shown in Table 2. It is desired that the automatically identified result is also similar to that shown in FIG.

[Table] An example of a method conventionally used to do this is described below. The microscopic image in FIG. 1 is divided into sufficiently fine two-dimensional grids, and the primary color components of brightness (for example, the three primary color components of red, green, and blue) at each grid point (pixel) are digitized to produce multispectral image data. Next, the brightness and chromaticity values of each pixel are calculated using equation (1). (R, G, B are the three primary color components, a ₁ , a ₂ , a ₃ are appropriate constants, l is the brightness, and r, g are the chromaticity.) The color of each pixel is determined using l, r, and g as the coordinate axes. is expressed as a point in Euclidean space. (In the following text, this will be referred to as a color space.) The color distribution of the multispectral image data is represented by a three-dimensional histogram in which each point in the color space has the number of pixels that have the color corresponding to that point. I can do it. FIG. 2 is a schematic diagram of a three-dimensional histogram with coordinate axes of l, r, and g, similar to the color space for the image in FIG. 1. Specifically, the pixels in the central part dyed red in group 11, etc. are near Figure 2 and 3, and in group 1 dyed blue.
The pixels in the central part of 2 are distributed around 4, and the peripheral parts of groups 11, 12, etc., and the pixels of group 13 are distributed around 5. In order to identify the microscopic image shown in FIG. 1 in the prior art, regions are divided while taking into account the three-dimensional histogram shown in FIG. 2, and pixels existing in each region are classified into classes. In the example in Figure 3, 21, 22, 2
3 and 24 are areas set to identify red, blue, light red, and light blue, respectively, and the classes are respectively
(a), (b), (c), and (d). First, if we classify by region 21, we get Figure 5a-111, 114, 115, and if we classify by region 22, we get Figure 5b-
112 is obtained. Next, when classifying by area 23, Fig. 5 c-113 and group 11 with the central part missing.
11, 1114 is obtained, and when classified by region 24, d1112 in FIG. 5 is obtained. The identification result is shown in Table 3, which is shown in FIG. 5e by superimposing FIGS.

[Table] Comparing this result with Figure 4, Figure 11, 1
When there is a difference in color density between the center and the periphery, as shown in numbers 2, 14, and 15, the classes are separated between the inside and outside of the density boundary, and the groups that touch, as shown in numbers 14 and 15, , two drawbacks can be found in the recognition of one group. In an attempt to eliminate the former drawback, a large area including 21 and 23 in FIG. 3, that is, an area where red and light red can be extracted, and 22, 24
Even if an area containing blue and light blue is set and identified at the same time, the result will be as shown in Figure 6. Table 4 clearly shows that the classes are divided based on density differences. However, the object class in FIG. 13 is different from that in FIG. 4.

