JP7666292B2 - Determination device, determination method, and determination program - Google Patents

Description

The present invention relates to a determination device, a determination method, and a determination program.

There are techniques for determining whether the color of a manufactured product matches a predetermined color sample. For example, Patent Document 1 discloses a technique for determining color by quantifying the color of a two-dimensional printed material or a three-dimensional solid object using the average and frequency values of brightness and hue.

JP 2017-188790 A

However, due to the influence of grain or warping, the apparent color of the object being judged (object to be judged) may differ from its original color. Conventionally, evaluation has been performed using simple statistical indicators (average value, mode, area ratio) as in the technology disclosed in Patent Document 1, but evaluation using simple statistical indicators is less robust.

The present invention has been made in consideration of the above points, and aims to provide a judgment device, judgment method, and judgment program that can judge the color of an object to be judged with high robustness compared to evaluation using simple statistical indicators.

The determination device according to the first aspect of the present invention includes a sampling unit that performs sampling by a bootstrap method on first image data in which a color sample serving as a color reference is captured and second image data in which an object to be determined is captured; a feature generation unit that performs conversion of the sampling results of the first image data and the second image data into M types of color spaces (M is a natural number) to generate a first feature pool and a second feature pool each consisting of M types of features; a spatial variation calculation unit that obtains a spatial variation by calculating a standard deviation for the M types of features in the first feature pool; a difference acquisition unit that obtains K difference features by taking the difference between the features of the first feature pool and the second feature pool using K types (K is a natural number) of distance calculation methods; and a learning unit that performs ensemble learning using a categorical variable corresponding to the determination result of the object to be determined for the K difference features as a target variable and the spatial variation as a loss function.

According to the first aspect of the present invention, by using differential features and spatial variation calculated based on features obtained from color samples and samples, it is possible to determine the color of an object to be determined with high robustness compared to evaluation using simple statistical indicators.

The determination device according to the second aspect of the present invention is the determination device according to the first aspect, and the feature amount is a value representing a color space after conversion of the first image data and the second image data.

According to the second aspect of the present invention, the difference feature amount and spatial variation amount can be obtained using values representing the converted color space.

The determination device according to the third aspect of the present invention is the determination device according to the first or second aspect, and the spatial variation calculation unit multiplies the standard deviation calculated for each type of feature by a predetermined weight and adds the multiplied values to obtain the spatial variation.

According to the third aspect of the present invention, it is possible to quantitatively evaluate the degree of variation in features by weighting the types of features.

In the determination method according to the fourth aspect of the present invention, a computer executes the following processes: sampling first image data in which a color sample serving as a color reference is captured and second image data in which an object to be determined is captured, using a bootstrap method; converting the sampling results of the first image data and the second image data into M types of color spaces (M is a natural number) to generate a first feature pool and a second feature pool each consisting of M types of features; calculating a standard deviation for the M types of features in the first feature pool to obtain a spatial variation; calculating the difference between the features in the first feature pool and the second feature pool using K types of distance calculation methods (K is a natural number) to obtain K difference features; and performing ensemble learning using a categorical variable corresponding to the determination result of the object to be determined for the K difference features as a target variable and the spatial variation as a loss function.

According to the fourth aspect of the present invention, by using differential features and spatial variation calculated based on features obtained from color samples and samples, it is possible to determine the color of an object to be determined with high robustness compared to evaluation using simple statistical indicators.

The judgment program according to the fifth aspect of the present invention causes a computer to execute the following process: sampling first image data in which a color sample serving as a color reference is captured and second image data in which an object to be judged is captured, using a bootstrap method; converting the sampling results of the first image data and the second image data into M types of color spaces (M is a natural number) to generate a first feature pool and a second feature pool each consisting of M types of features; calculating a standard deviation for the M types of features in the first feature pool to obtain a spatial variation; calculating the difference between the features of the first feature pool and the second feature pool using K types of distance calculation methods (K is a natural number) to obtain K difference features; and performing ensemble learning using a categorical variable corresponding to the judgment result of the object to be judged for the K difference features as a target variable and the spatial variation as a loss function.

According to the fifth aspect of the present invention, by using differential features and spatial variation calculated based on features obtained from color samples and samples, it is possible to determine the color of an object to be determined with high robustness compared to evaluation using simple statistical indicators.

