IMAGE IDENTIFICATION SYSTEM AND IMAGE IDENTIFICATION METHOD

Abstract

To build an image identification model having high accuracy of identification using divided training images obtained by dividing training images.SOLUTION: An image identification system includes: an image dividing unit 110 which divides training images of a first training data set and assigns a label assigned to an undivided training image as tentative labels of the divided training images; a texture index computing unit 120 which computes texture indexes for each of the divided training images; a tentative label prediction model building unit 130 which builds a tentative label prediction model to predict tentative labels assigned to the divided training images based on the texture indexes; and a label comparison unit 140 which compares first tentative labels assigned to the divided training images with second tentative labels predicted for the divided training images with the tentative label prediction model and extracts divided training images having a difference between the first and second tentative labels as divided training images for which it is highly necessary to modify tentative labels.SELECTED DRAWING: Figure 8

Description

[Technical field]

[0001] The present invention relates to an image identification system and an image identification method. [Background technology]

[0002] In the development and quality control of high-performance materials, it is common to identify the microscopic composition and material structure of materials from images captured by a microscope. Traditionally, humans have generally performed the identification visually, but with the recent rapid progress in image recognition technology using artificial intelligence (AI), it is expected that identification will be automatic from microscope images using models pre-constructed by machine learning techniques. This will eliminate the variability in human recognition and enable uniform identification at low cost.

[0003] However, in order to efficiently and accurately classify the material structure of imaged materials using such an image recognition model based on machine learning, the image recognition model needs to be appropriate and trained using a large amount of training data. Since the preparation of such training data, which requires both quality and quantity, requires highly skilled experts, it is desirable to keep the labor costs as low as possible.

[0004] Patent Document 1 discloses a technique for expanding a teacher database by dividing and processing the microscopic images in order to build a large amount of teacher data from a small number of microscopic images. Since dividing and processing the microscopic images may result in images that are different in structure from the original microscopic image and are easily confused with other structures, the feature amount of the divided images is calculated, the feature amount is compared with a threshold, and divided images whose feature amount does not satisfy the threshold are deleted from the teacher data.

[0005] In this embodiment, a system for estimating the microscopic composition and structure of a material from the texture appearing in a microscope image is disclosed, but in the field of remote sensing, estimation of the utilization status and environment of the earth's surface from the texture appearing in a satellite image is also carried out. Non-Patent Document 1 discloses texture indices used for texture analysis of satellite images. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2020-160543 A [Non-patent literature]

[0007] [Non-Patent Document 1] Jingjing Zhou et. al., “The Effects of GLCM parameters on LAI estimation using texture values ​​from Quickbird Satellite Imagery” Scientific Reports volume 7, Article number: 7366 (2017) Summary of the Invention [Problem to be solved by the invention]

[0008] In Patent Document 1, chromaticity, brightness, contrast, structural pattern, frequency, information amount (entropy), etc. are given as examples of image feature quantities for determining suitability of a segmented image as training data, and feature quantities according to the purpose of the model are selected, and threshold values ​​for feature quantities for determining suitability are determined in advance. It is considered possible to select appropriate feature quantities and their threshold values ​​if sufficient prior knowledge is provided in advance about the characteristics of the microscope image of the observation target, but it is desirable to appropriately determine the suitability of a segmented image even in the absence of such prior knowledge. [Means for solving the problem]