[Table] In the end, identification by dividing the color space means that even if the size and position of the region are varied in the color space, the peripheral part of the group mentioned earlier Due to both the disadvantages of not extracting the image and not being able to separate groups that are in contact, it is not possible to obtain a result equivalent to the discrimination result obtained by the human eye in FIG. 4. An object of the present invention is to eliminate such conventional drawbacks and to provide a method for obtaining identification results similar to those obtained by the human eye using the colors and positional relationships of stained images. The present invention selects a processing subspace in a three-dimensional histogram, extracts a histogram part (region) with a value larger than a specified threshold value, performs labeling, and detects adjacency with the region in the immediately preceding processing subspace. The extraction control process includes an extraction control process, an extraction/labeling process that extracts pixels of an image corresponding to the area and performs labeling, and an expansion process that expands the result of the extraction/labeling process. While sequentially moving the processing subspace of the extraction/labeling process to subspaces with higher brightness, the extraction results are superimposed and the above overlap is performed using the adjacency status between regions in adjacent subspaces detected by the extraction control process. A color image classification method is obtained that performs expansion processing on the combined results to extract the shape of the object and identify the class. The principle of the present invention will be explained below. In the case of a stained image of something like a microorganism in which minute objects that have absorbed dye can be seen on a bright background, the degree of dye absorption is large in the center and small in the periphery. Therefore, when each microscopic object is observed by a human using a microscope, even if the microscopic objects are in contact at the periphery, the central portions are separated and can be distinguished. The present invention takes advantage of this feature. In other words, in the example shown in Figure 1, the three-dimensional histogram of the color space is
Although the distribution is as shown in the figure, for example, part 4 is a blue part with low brightness and corresponds to the central pixel of the blue minute object 12 in FIG. Similarly, the part 3 is the blue part with low brightness and is the first
This corresponds to the pixels at the center of the red minute objects 11, 14, and 15 in the figure. On the other hand, the part 5 is a light colored part and corresponds to the blue minute object 12, the peripheral parts of the red minute objects 11, 14, and 15, and the entire pale minute object 13, and it is difficult to classify it using color information alone. This is the part that cannot be done. Therefore, the positional relationship in the digitized image must be used to classify the light-colored portions. In other words, the part that exists around the red minute object is part of class (a), the part that exists around the blue minute object is part of class (b), and the part that exists independently without either is part of class (a). It is part of class (c). In order to perform such recognition, it is sufficient to first extract the central part of the object, and then perform processing to determine the correct shape of the object by propagating it to the surroundings. Furthermore, as mentioned earlier, it is known that the central part of an object is dyed darkly and has low brightness, so extraction of the central part can be achieved by detecting and classifying pixels with low brightness. . Specifically, this is done by dividing the three-dimensional histogram of FIG. 2 into layers of different brightness as shown in FIG. 7, and tracing the portions having values from the low brightness layer. To explain this in detail, Figure 7 shows a three-dimensional histogram divided into three layers with different brightness, and Figure 8 shows each layer separated for easy understanding. The results are shown in Figure 9. First, we will process the layer shown in Figure 8, 31, and the high histogram values of this layer are 41, 4.
Part 2 shows the chromaticity regions corresponding to red pixels and blue pixels, respectively. Chromaticity region 4 in a low brightness layer like 31
The pixels corresponding to 1 and 42 are pixels in the center of the microscope image where the dye is concentrated. Classes (a) and (b) are assigned to regions 41 and 42, the corresponding pixels of the microscope image are extracted, and the results of classification are shown in FIG. 9a and Table 5.

[Table] Next, 32 layers are processed. When pixels corresponding to areas 43 and 44 in the three-dimensional histogram of layer 32 are extracted, the result is shown in FIG. 9b. Area 4
3, groups 131, 134, 135, area 4
Group 132 corresponds to number 4. When classes (c) and (d) are assigned to areas 43 and 44, FIG. 9b is divided into classes as shown in Table 6.

[Table] Here, in order to detect the positional relationship in the image, the identification result in layer 31 (FIG. 9a) and the identification result in layer 32 are superimposed. Figure 9b shows the result. At the same time, regions 43 and 44 extracted in the histogram are adjacent to region 4 in layer 31.
1 and 42 are also detected. Subsequently, 33 layers are processed. When a class (e) is given to the region 45 of the three-dimensional histogram and the corresponding pixels are extracted, the result is FIG. 9c, which is superimposed with FIG. 9d to obtain FIG. 9e. Also area 45
are dissociated. Adjacent to layers 43 and 44 of layer 32 is also detected. FIG. 9e shows the results obtained in layers 31 to 33, which are identified as shown in Table 7 by the classes given to the regions in each layer.

[Table] Here, expansion processing is performed using the positional relationship of the image and the positional relationship of the areas extracted in the histogram of the adjacent layer. For example, the group 141 is identified by classes (A), (C), and (E), and the adjacency of regions 41, 43, and 45 in these three-dimensional histograms is detected. Since the areas 45 and 43 are adjacent, and the areas 43 and 41 are also adjacent, the class (A) for the central pixel is expanded to the periphery. Groups 142, 144,
Group 145 is expanded in the same way, and group 143 is not expanded since there is no corresponding chromaticity region in the previous layer, and is identified by class (e) as it is. The results are shown in Figure 9f and each group is identified as shown in Table 8.