According to the present invention, by creating robust judgment criteria from image data, it is possible to provide a judgment device, judgment method, and judgment program that can judge the color of an object to be judged with high robustness compared to evaluation using simple statistical indicators.

1 is a diagram showing a schematic configuration of a determination system including a determination device according to an embodiment of the present invention; FIG. 2 is a block diagram showing a hardware configuration of the determination device. FIG. 2 is a block diagram showing an example of a functional configuration of a determination device. FIG. 13 is a diagram showing an example of sampling by a bootstrap method in a sampling unit. 13A and 13B are diagrams illustrating examples of histograms obtained by hue conversion performed by a feature generating unit. FIG. 13 is a diagram illustrating the amount of spatial variation. 11 is a diagram illustrating an example of a method for calculating a difference feature amount by a difference acquisition unit. FIG. 2 is a diagram illustrating the GBDT used by the determination device. 4 is a flowchart showing a flow of a determination process performed by the determination device.

Below, an example of an embodiment of the present invention will be described with reference to the drawings. Note that the same reference symbols are used in each drawing to identify identical or equivalent components and parts. Also, the dimensional ratios in the drawings have been exaggerated for the convenience of explanation and may differ from the actual ratios.

Figure 1 is a diagram showing the schematic configuration of a judgment system equipped with a judgment device according to this embodiment. Shown in Figure 1 is the schematic configuration of a judgment system that compares the color of a manufactured sample with a model color and outputs a judgment result for the color of the sample.

The determination system 1 shown in FIG. 1 includes a determination device 10 and an imaging device 200.

The determination device 10 is a device that uses images of the color sample 210 and the sample 211 captured by the imaging device 200 to output a determination result for the color of the sample 211. The imaging device 200 is fixed to a tripod (not shown) or the like, and captures images of the color sample 210 and the sample 211 placed on the imaging stand 230. The mounting surface 231 of the imaging stand 230 is at an angle of 45 degrees from the horizontal. In this embodiment, the color sample 210 and the sample 211 are captured in a darkroom, and a light source 220 is installed in the darkroom. In this embodiment, the distance between the light source 220 and the bottom side of the imaging stand 230 is 765 mm, and the distance between the imaging device 200 and the color sample 210 and the sample 211 is 445 mm. The imaging device 200 captures the color sample 210 and the sample 211 by setting the shutter speed to 1/15 seconds, the aperture to f8.0, and the ISO sensitivity to 800.

When light from the light source 220 is reflected by the sample 211, the apparent color of the sample 211 being judged may differ from its original color due to the influence of grain or bending. The judgment device 10 according to this embodiment is a device that makes it possible to judge the color of the sample 211 with high robustness compared to evaluation using simple statistical indicators.

Specifically, the determination device 10 according to this embodiment performs sampling from one image using the bootstrap method, and acquires multiple feature amounts from one sampling result. The determination device 10 then analyzes the content of the multiple feature amounts and performs a learning process to determine whether the color of the sample 211 matches the color of the color sample 210. The determination device 10 according to this embodiment creates robust judgment criteria from the image data of the color sample 210 and the sample 211, and can determine the color of the object to be determined with high robustness compared to evaluation using simple statistical indicators.

Figure 2 is a block diagram showing the hardware configuration of the determination device 10.

As shown in FIG. 2, the determination device 10 has a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage 14, an input unit 15, a display unit 16, and a communication interface (I/F) 17. Each component is connected to each other via a bus 19 so as to be able to communicate with each other.

The CPU 11 is a central processing unit that executes various programs and controls each part. That is, the CPU 11 reads a program from the ROM 12 or the storage 14, and executes the program using the RAM 13 as a working area. The CPU 11 controls each of the above components and performs various calculation processes according to the program recorded in the ROM 12 or the storage 14. In this embodiment, the ROM 12 or the storage 14 stores a judgment program that judges the color of the sample 211 using an image captured by the imaging device 200.

The ROM 12 stores various programs and various data. The RAM 13 temporarily stores programs or data as a working area. The storage 14 is composed of a storage device such as a hard disk drive (HDD), a solid state drive (SSD) or a flash memory, and stores various programs including an operating system, and various data.

The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used to perform various input operations.

