Training assistance device

Abstract

The purpose of the present invention is to assist in creation of an image classification model with a small number of teacher information items by training the model with separable feature expressions while paying attention to a minute detection region. A training assistance device according to the present invention: creates a set of detection region images for inspecting whether or not an object has an abnormality; updates a feature amount extraction model so that the similarity degree within the same detection region becomes high; and updates the feature amount extraction model so that the similarity degree between two image regions including detection regions different from each other becomes low (see fig. 1).

Description

Learning support device

[0001] The present invention relates to a technique for assisting in learning, by machine learning, a classification model for classifying an abnormality of an object by an image of the object.

[0002] In machine learning, an operation (labeling or annotation) for imparting information to data to be learned is required. For example, in order to create an image classification model using deep learning (DL: Deep Learning), a huge amount of teacher information is required, and the annotation cost is high.

[0003] Patent Document 1 describes a technique related to labeling. This document describes "training a standard DL-based object detection system with less labeled data achieved using self-supervised learning representations" (see paragraph 0009 of this document).

[0004] Japanese Unexamined Patent Application Publication No. 2023-126130

[0005] In the prior art such as Patent Document 1, training of an object detection system for minute objects is not assumed. Patent Document 1 applies data augmentation including random crop (randomly cutting out a region from the original image) during self-supervised learning. However, due to randomly cropping, this method learns the expression of the entire image, so it is difficult to focus on minute objects that occupy a small area within the image, and instead focuses on dominant objects within the image. Therefore, the expression learning of minute objects fails, and the annotation cost increases.

[0006] The present invention has been made in view of the above problems, and aims to assist in creating an image classification model with a small amount of teacher information by learning separable feature expressions while focusing on minute detection regions.

[0007] The learning support device according to the present invention creates a set of detection region images for inspecting whether an object has an abnormality, updates the feature extraction model so that the similarity between identical detection regions is high, and updates the feature extraction model so that the similarity between two image regions containing different detection regions is low.

[0008] The learning support device according to the present invention can learn separable feature representations while focusing on minute detection regions, and create an image classification model with a small amount of training information. Other issues, configurations, advantages, etc. will be clarified by the following description of the embodiments.

[0009] This is a diagram of the learning support device 100. This is an internal diagram of the detection region image creation unit 103. This shows specific examples of the detection region image set 1037 and the similar normal region image set 1038. This is an internal diagram of the image similarity calculation unit 104. This is an internal diagram of the image distribution display unit 105. This shows an example of a user interface for the similarity quality determination unit 1051 to present a scatter plot to the user. This shows an example of a user interface for the annotation candidate presentation unit 1052 to present annotation candidates to the user. This is a flowchart explaining the process by which the detection region image creation unit 103 creates the detection region image set 1037 and the similar normal region image set 1038 by inputting the learning image data 101 and the detection region coordinate data 102. This is a flowchart explaining the process by which the image similarity calculation unit 104 determines the image similarity by inputting the detection region image set 1037 and the similar normal region image set 1038, and the image distribution display unit 105 saves the annotation image set 1053.

[0010] The embodiments of the present invention described below are described in relation to 2D data (e.g., image data), but embodiments of the present invention can also be applied to 3D data. That is, the word "image" includes both 2D images (e.g., represented by pixels) and 3D scenes (e.g., represented by voxels or point clouds).

[0011] Figure 1 is a configuration diagram of a learning support device 100 according to an embodiment of the present invention. Embodiment 1 describes a case in which a semiconductor is inspected using images captured by a scanning electron microscope or the like. In the inspection process, minute defects are classified using abnormality candidates detected by abnormality detection.

[0012] The learning support device 100 receives as input (a) training image data 101 which is an inspection image of a product, and (b) detection region coordinate data 102 which indicates the position information of a detection region detected as an anomaly candidate by anomaly detection within the training image data 101. The learning support device 100 further includes: a detection region image creation unit 103 which creates a partial image from the training image data 101 based on the detection region; an image similarity calculation unit 104 which calculates the similarity between vectors of the partial images; and an image distribution display unit 105 which displays the similarity results.

[0013] Figure 2 is an internal configuration diagram of the detection region image creation unit 103. The detection region image creation unit 103 comprises the following: a detection region image creation unit 1031 that creates a partial image from training image data 101 and detection region coordinate data 102; a detection region specification parameter 1032 that describes the conditions under which the detection region image creation unit 1031 creates a partial image; a similar normal region image creation unit 1033 that searches for a normal region similar to the detected region from the training image data 101 and detection region coordinate data 102 and creates a partial image of it; a similar normal region specification parameter 1034 that describes the conditions under which the similar normal region image creation unit 1033 creates a partial image; a preprocessing unit 1305 that creates a detection region image set 1037, which is a partial image derived from the detected region, and a similar normal region image set 1038, which is a partial image derived from the normal region, by preprocessing the created partial image by adding noise, etc.; and an image extension parameter 1036 that describes the parameters of the preprocessing.