[0009] An image recognition system according to one embodiment of the present invention includes a model construction unit that constructs an image recognition model for classifying a recognition target image, and a teacher data creation unit that assigns labels to divided teacher images obtained by dividing a teacher image of a first teacher data set, the first teacher data set being a set of combinations of teacher images and labels assigned to the teacher images, and creates a second teacher data set being a set of combinations of the divided teacher images and the labels assigned to the divided teacher images, The model construction unit includes a learning unit that constructs an image recognition model using the first teacher data set or the second teacher data set; The teacher data creation unit includes an image division unit that divides teacher images of a first teacher dataset and assigns labels assigned to the teacher images before division as tentative labels for the divided teacher images; a texture index calculation unit that calculates a texture index for each of the divided teacher images; a tentative label prediction model construction unit that constructs a tentative label prediction model that predicts the tentative label assigned to the divided teacher image based on the texture index; a label comparison unit that compares a first tentative label assigned to the divided teacher image with a second tentative label predicted for the divided teacher image by the tentative label prediction model and extracts divided teacher images in which the first tentative label and the second tentative label are different as divided teacher images in high need of tentative label correction; and a label correction unit that corrects the label to be assigned to the divided teacher image extracted as the divided teacher image in high need of tentative label correction. Effect of the Invention

[0010] The present invention makes it possible to construct an image recognition model with high recognition accuracy by using divided teacher images obtained by dividing a teacher image. Other objects and novel features will become apparent from the description of this specification and the accompanying drawings. [Brief description of the drawings]

[0011] [Figure 1] 1 is a flowchart for constructing an image recognition model. [Diagram 2]1 is a flowchart for identifying an unknown image (image to be identified). [Figure 3A] This is an example of the first training dataset. [Figure 3B] This is an example of the second training dataset (with provisional labels). [Figure 3C] This is an example of a provisional second training dataset. [Figure 3D] This is an example of the second training dataset (with corrected labels). [Figure 4] 13 is a flowchart for extracting a divided teacher image that is highly in need of correction of a provisional label. [Figure 5A] 13 is an example of a distribution of nodes in an important texture index space. [Figure 5B] This is the distribution of predicted labels predicted using the model (decision tree). [Figure 6] 1 is a flowchart for refining temporary labels using an image identification model. [Figure 7] 13 is a flowchart showing how a user individually modifies labels. [Figure 8] FIG. 1 is a functional block diagram of a microscopic image texture analysis system. [Figure 9A] 13 is an example of management data of the first teacher data set. [Figure 9B] 13 is an example of management data of the second teacher data set. [Figure 10] 1 is an example of a GUI screen. [Figure 11] 13 is a flowchart for selecting an optimal number of divisions of a teacher image. [Figure 12] 1 is an example of a data structure of a discrimination model management database. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] A functional block diagram of the microscope image texture analysis system of this embodiment is shown in Figure 8. The microscope image texture analysis system mainly comprises an information processing device 10 such as a PC or a server, an image display device 20, and an input device 30. The information processing device 10 includes a processing device (processor) and a storage device.

[0013] The functions of the microscope image texture analysis system are realized by a processing device (processor) executing a program stored in a storage device of the information processing device 10. The physical configuration of the information processing device 10 may be a single server or PC, or any part of the processing device or storage device may be configured as multiple servers connected via a network. Furthermore, functions equivalent to all or part of the functions configured by software may be realized by hardware such as an FPGA (Field Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit).

[0014] In Fig. 8, the functional features realized by the microscopic image texture analysis system are shown as a functional block diagram. The details of each function will be described later.

[0015] The storage device 400 is a non-volatile memory such as a hard disk drive (HDD) or a solid state drive (SSD), and stores data necessary for system processing and data obtained as a result of processing. The details of this data will be described later.

[0016] The information processing device 10 is connected to an image display device 20 and an input device 30 such as a keyboard or a pointing device. A user inputs instructions to a GUI (Graphical User Interface) displayed on the image display device 20 using the input device 30.