The display unit 16 is, for example, a liquid crystal display, and displays various information. The display unit 16 may also function as the input unit 15 by adopting a touch panel system.

The communication interface 17 is an interface for communicating with other devices such as the imaging device 200, and uses standards such as Ethernet (registered trademark), FDDI, and Wi-Fi (registered trademark).

When executing the above-mentioned determination program, the determination device 10 realizes various functions using the above-mentioned hardware resources. The functional configuration realized by the determination device 10 will be described below.

Figure 3 is a block diagram showing an example of the functional configuration of the determination device 10.

As shown in FIG. 3, the judgment device 10 has, as its functional components, a sampling unit 101, a feature generation unit 102, a spatial variation calculation unit 105, a difference acquisition unit 106, a learning unit 107, and a judgment unit 108. Each functional component is realized by the CPU 11 reading and executing a judgment program stored in the ROM 12 or the storage 14.

The sampling unit 101 samples image data of the color sample 210 and image data of the sample 211 captured by the imaging device 200. The image of the color sample 210 is an example of the first image data of the present invention, and in this embodiment, the sampling unit 101 samples N sets (N is a natural number) of images of the color sample 210 and the sample 211 using the bootstrap method. The bootstrap method is a method of randomly obtaining many samples from one specimen, allowing overlaps. The number of sets N of sampling using the bootstrap method in the sampling unit 101 may be determined by the user. A user interface for inputting the number of sampling sets N may be displayed on the display unit 16.

Figure 4 is a diagram showing an example of sampling by the bootstrap method in the sampling unit 101. In this embodiment, the sampling unit 101 randomly extracts sub-regions 320 from the image data 310 of the color sample 210 and the image data 311 of the sample 211, allowing overlap. In this embodiment, about 200 sub-regions 320 (i.e. N = 200) are extracted from each image data. Note that the size of the sub-regions 320 is not limited to a specific size.

The feature generation unit 102 converts the image data of the color sample 210 sampled by the sampling unit 101 and the image data of the sample 211 into M types of color spaces (M is a natural number). The feature generation unit 102 generates a first feature pool 103 consisting of M types of features for the image data of the color sample 210, and generates a second feature pool 104 consisting of M types of features for the image data of the sample 211.

Which feature quantity can be used to perform the most robust evaluation depends on the circumstances when the color sample 210 and the sample 211 are captured, and it is difficult to determine a specific feature quantity. Therefore, the feature quantity generation unit 102 converts all possible feature quantities and pools the feature quantity information. In this embodiment, the feature quantity generation unit 102 obtains feature quantities for each image data by performing hue conversion such as L*a*b* conversion, YCrCb conversion, Luv conversion, hls conversion, hsv conversion, and xyz conversion. Note that the type of conversion to be performed on each image data may be specified by the user of the determination device 10.

Figure 5 shows an example of a histogram obtained by hue conversion by the feature generator 102. In Figure 5, as an example, the results of hue conversion to the L*a*b* color space for two samples are shown in histograms. By converting the hue to the L*a*b* color space, features representing brightness, features representing red and green, and features representing yellow and blue are obtained. By looking at the features represented in the histograms and the reference (ref) features, the evaluator determines which features should be adopted to make an accurate judgment. In other words, by adopting features having a histogram that deviates from the histogram of the reference features, it becomes possible to accurately judge the color of the sample 211 using the color sample 210.

For example, in the example shown in FIG. 5, the NG histogram of a* in sample 1 deviates from the shape of the histogram of the reference feature. Therefore, by using a* as the feature for sample 1, it becomes possible to accurately determine the color of sample 211 using color sample 210.

The spatial variation calculation unit 105 calculates the spatial variation from the first feature pool 103 generated by the feature generation unit 102. That is, the spatial variation calculation unit 105 calculates the spatial variation based on the sampling result of the color sample 210. In this embodiment, the spatial variation refers to an amount indicating the degree to which the feature varies. In this embodiment, the spatial variation calculation unit 105 obtains the spatial variation by calculating the standard deviation for the feature accumulated in the first feature pool 103.

The spatial variation calculation unit 105 may take the sum of the standard deviations calculated for each type of feature quantity multiplied by a predetermined weight as the spatial variation. By taking the sum of the standard deviations calculated for each type of feature quantity multiplied by a predetermined weight as the spatial variation, it is possible to quantitatively evaluate the degree of variation in the feature quantities after weighting the types of feature quantities.