[Table] In this way, by using the positional relationship of the extracted areas and the positional relationship of the extracted groups in the three-dimensional histogram and processing from the lowest to highest brightness layer, it is possible to classify the color information alone. Classification can be performed in areas that cannot be detected, and it is possible to extract even the periphery of a dyed image with different color densities in the center and periphery. Also, if they are in contact at the periphery, they can be separated. An embodiment using the color image identification method of the present invention will be shown and explained. Similar to FIG. 7, FIG. 10 is a diagram in which the three-dimensional histogram is divided into 61 to 66 layers, and each layer is hereinafter referred to as a subspace. The operation of the color image classification device using the method of the present invention is explained with reference to the block diagram of FIGS. 11a to 11c and the conceptual diagrams of the histograms of each subspace and the classification results in each subspace, etc., to FIGS. 12 to 21. I will explain. In addition, 1100-1 in Figure 11 a, b, c
1100-23 is a general OR circuit, 110
1-1 and 1101-2 are general AND circuits and are well known, so they will not be particularly explained in the text. A control section 999 in FIG. 11a controls the entire apparatus. This control unit outputs signals indicating brightness, chromaticity, histogram calculation mode, extraction control mode, extraction, labeling mode, and expansion processing mode, respectively.
s ₁ , s ₂ , s ₃ , s ₄ , specifying the address p ₁ , p ₂ ,
p ₃ , p ₄ specifying brightness and auxiliary signals s ₂₀ , s ₂₁ ,
s ₂₂ , s ₂₃ , s ₃₀ , s ₃₁ , s ₃₂ , s ₄₀ are issued. First, a three-dimensional histogram is created by counting the brightness, chromaticity, and number of pixels with the same brightness and chromaticity for each pixel of the input image.
Histogram calculation mode is executed. When the analog signal 499 of R, G, and B components of a pixel is input to the A/D converter 1000 in FIG. Arithmetic unit 1001
is a signal 501 consisting of brightness l, chromaticity r, and g calculated according to equation (1) on the X axis in the horizontal direction with respect to the image.
The image is stored in the brightness/chromaticity image memory 1002 at an address 498 based on the coordinate values (X, Y) of each pixel when the Y axis is taken in the vertical direction and each pixel is expressed in two-dimensional coordinates. At the same time, the brightness and chromaticity signals 501
Histogram memory 1003 as address
The adder 1004 adds “1” to the result.
A histogram is created by adding the value to the histogram memory 1003 at the same address. The results obtained in this way are distributed as shown in the conceptual diagram shown in FIG.
It will be stored in 003. The area surrounded by the curve indicates the area where the cumulative value is large. Chromaticity obtained from the histogram as shown in FIG. 10 stored in the histogram memory 1003 according to the address signal p ₂ and the brightness signal p ₃ that changes with the brightness width of the selected subspace according to the control mode s ₂ The histogram value of the lattice of the subspace corresponding to the brightness is read out. The histogram values of the respective subspaces 61 to 65 in FIG. 10 that are read out are shown as conceptual diagrams as shown in FIGS. 12 a to 12, for example. The extraction control mode will be explained using the diagram in FIG. 11b. First, subspace 61 in Fig. 10 is specified,
The histogram values corresponding to the subspace 61 shown in FIG. 12a are read out. The comparator 1008 outputs "1" if the threshold value T (10 as an example) is greater than or equal to the threshold value T, and outputs "0" if it is less than the threshold value T (here, 10 as an example). Histogram value of subspace 61 Fig. 12a
Since there is no histogram value of 10 or more and therefore no corresponding pixel, the control unit 999 uses the extraction control mode signal s ₂ and the address signal p ₂ , to read out the histogram in the next subspace 62.
Send p ₃ . The histogram of the subspace 62 in FIG. 12b is read out to the comparator 1008 and compared with the threshold value "10". The binary pattern of the comparison result "0" or "1" is stored in the histogram binary memory 1009 using the chromaticity signal (r, g) given as the signal _p2 as an address. The histogram binary pattern in the specified subspace obtained in this way is the extraction control mode signal
It is read out according to the readout signal _s20 at _s2 and the address signals of the target point and its surrounding points sequentially generated by the labeling processor 1010. The read binary pattern signal 503 is subjected to labeling processing in order to recognize the concatenated portion of "1" (the method of labeling processing is described in Digital Picture Processing 9, 1, 3, published by Academic Press, 1976). (as shown in detail). The labeling process will be explained below. In the labeling process, when 70 is the target point in Figure 13a, if the target point is "1" as shown in the figure, its surroundings 71 are referred to, and if all are "0", it is set to "0" or " Attach a label other than 1” and other than the previously used label. Here, for example, "102" is assigned to the 70th point. Therefore, the result is shown in FIG. 13b. If the target point is shifted to the right and set to 72, its surroundings 73 are referred to, and if the target point is "1" and there is a label other than "0" or "1" around it as shown in the figure, that label is attached. If the target point is "0", do nothing and shift the position of the target point by one. Proceeding further, if there is a new "1" as the target point and there are no labels other than "0" and "1" around it as described above, "103" is added here as the next label. (In this device, as an example, labels are given in ascending numbers of 102 or more.) When the labeling process is performed on the whole, the result is as shown in FIG. 13c. This labeled part is the chromaticity region of a certain color, and the labels "102" and "103" are the classes given to it. The labeling processor 1010 calculates r and g of the target point and its surrounding points.
Address signals _p21 based on coordinates are sequentially generated and labeling processing is performed. The result is stored in the first histogram labeling result memory 1011 by the address signal _p21 from the labeling processor 1010. When the labeling is completed, the first histogram labeling result memory 101 is read out by the readout signal _s21 .
The labeling results are read from 1 to ₁ in accordance with the address signal p2, and a class table is created in the first glass label memory 1012. Table 9 is an example of the contents of the first class label memory 1012, in which the class label is used as an address, and the lightness l at which it begins to be distributed and its coordinates (r, g) on the chromaticity plane are stored. At the same time, the adjacency state between the part to which a class label was attached in the subspace processed immediately before and the part in the subspace currently being processed is determined, and if they are adjacent, the class label in the subspace immediately before is determined.
If there is no adjacency, "0" is stored.