[0017] FIG. 1 shows a flowchart for constructing an image identification model in a microscopic image texture analysis system. In this flowchart, an image identification model (second image identification model) is constructed by learning divided images obtained by dividing a teacher image as new teacher images. The image identification model uses a known method such as CNN (convolutional neural network). Here, if the teacher image to be divided is an image showing a uniform texture overall, the label given to the divided image may be considered to be the same as the label of the original teacher image. However, there are also cases where the teacher image is an uneven image in which multiple textures are mixed. In this case, the class classification of the teacher image is performed according to the texture that occupies the majority of the image, the characteristic texture, or a characteristic combination of textures. The divided images obtained by dividing such a teacher image do not necessarily have the characteristics that the original teacher image had as a whole, and therefore, if an image identification model is constructed using teacher data in which the same label as the original teacher image is mechanically assigned to the divided images, there is a risk that the classification accuracy of the image identification model will be reduced. In this embodiment, it is possible to create a teacher image set based on high-quality divided images while reducing the burden on the user to reclassify the classes of the divided teacher images.

[0018] First, the model construction unit 200 constructs an image recognition model using a first teacher data set, which is the original teacher image. The learning unit 210 refers to the teacher data management database 410, reads the teacher image of the first teacher data set from the teacher image storage unit 430, and constructs the first image recognition model (S01).

[0019] A set of combinations of teacher images and labels assigned to the teacher images is called a teacher dataset, and FIG. 9A shows an example of management data of a first teacher dataset registered in the teacher data management database 410. In the database 410, an image ID 601, a file name 602, and a label 603 are registered for the teacher images constituting the first teacher dataset. The image ID 601 is an ID that uniquely identifies the teacher image. The file name 602 is the file name of the teacher image stored in the teacher image storage unit 430. The label 603 is a label assigned to the teacher image. FIG. 3A shows an example of the first teacher dataset shown in FIG. 9A.

[0020] Next, the classification accuracy calculation unit 220 obtains the classification accuracy of the first image classification model by cross-validation, and stores it in the classification model management database 420 (S02). FIG. 12 shows an example of the data structure of the classification model management database 420. The database 420 includes columns such as a model ID 801, a creation date 802, a file name 803, a teacher dataset 804, a division number 805, and a classification accuracy 806. The model ID 801 is an ID that uniquely identifies a record in the database 420. The creation date 802 is the creation date of the record. The file name 803 is the file name of the image classification model stored in the classification model storage unit 440. The teacher dataset 804 is a teacher dataset used to build the image classification model. The division number 805 is the number of divisions of the original teacher image. The classification accuracy 806 is the classification accuracy obtained by performing cross-validation on the image classification model.

[0021] 12 is a record of the first image recognition model, and since the first image recognition model is a model constructed using the first training data set which is the original training image as is, the division number 805 is set to "1 x 1." The recognition accuracy of the first image recognition model obtained in step S02 is stored in column 806.

[0022] Next, a process of dividing an original teacher image to create a new teacher data set will be described. In the following, an example will be described in which a teacher image constituting the first teacher data set (FIG. 3A) is divided vertically and horizontally into two equal parts ("2×2") to create a second teacher data set. FIG. 9B shows an example of management data of the second teacher data set registered in the teacher data management database 410 in this process. In this process, an image ID 611, a file name 612, an image ID before division 613, a division number 614, a tentative label 615, a predicted label 616, an X / Y image group 617, and a corrected label 618 are registered for the teacher images constituting the second teacher data set.

[0023] First, the image division unit 110 divides the teacher image and stores it in the teacher image storage unit 430, writes the number of image divisions in the teacher data management database 410 (S03), and assigns the same labels as those in the first teacher data set as provisional labels to the divided teacher images (S04). At this stage, columns 611 to 615 of the management data shown in FIG. 9B are registered. FIG. 3B shows the second teacher data set at step S04. As is clear from a comparison between FIG. 3A and FIG. 3B, the labels of the pre-division teacher images are assigned to the provisional labels of the divided teacher images.

[0024] 9B, image ID 611 is an ID that uniquely identifies a divided teacher image, and here, it is identified by adding a sub-number to pre-segmentation image ID 613. File name 612 is the file name of the divided teacher image stored in teacher image storage unit 430. Division number 614 indicates the number of divisions from the pre-segmentation image. In tentative label 615, the label that was assigned to the pre-segmentation image is mechanically assigned to the divided teacher image.