Figure 6 is a diagram illustrating the spatial variation amount using a histogram. A wide width of each feature amount shown in the histogram in Figure 6 means a large standard deviation. In other words, a large standard deviation means a large spatial variation. And a large spatial variation means that it is not robust. On the other hand, a narrow width of a feature amount means a small standard deviation. A small standard deviation means a small spatial variation. A small spatial variation means that it is robust. Therefore, by the spatial variation amount calculation unit 105 calculating the spatial variation amount, it is possible to find a robust index. In other words, by the spatial variation amount calculation unit 105 calculating the spatial variation amount, it is possible to find a feature amount that has a sharp shape when it is made into a histogram.

For example, if a sub-region is 100 pixels long and 100 pixels wide, RGB values are input to each pixel, so each of the three features R, G, and B has 10,000 data points. The spatial variation calculation unit 105 calculates the standard deviation for these 10,000 pieces of data, thereby obtaining three standard deviation values for one image.

The difference acquisition unit 106 calculates the difference between the features stored in the first feature pool 103 and the second feature pool 104 for each type of feature to obtain differential features. In this embodiment, the difference acquisition unit 106 uses K types (K is a natural number) of distance calculation methods to obtain the differences between the features in the first feature pool 103 and the second feature pool 104 to obtain K differential features.

Which distance calculation method can be used to perform the most robust evaluation depends on the situation when the color sample 210 and the sample 211 are captured, and it is therefore difficult to determine a specific distance calculation method. Therefore, the difference acquisition unit 106 calculates the difference using as many known distance calculation methods as possible. Examples of distance calculation methods include Euclidean distance, standardized Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance, Canberra distance, Bray-Curtis distance, Haversine distance, Mahalanobis distance, Minkowski distance, cosine distance, interlayer distance, Hamming distance, Jaccard coefficient, Dice coefficient, Russell-Rao coefficient, Rogers-Tanimoto coefficient, sokalmichener, sokalsneath, yule, etc. Note that the user may specify which distance calculation method is used for each feature pool.

The difference acquisition unit 106 calculates the histogram difference for each of the M features by applying K types of distance calculation methods. Therefore, the difference acquisition unit 106 creates M×K difference features from one set. Therefore, the difference acquisition unit 106 can obtain a difference feature matrix of N rows and M×K columns.

Figure 7 is a diagram showing an example of a method for calculating a difference feature by the difference acquisition unit 106. Figure 7 shows an example of calculating the Euclidean distance between the feature of the color sample 210 stored in the first feature pool 103 and the feature of the sample 211 stored in the second feature pool 104. The Euclidean distance is the square root of the sum of the squares of the difference in the properties of the data. By calculating the Euclidean distance over the entire spectrum, it is possible to grasp not only the difference between the peaks of the two spectra, but also the difference in the shape of the spectral distribution. The difference acquisition unit 106 calculates the Euclidean distance between the feature of the color sample 210 and the feature of the sample 211 at each point. In other words, the greater the difference between the feature of the color sample 210 and the feature of the sample 211, the greater the Euclidean distance.

The learning unit 107 learns the learning model 20 for outputting the judgment result of the sample 211 by machine learning. There are many machine learning algorithms that automatically classify data, but in this embodiment, a method called GBDT (Gradient Boosting Decision Tree) is used. The reason for using GBDT is that the judgment criteria can be flexibly changed because the data has characteristics that change the judgment criteria for each sample.

Figure 8 is a diagram explaining the GBDT used by the determination device 10. The determination device 10 determines whether the color of the sample 211 matches the color of the color sample 210 by majority voting of a large number of discriminators. Figure 8 shows an example in which it is determined whether the color of the sample 211 matches the color of the color sample 210 by majority voting of the discrimination results by discriminators using two indexes. Differential feature amounts calculated by different distance calculation methods are used for index 1 and index 2.

The learning unit 107 performs ensemble learning using a categorical variable (binary values of OK and NG) corresponding to the judgment result of the sample 211 for the N rows, M x K columns, differential feature matrix as the objective variable and the spatial variability calculated by the spatial variability calculation unit 105 as the loss function. In ensemble learning, learning is performed so that the estimation error of OK/NG is minimized using the categorical variable as the objective variable. Then, the learning unit 107 sets the spatial variability as the loss function for ensemble learning.