[Table] The example shown in Table 9 was created for the class label in the subspace 62. First, the labeling result is read out from the first histogram labeling result memory 1011 using the readout signal _s21 according to the address signal _p2 . . In the area value setting circuit 1013, if the label is sent for the first time, the signal _p2 indicating the chromaticity values r and g and the signal _p3 indicating the brightness of the subspace are set as the first class label using the label as an address signal. memory 101
2 is stored. Next, in accordance with the address signal _p2 , the labeling result in the histogram of _the immediately preceding processing subspace is stored in the first histogram labeling result memory 1011 by the signal s22 that starts checking the adjacent state of the class label. The second histogram labeling result memory 1014 is sequentially read and compared by the comparator 1015, and if different class labels are attached at the same address, the class label in the subspace currently being processed is used as the address for the adjacent class in the immediately previous processing subspace. By adding labels, the contents of the first class label memory 1012 are created. However, in this example, the number of adjacent classes to one class is 3 or less. In the subspace 62, the labeling result in the subspace processed immediately before is a label of all "0" because there is no connected deposit area of "1" to be labeled, so when checking the adjacency state, it is a label indicating adjacency. "0" is written as "0" to indicate that there is no adjacency, and the contents of Table 9 are stored in the first class label memory 1012. The contents of the first class level memory 1012 will be required for later expansion processing. After comparing the histogram labeling results, a readout signal _s23 sends the contents of the first histogram labeling result memory 1011 to the second histogram labeling result memory 1014.
The contents of the second histogram labeling result memory 1014 are updated. This operation is performed by the read signal _s23 and address signal _p2 . The labeling processor 1010 also stores the number and maximum value of labels used during execution of the extraction control mode. The above is the operation of extraction control mode _s2 . When the extraction control mode _s2 is completed, the extraction/labeling mode _s3 is selected, and pixels corresponding to the chromaticity regions to which class labels are attached in the histogram are extracted and labeling processing is performed. However, extraction is performed for each chromaticity region,
The labeling processor 1010 extracts the number of class labels stored in the labeling processor 1010, and repeats the labeling mode _s3 . For example, if you try to extract the pixel corresponding to the class label "102", the first
Brightness histogram image memory 1002 in Figure 1a
The lightness 505 and chromaticity 504 of the pixel are sequentially read out in accordance with the address signal _p1 by the readout signal _s3 . First, the comparator 1016 calculates the brightness 505 and the subspace 504.
is read out and compared with the brightness range _p4 , only the pixels distributed in the subspace 62 are extracted, and the contents of the first histogram labeling result memory 1011 are read out using the chromaticity as an address. The comparator 1017 compares the signal 508 indicating the label value output from the labeling processor 1010, here "102", and if the content of the first histogram labeling result memory is "102", "1" becomes "0". ” or if it is a different label, “0” is output, and stored in the binary image memory 1018 according to the same address p ₁ as read from the brightness and chromaticity image memory 1002. When all the pixels in the brightness/chromaticity image memory 1002 are read out, the pixels corresponding to the class label "102" are extracted as shown in FIG. 14a. Next, regarding the contents of the binary image memory 1018, a labeling processor 1019 is applied in the same way as the labeling process for the contents of the histogram binary memory in the extraction control mode _s2 in order to recognize the connected portion of "1". The same processing is performed in , but in this case, the signal _p11 generated by the labeling processor 1019 as an address signal indicates the coordinate value for the image. In addition, labels are assigned in ascending order of numbers greater than or equal to "2" so as not to be mixed with class labels, and the value of the maximum label is assigned to the labeling processor 101.
Manage it with 9. In FIG. 14a, a labeling processor 1019 performs processing according to the readout signal for labeling at s ₃₀ and the result is stored in a labeling result image memory 1020.
Each extracted group and its label in Figure 14a are displayed.
10. When labeling is completed, a readout signal is sent to create a table showing the correspondence between the labels used in this process and the class labels.
At _s31 , the label 509 is read out from the labeling result memory 1020 in accordance with the address signal _p1 . A table as shown in Table 11 is created by the table creator 1021, and the class label signal 506, that is, "102" is assigned to the second class label memory, 10 with the label 509 as the address.
22.