[0025] Next, from among the divided teacher images, those for which the provisional labels are likely to need to be corrected are extracted (S05). For this purpose, texture indices of the images are extracted, an analysis model based on the texture indices of the images is constructed, and the provisional labels are compared with the labels predicted by the analysis model to extract provisional labels that are likely to need correction. Details of step S05 are shown in FIG.

[0026] The texture index calculation unit 120 calculates N predetermined texture indices for each divided teacher image (S21). The texture index is an index of the texture of the image, and may be, for example, a texture index such as brightness, frequency, uniformity, local uniformity, contrast, entropy, non-uniformity, edge strength, intensity of change in brightness and darkness, correlation, or a linear combination of these. These are all examples, and other texture indices may be used.

[0027] The tentative label prediction model construction unit 130 constructs a model (decision tree) that predicts the tentative label by selecting a texture index that is highly correlated with the tentative label from among the N texture indices (S22), and stores the texture index (called important texture index) and threshold value used in the model (decision tree) (S23). FIG. 5A shows an example of distribution of nodes in the important texture index space. Each node corresponds to a divided teacher image, and is shown in different colors according to the tentative label assigned to it. In the texture index space, in order to make the tentative labels of the nodes distributed within the area divided by the threshold value shown by the dashed line as similar as possible, in the construction of the model (decision tree), a texture index that is highly correlated with the tentative label is selected as the texture index used in the model construction, and the threshold value for dividing the space is appropriately selected so that nodes with different tentative labels are not mixed as much as possible. Here, an example of an important texture index space defined by two texture indices that are easy to illustrate is shown, but a model based on an important texture index space defined by two or more indices may be used.

[0028] The model (decision tree) created in step S22 is used to predict the labels of the divided teacher images (S24). The tentative labels predicted using the model are called predicted labels. The distribution of predicted labels predicted using the model (decision tree) of FIG. 5A is shown in FIG. 5B. The distribution of nodes is the same in FIG. 5A and FIG. 5B, and the predicted labels shown in FIG. 5B differ from the tentative labels in FIG. 5A in that the nodes distributed within the area divided by the threshold value shown by the dashed line are the same.

[0029] The label comparison unit 140 compares the predicted label predicted by the texture index with the tentative label, classifies images where the two match as image group X, and images where the two do not match as image group Y (S25), and registers the image group Y in the teacher data management database 410 as teacher images for which the tentative labels need to be corrected (S26). In the example of Fig. 5B, the divided teacher images corresponding to the nodes indicated by the arrows are classified into image group Y, and the divided teacher images corresponding to the other nodes are classified into image group X.

[0030] At the stage where step S05 is completed, columns 616 to 617 of the management data shown in Fig. 9B are registered. The predicted label predicted in step S24 is registered in predicted label 616, and the classification of the image group classified in step S25 is registered in X / Y image group 617. In this way, by performing a process of selecting texture indices that closely reproduce the tentative label rather than selecting texture indices by default, it becomes possible to appropriately extract divided teacher images that are highly necessary to correct the tentative label for any type of microscope image.

[0031] In step S05 (see FIG. 1) described above, for the divided teacher images extracted as having a high need for correction of the tentative labels, the label correction unit 150 checks and corrects the labels, and creates a second teacher data set (S06). The label correction may be performed automatically using an image recognition model, or may be performed individually by the user using a GUI. Each correction method will be described.

[0032] FIG. 6 is a flowchart for correcting the provisional labels using a model. Based on the classification result of step S25, a combination of the image group X and the provisional labels is set as a provisional second teacher data set (S31). FIG. 3C shows the provisional second teacher data set. The learning unit 210 of the model construction unit 200 constructs a provisional second image discrimination model using the provisional second teacher data set (S32), and the discrimination unit 330 of the image discrimination unit 300 estimates labels to be assigned to the divided teacher images of the image group Y using the provisional second image discrimination model (S33). The labels estimated by the provisional second image discrimination model are corrected to the labels of the divided teacher images of the image group Y (S34).