In this embodiment, the learning unit 107 performs ensemble learning using features that minimize the spatial variation, and differential features calculated by a distance calculation method that maximizes the differential features. In the example of Figure 8, differential features calculated by different distance calculation methods are used for index 1 and index 2, and the spatial variation is used as a criterion for determining whether the index is appropriate. The determination device 10 can improve the learning accuracy of ensemble learning by setting the spatial variation in the loss function of ensemble learning.

The judgment unit 108 uses the results of ensemble learning by the learning unit 107 to judge whether the color of the sample 211 matches the color of the color sample 210. For example, the judgment unit 108 outputs OK as the judgment result if the color of the sample 211 matches the color of the color sample 210, and outputs NG as the judgment result if the color of the sample 211 does not match the color of the color sample 210. The judgment unit 108 may display the judgment result on the display unit 16.

Next, the operation of the determination device 10 will be explained.

Figure 9 is a flowchart showing the flow of the determination process by the determination device 10. The CPU 11 reads out a determination program from the ROM 12 or storage 14, expands it into the RAM 13, and executes it to perform the determination process.

In step S101, the CPU 11 samples image data of the color sample 210 and image data of the sample 211 captured by the imaging device 200. In this embodiment, the CPU 11 samples N sets of images of the color sample 210 and images of the sample 211 using the bootstrap method. In this embodiment, the CPU 11 samples approximately 200 sets of images of the color sample 210 and images of the sample 211 using the bootstrap method.

After sampling the image data of the color sample 210 and the image data of the sample 211 in step S101, the CPU 11 then generates a feature pool from the sampling results in step S102. The CPU 11 converts the image data of the color sample 210 and the image data of the sample 211 into M types of color spaces. The CPU 11 generates a first feature pool 103 consisting of M types of feature amounts for the image data of the color sample 210, and generates a second feature pool 104 consisting of M types of feature amounts for the image data of the sample 211. In this embodiment, the CPU 11 obtains feature amounts for each image data by performing hue conversion such as L*a*b* conversion, YCrCb conversion, Luv conversion, hls conversion, hsv conversion, and xyz conversion.

Next, in step S103, the CPU 11 calculates the spatial variation from the first feature pool 103 obtained from the image data of the color sample 210. The spatial variation is an amount that indicates the degree to which the feature varies, and in this embodiment, the CPU 11 obtains the spatial variation by calculating the standard deviation for the feature accumulated in the first feature pool 103.

Next, in step S104, the CPU 11 calculates the difference between the features stored in the first feature pool 103 and the second feature pool 104 for each type of feature to obtain differential features. In this embodiment, the CPU 11 uses K types (K is a natural number) of distance calculation methods to calculate the differences between the features in the first feature pool 103 and the second feature pool 104 to obtain K differential features. Note that the CPU 11 may perform steps S103 and S104 in different orders, or may perform steps S103 and S104 in parallel.

Next, in step S105, the CPU 11 performs machine learning of the learning model 20 to output the judgment result of the sample 211. There are many machine learning algorithms that automatically classify data, but in this embodiment, the CPU 11 performs machine learning of the learning model 20 using the GBDT to output the judgment result of the sample 211.

Then, in step S106, the CPU 11 outputs a judgment result of the sample 211 using the learning model 20 learned by machine learning. For example, if the color of the sample 211 matches the color of the color sample 210, the CPU 11 outputs OK as the judgment result, and if the color of the sample 211 does not match the color of the color sample 210, the CPU 11 outputs NG as the judgment result. The CPU 11 may display the judgment result on the display unit 16.

The determination device 10 executes a series of processes to create robust determination criteria from image data. By creating robust determination criteria from image data, the determination device 10 is able to determine the color of the object to be determined with high robustness compared to evaluation using simple statistical indicators.

The effect of the determination device 10 according to this embodiment will be described. Table 1 shows the evaluation results of the determination accuracy of the sample 211 during cross-validation depending on whether or not sampling is performed using the bootstrap method and whether or not the amount of spatial variation is calculated.