【table】

[Table] Once the table is created, the signal 509 indicating the label is sequentially read out from the labeling result image memory 1020 in accordance with the address signal _p1 using the readout signal _s32 . Label 51 at the same address from the first output image memory 1024 in FIG. 11c
0 is read out, the adder 1023 adds the label signals 509 and 510, and the result is stored in the first output image memory 1024 according to the address. Next, similar processing is performed to extract pixels corresponding to the class label "103". FIG. 14b shows the extraction results. On the other hand, when labeling processing is performed, each group is labeled as shown in Table 12, and the contents of the second class label memory 1022 are as shown in Table 13. Can be added.

[Table] Further, the results shown in FIG. 14b are sequentially added by the adder 1023 in response to the readout signal s ₃₂ , and the first output image memory 1024 is updated as shown in FIG. 15. This completes the processing for the subspace 62,
Next, the process returns to the extraction control mode and performs similar processing for the partial space 63. The histogram in subspace 63 is compared with the threshold "10" in FIG. 12c, and regions 50 and 51 are extracted. Areas 50 and 51 are labeled with class labels "104" and "105", respectively, and the contents of the first class label memory 1012 are updated as shown in Table 14. Furthermore, by comparing the labeling result of the histogram in the subspace 62, that is, the contents of the second histogram labeling memory 1014, it becomes clear that the subspace 62 is adjacent to the labeled portions 48 and 49 of the histogram in FIG. 12b. Therefore, the adjacent class label “102”,
“103” is also written to the first class label memory 1012.