[0033] Fig. 7 is a flowchart showing how a user corrects labels individually using a GUI. The divided teacher images classified into image group Y in step S25 and their provisional labels are presented to the user via the GUI, and the user is asked whether or not to correct the labels (S41). Fig. 10 shows an example of the GUI screen. The user corrects the labels by visually selecting the corresponding labels for the divided teacher images displayed on the GUI screen 700 (S42).

[0034] 3D shows the second teacher data set in which the labels of the divided teacher images of the image group Y have been corrected. When the labels assigned to the divided teacher images of the image group Y are corrected by either correction method, the labels assigned to each divided teacher image are registered in column 618 (corrected label 618) of the management data shown in FIG. 9B (S07). In this way, log data until the labels assigned to the divided teacher images of the second teacher data set are finally determined remains in the teacher data management database 410, and can be used to investigate the cause of a problem such as a failure to improve the classification accuracy of the second image classification model constructed using the second teacher data set.

[0035] Using the second teacher data set created in this way, the learning unit 210 of the model construction unit 200 constructs a second image recognition model (S08), and the recognition accuracy calculation unit 220 calculates the recognition accuracy of the second image recognition model by cross-validation and stores it in the recognition model management database 420 (S09). The processing here is similar to steps S01 and S02, so a duplicated explanation will be omitted.

[0036] Next, the model selection unit 310 of the image classification unit 300 compares the classification accuracy of the first image classification model and the second image classification model, and selects the image classification model with the higher classification accuracy as the image classification model to classify the unknown image (image to be classified) (S10).

[0037] 2 shows a flowchart for classifying an image to be classified using the image classification model selected in step S10. This flowchart is executed by the image classification unit 300.

[0038] First, the classification unit 330 reads out the image classification model selected by the model selection unit 310 from the classification model storage unit 440 (S11).

[0039] Meanwhile, the image dividing unit 320 refers to the discrimination model management database 420 and reads the division number 805 of the teacher dataset used by the image discrimination model selected by the model selection unit 310 (S12). Next, the discrimination target image is read (S13), and the read discrimination target image is divided by the division number read in step S12 (S14). For example, if the second discrimination model is selected in step S10, the discrimination unit 330 uses the second image discrimination model to discriminate the discrimination target image divided into 2×2 (S15), and stores the discrimination result in the discrimination result storage unit 450 (S16).

[0040] In the above explanation, an example was explained in which the teacher image was divided into 2×2, but the number of divisions is not limited to this. A plurality of division methods may be tried to select the optimal image recognition model. A flowchart of this method is shown in FIG. 11. In FIG. 11, the teacher image is divided into 2×2. nThe n+1th teacher data set is created using the divided teacher image divided equally. The first to m+1th teacher data sets are created by sequentially adding 1 from 1 until n becomes m (S58, S59), and steps S54 to S57 are repeated to construct the first to m+1th image classification models using these teacher data sets and calculate the classification accuracy. Steps S51 to S57 when n=1 are the same processes as S01 to S09 in the flowchart of FIG. 1, and the flow of FIG. 11 repeats the same processes by changing the number of divisions of the teacher image, so duplicated explanations will be omitted. Information on the classification models constructed in this process is stored in the classification model management database 420 as shown in FIG. 12.

[0041] The model selection unit 310 of the image classification unit 300 compares the classification accuracy of the first to (m+1)th classification models, and selects the image classification model with the highest classification accuracy as the image classification model that will classify the classification target image (S60).