As shown in Table 1, when neither bootstrap sampling nor spatial variation calculation is performed, the judgment accuracy is roughly half, which is almost the same as when OK and NG are selected randomly. When bootstrap sampling is not performed and only spatial variation calculation is performed, the judgment accuracy improves to 64%, but the calculation still outputs an incorrect judgment result about once in three times. When bootstrap sampling is performed and spatial variation is not calculated, the judgment accuracy improves to 85%. Furthermore, when bootstrap sampling is performed and spatial variation is calculated, the judgment accuracy improves further to 92%.

Therefore, the determination device 10 according to this embodiment performs sampling using the bootstrap method and calculates the amount of spatial variation, so that it can output a correct determination result with a probability of more than 9 times out of 10.

In addition, the judgment process executed by the CPU by reading the software (program) in each of the above embodiments may be executed by various processors other than the CPU. Examples of processors in this case include PLDs (Programmable Logic Devices) such as FPGAs (Field-Programmable Gate Arrays) whose circuit configuration can be changed after manufacture, and dedicated electrical circuits such as ASICs (Application Specific Integrated Circuits) that are processors with circuit configurations designed exclusively to execute specific processes. The judgment process may be executed by one of these various processors, or by a combination of two or more processors of the same or different types (e.g., multiple FPGAs, a combination of a CPU and an FPGA, etc.). The hardware structure of these various processors is, more specifically, an electrical circuit that combines circuit elements such as semiconductor elements.

In addition, in each of the above embodiments, the determination process program is described as being pre-stored (installed) in ROM or storage, but this is not limiting. The program may be provided in a form recorded on a non-transitory recording medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a USB (Universal Serial Bus) memory. The program may also be downloaded from an external device via a network.

10 Determination device 200 Imaging device 210 Color sample 211 Sample 230 Light source

Claims (5)

a sampling unit that performs sampling by a bootstrap method on each of first image data in which a color sample serving as a color reference is imaged and second image data in which an object to be determined is imaged;
a feature generating unit that converts sampling results of the first image data and the second image data into M types of color spaces (M is a natural number) to generate a first feature pool and a second feature pool each including M types of feature amounts;
a spatial variation calculation unit that calculates a standard deviation for the M types of feature amounts in the first feature amount pool to obtain a spatial variation amount;
a difference acquisition unit that acquires K difference features by calculating differences between the features in the first feature pool and the second feature pool using K types (K is a natural number) of distance calculation methods;
a learning unit that performs ensemble learning using a categorical variable corresponding to a determination result of the object to be determined for the K difference feature amounts as a response variable and the spatial variation amount as a loss function;
A determination device comprising: The determination device according to claim 1, wherein the feature amount is a value representing a color space after conversion of the first image data and the second image data. The determination device according to claim 1 or 2, wherein the spatial variation calculation unit multiplies the standard deviation calculated for each type of feature by a predetermined weight and adds the multiplied values together to obtain the spatial variation. Sampling the first image data in which a color sample serving as a color reference is imaged and the second image data in which the object to be determined is imaged by using a bootstrap method;
converting the sampling results of the first image data and the second image data into M types of color spaces (M is a natural number) to generate a first feature pool and a second feature pool each consisting of M types of feature amounts;
calculating a standard deviation for the M types of features in the first feature pool to obtain a spatial variation amount;
calculating differences between the features in the first feature pool and the second feature pool using K types of distance calculation methods (K is a natural number), to obtain K difference features;
A determination method, comprising: a process of performing ensemble learning for the K difference feature amounts, the categorical variable corresponding to the determination result of the object to be determined being used as a response variable, and the spatial variation being used as a loss function, the process being performed by a computer. On the computer,
Sampling the first image data in which a color sample serving as a color reference is imaged and the second image data in which the object to be determined is imaged by using a bootstrap method;
converting the sampling results of the first image data and the second image data into M types of color spaces (M is a natural number) to generate a first feature pool and a second feature pool each consisting of M types of feature amounts;
calculating a standard deviation for the M types of features in the first feature pool to obtain a spatial variation amount;
calculating differences between the features in the first feature pool and the second feature pool using K types of distance calculation methods (K is a natural number), to obtain K difference features;
a categorical variable corresponding to a judgment result of the object to be judged for the K difference features is used as a target variable, and the spatial variation amount is used as a loss function to perform ensemble learning.