[0014] The detection region image creation unit 1031 selects a detection region based on the detection region coordinate data 102 and acquires multiple partial images from the training image data 101 by cropping so that the selected detection region is included. The method by which the detection region image creation unit 1031 acquires partial images is not limited to this.

[0015] The detection region specification parameter 1032 is a parameter that specifies the crop area when the detection region image creation unit 1031 creates a partial image. For example, it describes parameters such as the image size of the partial image, and the maximum and minimum area of ​​the region shared by multiple partial images. Other parameters used to specify the partial image may also be described.

[0016] The similar normal region image creation unit 1033 searches within the training image data 101 for normal regions that have similar features to the partial images created by the detection region image creation unit 1031 and for which no abnormalities have been detected. Multiple partial images are obtained from the normal regions obtained through the search by cropping. Template matching is one method of search. Other search methods may also be used. For example, if there is a reference image such as design data, it may be used as a similar normal region image.

[0017] The similar normal region specification parameter 1034 is a parameter used by the similar normal region image creation unit 1033 to specify the crop area when creating a partial image. For example, it describes parameters such as the image size of the partial image and the similarity threshold between the detected region and the normal region. Other parameters used to specify the partial image may also be described.

[0018] The preprocessing unit 1305 performs preprocessing on the created partial image, such as adding noise. Examples of preprocessing include color jittering, grayscale conversion, Gaussian blur, horizontal flipping, vertical flipping, random cropping, and resizing. Other preprocessing may also be performed.

[0019] The image enhancement parameter 1036 controls the probability by which the preprocessing unit 1305 performs each preprocessing step by specifying the application probability. For example, by setting the application probability of Gaussian blur to 50% and the application probability of vertical flipping to 50%, either Gaussian blur or vertical flipping will be applied to the partial image with equal probability. It is also possible to perform no preprocessing at all by setting all application probabilities to 0%.

[0020] The detection region image set 1037 is a set of images in which the preprocessing unit 1305 has preprocessed a partial image created by the detection region image creation unit 1031.

[0021] The similar normal region image set 1038 is a set of images in which the preprocessing unit 1305 has preprocessed a partial image created by the similar normal region image creation unit 1033.

[0022] Figure 3 shows specific examples of the detection region image set 1037 and the similar normal region image set 1038. The detection region 301 is a detection region obtained from the detection region coordinate data 102. Partial image 302 is an example of a detection region image created by the detection region image creation unit 1031 selecting the detection region 301. Partial image 303 is another detection region image that includes the detection region 301. Partial image 304 is an example of a similar normal region image created by the similar normal region image creation unit 1033 selecting part image 302 as the detection region image. Partial image 305 is an example of a similar normal region image created by the similar normal region image creation unit 1033 selecting part image 303 as the detection region image.

[0023] Figure 4 is an internal diagram of the image similarity calculation unit 104. The image similarity calculation unit 104 comprises: a feature extraction unit 1041 that creates a detection region feature vector set 1042 and a similar normal region feature vector set 1043 by extracting features from the image data of the detection region image set 1037 and the similar normal region image set 1038; a vector comparison unit 1044 that calculates the similarity between the vectors between the detection region feature vector set 1042 and the similar normal region feature vector set 1043; and a model update unit 1045 that calculates model update parameters 1046 based on the vector similarity.

[0024] The feature extraction unit 1041 extracts features (feature vectors) from the images in the detection region image set 1037 and the similar normal region image set 1038. Methods for extracting features include deep learning methods such as convolutional neural networks and multilayer perceptrons, as well as other machine learning methods. Other methods may also be used.

[0025] The detected region feature vector set 1042 is a feature vector set extracted by the feature extraction unit 1041 from the detected region image set 1037.

[0026] The similar normal region feature vector set 1043 is a feature vector set extracted by the feature extraction unit 1041 from the similar normal region image set 1038.