[Table] After that, the extraction/labeling processing mode is executed. The pixel corresponding to 50 in the histogram of FIG. 12c is extracted and labeled as shown in FIG. 16a. Similarly, the pixel 16b corresponding to the histogram 51 is also extracted and labeled.
These are added to the first output image memory 1024, the contents of the first output image memory 1024 are updated as shown in FIG. 17, and each group is labeled as shown in Table 15.

[Table] At this time, the contents of the second class label memory 1022 are also updated as shown in Table 16.

[Table] Next, when the extraction control mode and the extraction/labeling mode are similarly executed for the subspace 64 using its histogram in FIG. 12d, the first output image memory 1024 is as shown in FIG. The label shown in 17 is attached.

[Table] The contents of the first class label memory 1012 and second class label memory 1022 are shown in the table.
18, updated as shown in Table 19.

【table】

[Table] Next, the subspace 64 is selected, and in the extraction control mode, the region 54 having a threshold value of "10" or more is extracted from the histogram (e) in FIG.
Labeled by “108” label. Portions 52 and 53 in the histogram of the subspace 63 that was processed immediately before in FIG. 12d are compared with 54, and since they are adjacent, the first class label memory 1012 is updated as shown in Table 20.

[Table] Next, the extraction/leveling mode starts.
When the pixel corresponding to 54 in Figure 2 e is extracted, the first
The binary pattern shown in FIG. 9 is obtained, labeling processing is performed on it, and the first output image memory 102
It will be added to 4. First output image memory 10
24 is as shown in FIG. Table 21 shows the labels in each group, and Table 22 shows the class labels corresponding to each label, that is, the contents of the second class label memory 1022.

【table】

[Table] Next, the extraction control mode is returned to and the subspace 66 is selected, but the brightness of this subspace 66 is known in advance to be the background, so there is no need to process it. When the extraction and labeling processing of pixels corresponding to the subspace in which pixels other than the background in the three-dimensional histogram are distributed is completed in this way, the process moves to execution of the expansion processing mode _s4 . In the expansion processor 1025 shown in FIG. _11c , a 3×3 mask such as 100 shown in FIG. . Among the labels of pixels read out sequentially, when the label of the pixel at the center of the mask is other than "0", the labels of the surrounding 8 pixels are "0" or labels other than the same label as the center pixel. Only in certain cases is the label signal 515 of the center pixel followed by the signal 516 indicating the maximum value of the labels of the surrounding pixels being read out. The class label signal 512 for the center pixel is read out from the second class label memory 1022 using the label signal 515 as an address,
A signal 513 indicating an adjacent class label is read from the first class label memory 1012 using the class label signal 512 as an address. Next, the class label signal 512 is read out from the second class label memory 1022 using the label signal 516 as an address, and the comparator 1026
are compared. The comparator 1026 generates a signal 518 when the class label signals 512 and 513 match, and a signal 517 when they do not match. Counter 1027 is cleared by mode signal _s4 and counts up each time signal 518 is generated. When the matching signal 518 is generated in the comparator 1026, the label signal 516 is generated, and when the signal 517 is generated, the label signal 515 is generated according to the address signal _p12 indicating the coordinates of the center pixel on the image generated by the expansion processor. 2nd output image memory 1
028. When the mask has scanned all pixels, counter 1
Refer to the value of 027. When the counter 1027 is "1" or more, that is, it indicates that a pixel has been expanded, so the counter 1027 is cleared with the readout signal _s41 , and then the contents of the second output image memory 1028 are sent to the address signal. p ₁ to the first output image memory 1024. When the transfer is completed, the expansion processor 1025 operates again and the same expansion process as described above is performed. This expansion process is repeated until the counter 1027 holds the value "0" when the mask has scanned all pixels. The above expansion process will be explained using FIGS. 21a, b, c, and d. Mask 100 as shown in FIG. 21a
is scanned, and the label "14" of the center pixel and the maximum value "10" of the labels other than "0" and "14" around it are read out. Figure 21b is an example of the contents of the second class label memory 1022, Figure 21c
is an example of the contents of the first class label memory 1012. First second class label memory 10
The class label “108” of “14” is read from the first class label memory 1012, and the adjacent class labels “106” and “107” of the class label “108” are read from the first class label memory 1012. Next, class label “106” of label “10” is read from second class label memory 1022. Since the class labels "106" match, the pixel with the label "14" in the center of the mask is adjacent to the surrounding pixels with the label "10" on the image, and the chromaticity areas are also adjacent. Therefore, the label "10" is expanded to the center pixel as shown in FIG. 21d. This expansion process is a method of expanding the labels of the inner pixels outward one by one. For example, even if the groups 2064 and 2065 in Figure 20 are connected by the outer label "17", in this case, the center labels "3" and "4" will expand, respectively, and no further expansion will occur at the boundary. Since there is no group 2064, 2065
can be separated, and finally the extraction results and table in Figure 22 are obtained.
The classification results shown in 23 are obtained.