[0042] The present invention has been described above based on the embodiments and modifications. The present invention is not limited to the above embodiments and modifications, and various modifications are possible without departing from the spirit and scope of the invention. [Explanation of symbols]

[0043] 10: information processing device, 20: image display device, 30: input device, 100: teacher data creation unit, 110: image division unit, 120: texture index calculation unit, 130: provisional label prediction model construction unit, 140: label comparison unit, 150: label correction unit, 200: model construction unit, 210: learning unit, 220: classification accuracy calculation unit, 300: image classification unit, 310: model selection unit, 320: image division unit, 330: classification unit, 400: storage device, 410: teacher data management database, 420: classification model management database database, 430: teacher image storage unit, 440: classification model storage unit, 450: classification result storage unit, 601: image ID, 602: file name, 603: label, 611: image ID, 612: file name, 613: pre-division image ID, 614: number of divisions, 615: provisional label, 616: predicted label, 617: X / Y image group, 618: corrected label, 700: GUI screen, 801: model ID, 802: creation date, 803: file name, 804: teacher dataset, 805: number of divisions, 806: classification accuracy.

Claims

1. a model construction unit that constructs an image recognition model for classifying a classification target image; a teacher data creation unit that assigns labels to divided teacher images obtained by dividing a teacher image of a first teacher data set, the divided teacher image being a set of combinations of the teacher image and the labels assigned to the teacher image, and creates a second teacher data set being a set of combinations of the divided teacher images and the labels assigned to the divided teacher images, The model construction unit includes a learning unit that constructs the image recognition model by using the first teacher data set or the second teacher data set, The teacher data creation unit is an image division unit that divides a teacher image of the first teacher data set and assigns a label that was assigned to the teacher image before division as a provisional label of the divided teacher image; a texture index calculation unit for calculating a texture index for each of the divided teacher images; a tentative label prediction model construction unit that constructs a tentative label prediction model that predicts a tentative label assigned to the divided teacher image based on the texture index; a label comparison unit that compares a first tentative label assigned to the divided teacher image with a second tentative label predicted for the divided teacher image by the tentative label prediction model, and extracts a divided teacher image in which the first tentative label and the second tentative label are different as a divided teacher image in which the tentative label needs to be corrected; a label correction unit that corrects a label to be assigned to a divided teacher image extracted as a divided teacher image having a high need for correction of the provisional label, 1. An image recognition system comprising:

2. In claim 1, The image recognition unit further comprises: the model construction unit includes a classification accuracy calculation unit that calculates classification accuracy for each of a first image classification model constructed using the first teacher data set and a second image classification model constructed using the second teacher data set; The image identification unit a model selection unit that compares the classification accuracy of the first image classification model with the classification accuracy of the second image classification model, and selects an image classification model with a higher classification accuracy as an image classification model to be used for classifying the classification target image; an image division unit that divides the classification target image according to the number of divisions of the teacher image of the image classification model selected by the model selection unit to create a divided classification target image; a classification unit that classifies the divided classification target image using the image classification model selected by the model selection unit, 1. An image recognition system comprising:

3. In claim 1, A teacher data management database for storing management data of a teacher data set, The management data of the second teacher data set includes log data indicating a process in which the teacher data creation unit created the second teacher data set.

1. An image recognition system comprising:

4. In claim 1, the texture index calculation unit calculates a plurality of predetermined texture indices for each of the divided teacher images; the tentative label prediction model construction unit selects a texture index having a high correlation with the tentative label from among the plurality of texture indices, and constructs the tentative label prediction model based on the selected texture indices; 1. An image recognition system comprising:

5. In claim 4, An image classification system characterized in that the texture indices include any of brightness, frequency, uniformity, local uniformity, contrast, entropy, non-uniformity, edge strength, strength of light-dark changes, correlation, or linear combinations thereof.