[0027] The vector comparison unit 1044 calculates the following similarities: a first similarity, which is the similarity between vector 1 originating from detection region 1 and vector 2 originating from detection region 1 (detection region 1 is the same detection region from which feature vectors included in the detection region feature vector set 1042 are extracted); a second similarity, which is the similarity between vector 1 or vector 2 and a vector originating from detection region 2 (a detection region different from detection region 1); a third similarity, which is the similarity between vector 1 and a vector originating from the similar normal region of detection region 1 among the feature vectors included in the similar normal region feature vector set 1043; and a fourth similarity, which is the similarity between a vector originating from the similar normal region of detection region 1 and a vector originating from the similar normal region of detection region 2. Examples of methods for calculating similarity include cosine similarity and Euclidean distance. Other methods may also be used.

[0028] The model update unit 1045 calculates the model update parameters 1046, which are the amount of model parameter updates, so that the first and fourth similarities calculated by the vector comparison unit 1044 are high, and the second and third similarities are low. The model update parameters 1046 are parameters that define the feature extraction model (such as a convolutional neural network in the above example) used by the feature extraction unit 1041 when extracting image features. The method for updating the model parameters is, but is not limited to, backpropagation.

[0029] Updating the model parameters to increase the first similarity, which is the similarity between vectors originating from images containing the same detection region, has the effect of extracting features that make the model focus on the detection region. Updating the model parameters to decrease the second similarity, which is the similarity between vectors originating from images containing different detection regions, has the effect of extracting features that make the model focus on differences in the detection regions. Updating the model parameters to decrease the third similarity, which is the similarity between an image containing the detection region and an image containing a similar normal region, has the effect of preventing the model from focusing on surrounding features and instead focusing on the detection region. Updating the model parameters to increase the fourth similarity, which is the similarity between images containing similar normal regions of different detection regions, has the effect of preventing the model from focusing on surrounding features.

[0030] Figure 5 is an internal configuration diagram of the image distribution display unit 105. The image distribution display unit 105 includes: a similarity quality determination unit 1051 that determines the learning status of the model based on the similarity calculated by the vector comparison unit 1044; an annotation candidate presentation unit 1052 that presents annotation candidates to the user based on the features determined by the similarity quality determination unit 1051; and an annotation image set 1053 that manages the data on which annotation has been performed.

[0031] The similarity quality determination unit 1051 receives vectors originating from the detection region and the similar normal region, respectively, along with the first, second, third, and fourth similarity values ​​from the vector comparison unit 1044, and visualizes these as a scatter plot. By checking whether the feature points are distributed in the scatter plot such that the first and fourth similarity values ​​are high and the second and third similarity values ​​are low, the quality of the model's training can be determined. If the determination is negative, the first, second, third, and fourth similarity values ​​are passed to the model update unit 1045, and the model is updated. As a method for visualizing the features as a scatter plot, for example, one method is to reduce the feature vector to two dimensions using algorithms such as UMAP (Uniform Manifold Approximation and Projection) or t-SNE (t-distributed stochastic neighbor embedding), and then plot each feature. Other methods may also be used.

[0032] The annotation candidate presentation unit 1052 selects annotation candidates from vectors originating from the detected region and presents them to the user when the similarity quality determination unit 1051 determines that the model training is satisfactory. Possible selection methods include presenting data in order of increasing distance from the normal region. Other selection methods may also be used.

[0033] The annotation image set 1053 is a dataset that manages annotation images when a user annotates an annotation candidate presented by the annotation candidate presentation unit 1052.

[0034] Figure 6 shows an example of a user interface for the similarity quality determination unit 1051 to present a scatter plot to the user. When button 601 is pressed, the similarity quality determination unit 1051 searches for the image to be presented. When button 602 is pressed, the similarity quality determination unit 1051 changes the image. The image display unit 603 displays some images from the detection region image set 1037 and the similar normal region image set 1038. The scatter plot 604 displays a scatter plot of similarity calculated by the vector comparison unit 1044 after inputting the images in the image display unit 603 to the feature extraction unit 1041. When button 605 is pressed, the similarity quality determination unit 1051 determines that the learning status of the currently displayed image is good. When button 606 is pressed, the similarity quality determination unit 1051 determines that the learning status of the currently displayed image is bad.

[0035] Examples of the images displayed by the image display unit 603 are as follows: Images (1) and (2) are two detection region images containing the same detection region. Image (3) is a similar normal region image corresponding to image (1). Image (5) is a detection region image containing a different detection region from images (1) and (2). Image (4) is a similar normal region image corresponding to image (5).

[0036] The arrows displayed in scatter plot 604 correspond to the following similarities: The arrow between images (1) and (2) represents the first similarity. The arrow between images (2) and (5) represents the second similarity. The arrow between images (1) and (3) represents the third similarity. The arrow between images (3) and (4) represents the fourth similarity. Users can determine that the learning results are good if the first and fourth similarities are high (i.e., the distance between the features represented by the black circles is small) and the second and third similarities are low (i.e., the distance between the features represented by the black circles is large).