[Table] In the embodiment, the above results are obtained from the second output image memory 1028 and the second class label memory 1022, and at the same time, information regarding the chromaticity and brightness of each class, for example, the chromaticity corresponding to the center pixel, Brightness is the first class label memory 1012
obtained from. Further, the embodiment can be constructed with a relatively simple circuit. Note that the brightness range of the processing subspace in the description of the embodiment is arbitrary. However, when the brightness range is increased, in addition to labeling processing for recognizing two-dimensional connected parts, connections in the brightness direction must also be determined. In other words, three-dimensional labeling processing is required. This is similar to two-dimensional labeling processing. Further, the threshold value specified for the three-dimensional histogram is appropriately determined based on the characteristics of the image. Moreover, it can also be changed for each processing subspace.

[Brief explanation of the drawing]

Figure 1 is a diagram showing an example of a microscopic image of a stained minute object, Figure 2 is a three-dimensional histogram for Figure 1, Figure 4 is a diagram to explain the identification results by the human eye, and Figure 3. , FIGS. 5 and 6 are diagrams for explaining identification processing in the conventional technology, and FIG.
9 shows a three-dimensional histogram and the process of classification, FIGS. 11a, b, and c are block diagrams showing an embodiment of the color image classification method of the present invention, and FIGS. FIGS. 12 to 22 illustrate the processing subspace of the identification process for FIG. 1 according to the embodiment, as well as intermediate results of the process and identification results. In the figure, 999 is the control unit, 1000 is the A/
D converter, 1001, brightness chromaticity calculator, 1002
1003 is a brightness/chromaticity image memory, 1003 is a histogram memory, 1004 is an adder, 1008 is a comparator, 1009 is a histogram value memory, 10
10 is a labeling processor, 1011 is a first histogram labeling result memory, 1012 is a first class label memory, 1013 is a region value setting circuit, 1014 is a second histogram labeling result memory, 1015 is a comparator, 1016
is a comparator, 1017 is a comparator, 1018 is a binary image memory, 1019 is a labeling processor, 1
020 is a labeling result image memory, 102
1 is a table creator, 1022 is a second class label memory, 1023 is an adder, 1024 is a first output image memory, 1025 is an expansion processor, 1
026 is a comparator, 1027 is a counter, 102
8 indicates a second output image memory.

Claims (1)

[Claims] 1. Select a processing subspace in a multidimensional histogram using the color of an image of a stained object, extract a histogram portion (region) with a value larger than a specified threshold value, and perform labeling. an extraction control process that detects adjacency with a region in the immediately preceding processing subspace; an extraction/labeling process that extracts pixels of an image corresponding to the region and performs labeling; It consists of an expansion process that expands the extraction control process, extraction and
While sequentially moving the processing subspace of the labeling process to subspaces with higher brightness, the extraction results are superimposed and the above superposition is performed using the adjacency situation between regions in adjacent subspaces detected by the extraction control process. A color image identification method characterized by performing expansion processing on the result to extract the shape of an object and identify its class.