6. In claim 1, The label correction unit includes: correcting the label of a divided teacher image extracted as a divided teacher image for which the provisional label is highly in need of correction, by using a provisional second image recognition model constructed by the model construction unit using a provisional second teacher data set which is a set of combinations of divided teacher images in which the first provisional label and the second provisional label are equal, and the first provisional label, to identify divided teacher images in which the first provisional label and the second provisional label are different; 1. An image recognition system comprising:

7. In claim 1, The label correction unit includes: displaying a divided teacher image in which the first tentative label and the second tentative label are different on an image display device, and allowing the user to select a label to be assigned to the divided teacher image; 1. An image recognition system comprising:

8. In claim 2, the teacher data creation unit creates a plurality of the second teacher data sets having different numbers of divisions; The model selection unit of the image discrimination unit compares the discrimination accuracy of the first image discrimination model with the discrimination accuracy of the plurality of second image discrimination models, and selects the image discrimination model with the highest discrimination accuracy as the image discrimination model to be used for discriminating the discrimination target image.

1. An image recognition system comprising:

9. In claim 1, 11. An image recognition system according to claim 10, wherein the image to be recognized is a microscope image.

10. An image classification method using an image classification system including: a model construction unit that constructs an image classification model for classifying a classification target image; and a teacher data creation unit that assigns labels to divided teacher images obtained by dividing a teacher image of a first teacher data set, the first teacher data set being a set of combinations of teacher images and labels assigned to the teacher images, and creates a second teacher data set being a set of combinations of the divided teacher images and the labels assigned to the divided teacher images, The teacher data creation unit is Dividing a teacher image of the first teacher data set, and assigning a label assigned to the teacher image before division as a provisional label of the divided teacher image; Calculating a texture index for each of the divided teacher images; A tentative label prediction model is constructed to predict a tentative label assigned to the divided teacher image based on the texture index; comparing a first tentative label assigned to the divided teacher image with a second tentative label predicted for the divided teacher image by the tentative label prediction model, and extracting a divided teacher image in which the first tentative label and the second tentative label are different as a divided teacher image in which the tentative label needs to be corrected highly; correcting the labels to be assigned to the divided teaching images extracted as divided teaching images for which the provisional labels are highly in need of correction; The model construction unit constructs a first image recognition model and a second image recognition model by using the first teacher data set and the second teacher data set, respectively.

13. An image identification method comprising:

11. In claim 10, The image classification system further includes an image classification unit that classifies the classification target image using the image classification model, The image identification unit comparing the classification accuracy of the first image classification model with the classification accuracy of the second image classification model, and selecting an image classification model with a higher classification accuracy as an image classification model to be used for classifying the classification target image; Divide the classification target image according to the number of divisions of the teacher image of the selected image classification model to create a divided classification target image; classifying the segmented classification target image using the selected image classification model; 13. An image identification method comprising:

12. In claim 10, the teacher data creation unit calculates a plurality of texture indices determined in advance for each of the divided teacher images, selects a texture indices having a high correlation with the tentative label from the plurality of texture indices, and constructs the tentative label prediction model based on the selected texture indices; 13. An image identification method comprising:

13. In claim 10, the teacher data creation unit corrects the label of the divided teacher image extracted as the divided teacher image for which the provisional label is highly in need of correction by classifying a divided teacher image for which the first provisional label and the second provisional label are different from each other using a provisional second image recognition model constructed using a provisional second teacher data set which is a set of combinations of the divided teacher image and the first provisional label, the divided teacher image having the first provisional label and the second provisional label being equal; 13. An image identification method comprising:

14. In claim 10, the teacher data creation unit displays a divided teacher image, in which the first tentative label and the second tentative label are different, on an image display device, and allows the user to select a label to be assigned to the divided teacher image; 13. An image identification method comprising:

15. In claim 11, the teacher data creation unit creates a plurality of the second teacher data sets having different numbers of divisions; The image discrimination unit compares the discrimination accuracy of the first image discrimination model with the discrimination accuracy of the plurality of second image discrimination models, and selects the image discrimination model having the highest discrimination accuracy as the image discrimination model to be used for discriminating the discrimination target image.

13. An image identification method comprising:

Images