[0037] Figure 7 shows an example of a user interface for the annotation candidate presentation unit 1052 to present annotation candidates to the user. When button 701 is pressed, the annotation candidate presentation unit 1052 changes the criteria used to select annotation candidates. When button 702 is pressed, the annotation candidate presentation unit 1052 changes the annotation candidate to the next candidate. The image display unit 703 displays the images from the detection region image set 1037. The scatter plot 704 displays a scatter plot in which the feature vectors of the partial images are distributed on a two-dimensional graph. When text is entered or a label is selected in the dropdown list 705, the annotation candidate presentation unit 1052 determines the label of the image displayed in the image display unit 703. When button 706 is pressed, the annotation candidate presentation unit 1052 saves the image from the image display unit 703 and the label from the dropdown list 705 in the annotation image set 1053.

[0038] The scatter plot 704 is a graph that maps the feature vectors of partial images onto a two-dimensional space. Depending on the type of defect, the feature vectors are clustered within a certain region. Therefore, if you want to annotate a specific type of defect, you just need to select any of the partial images within that region (black circles in the scatter plot 704). The selected partial image will be displayed in the image display unit 703.

[0039] Figure 8 is a flowchart illustrating the process by which the detection region image creation unit 103 creates a detection region image set 1037 and a similar normal region image set 1038 by inputting training image data 101 and detection region coordinate data 102.

[0040] Step S801: The user captures an image of the target for which an abnormality is to be detected and saves it as learning image data 101.

[0041] Step S802: The user specifies detection area coordinate data 102 that describes the coordinates where an abnormality was detected on the sample.

[0042] Step S803: The detection area image creation unit 103 selects a detection area from the detection area coordinate data 102 and crops a plurality of images from the learning image data 101 so that the selected detection area is included. The detection areas are sequentially selected, and a plurality of cropped images derived from different detection areas are created. As described using FIG. 2, the crop area is determined by the detection area specification parameter 1032. The detection area specification parameter 1032 is a parameter that determines the image size, the area size shared by different images obtained from the same detection area, and the like.

[0043] Step S804: The detection area image creation unit 103 identifies a normal area that is similar in characteristics to the surrounding area of the image created in step S803 and in which no abnormality has been detected, and crops the same number of images. There is no need to search for the normal area from the learning image data 101. As described using FIG. 2, the crop area is determined by the similar normal area specification parameter 1034. The similar normal area specification parameter 1034 is a parameter that determines the image size, the threshold for the similarity of the surrounding characteristics between the detection area and the normal area, and the like.

[0044] Step S805: The detection area image creation unit 103 performs preprocessing such as color dithering, grayscale conversion, Gaussian blurring, etc. on the images created in steps S803 and S804 to create a detection area image set 1037 and a similar normal area image set 1038. As described using FIG. 2, the preprocessing is determined by the image expansion parameter 1036. The image expansion parameter 1036 is a parameter that determines the application probability of each image expansion.

[0045] Figure 9 is a flowchart for explaining the process in which the image similarity calculation unit 104 determines the image similarity by inputting the detection region image set 1037 and the similar normal region image set 1038, and the image distribution display unit 105 saves the annotation image set 1053.

[0046] Step S901: By inputting the detection region image set 1037 and the similar normal region image set 1038, the feature extraction unit 1041 outputs the detection region feature vector set 1042 and the similar normal region feature vector set 1043. The feature extraction unit 1041 does not necessarily have to be a machine learning model.

[0047] Step S902: The vector comparison unit 1044 extracts vectors from the detection region feature vector set 1042 that are from the same detection region and vectors from different detection regions, and further extracts vectors from the similar normal region corresponding to the extracted detection regions from the similar normal region feature vector set 1043. The vector comparison unit 1044 calculates the first, second, third, and fourth similarities described in FIG. 3 using these vectors.

[0048] In step S903, it is checked whether the number of parameter updates of the feature extraction unit 1041 has reached the number set in advance by the user. The number of parameter updates indicates the progress of learning. If it can be regarded as an index indicating the progress of learning, other numerical values than the number of parameter updates of the feature extraction unit 1041 may be used. If the number of parameter updates has reached the specified number, proceed to S904; if not, skip to S906.

[0049] Step S904: The image distribution display unit 105 distributes the vectors used by the vector comparison unit 1044 as a scatter plot. The user checks the learning status using the scatter plot. It is not necessary to be a scatter plot as long as it is a method that can visualize the similarity calculated by the vector comparison unit 1044.

[0050] Step S905: The user checks the learning status and checks whether it is good. If it is not good, proceed to S906; if it is good, skip to S908.

[0051] Step S906: The model update unit 1045 calculates the model update parameters 1046 by calculating a gradient such that the first and fourth similarities increase and the second and third similarities decrease.

[0052] Step S907: The model update unit 1045 updates the feature extraction model using the model update parameters 1046 calculated in S906. Return to S901.

[0053] Step S908: The annotation candidate presentation unit 1052 calculates an anomaly score from the similarity between vectors. The annotation candidate presentation unit 1052 presents annotation candidate regions to the user based on the anomaly score. The anomaly score is a parameter that represents the severity of the defect. The anomaly score may be obtained from the defect inspection results, or the third similarity itself may be considered as the severity of the defect and used as the anomaly score.

[0054] Step S909: The user annotates the annotation candidate region. The annotation candidate presentation unit 1052 saves the result in the annotation image set 1053.

[0055] <Regarding Modifications of the Invention> The present invention is not limited to the embodiments described above, and includes various modifications. For example, the embodiments described above are described in detail to make the present invention easier to understand, and are not necessarily limited to those having all the described configurations. Furthermore, it is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. In addition, it is possible to add, delete, or replace parts of the configuration of each embodiment with other configurations.

[0056] Each of the above configurations, functions, processing units, and processing means may be implemented in hardware, in whole or in part, for example, by designing them as integrated circuits. Alternatively, each of the above configurations and functions may be implemented in software by having the processor interpret and execute programs that implement each function. Information such as programs, tables, and files that implement each function can be stored in memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD.

[0057] In the above embodiments, the image distribution display unit 105 may determine whether the learning status was good or bad (S905) based on user input on the user interface, or it may determine good or bad based on the similarity between the feature vectors of the detected region image and the similar normal region image (regardless of user input). Furthermore, the user may correct the learning status determined by the image distribution display unit 105 using buttons 605 or 606.

[0058] In the embodiments described above, an example configuration in which both the detected region image and the similar normal region image are used has been explained, but only the detected region image may be used. In this case, the similar normal region image creation unit 1033 is not necessary.

[0059] In the above embodiments, the detection region image creation unit 103, the image similarity calculation unit 104, and the image distribution display unit 105 (and each of the functional units they comprise) can be configured by hardware such as circuit devices that implement these functions, or by a computing device such as a CPU (Central Processing Unit) executing software that implements these functions.

[0060] 100: Learning support device 101: Training image data 102: Detection region coordinate data 103: Detection region image creation unit 104: Image similarity calculation unit 105: Image distribution display unit

Claims

1. A learning support device that assists in learning a classification model for classifying abnormalities in an object using an image of the object, comprising: a detection region image creation unit that creates a set of detection region images including detection regions for checking for the presence or absence of abnormalities from training image data used for machine learning; an image similarity calculation unit that calculates the similarity between detection region images included in the set of detection region images; and an image distribution display unit that displays the distribution of the similarity and presents candidate annotation regions among the detection regions that are candidates for inputting annotations to be used in machine learning, wherein the image similarity calculation unit updates the feature extraction model used by the image similarity calculation unit to calculate the similarity so that the similarity between identical detection regions is high, and updates the feature extraction model so that the similarity between two image regions containing different detection regions is low.

2. The learning support device according to claim 1, wherein the detection region image creation unit acquires detection region coordinate data that describes the coordinates of the detection region in the learning image data, and the detection region image creation unit creates the detection region image set based on the coordinates described in the detection region coordinate data.

3. The learning support device according to claim 1, wherein the detection region image creation unit obtains a detection region specification parameter that defines the size of the detection region, and creates the detection region image set according to the detection region specification parameter, wherein the detection region specification parameter defines the image size of the detection region and defines the size of the region shared between different detection region images that include the same detection region.

4. The learning support device according to claim 1, characterized in that the detection region image creation unit identifies a normal region in the training image data whose surrounding features are similar to the detection region and in which no abnormality has been detected; the detection region image creation unit creates a set of similar normal region images by extracting a partial image including the normal region from the training image data; and the image similarity calculation unit calculates the similarity between the detection region image included in the set of detection region images and the similar normal region image included in the set of similar normal region images.

5. The learning support device according to claim 4, wherein the detection region image creation unit acquires a similar normal region designation parameter that defines the size of the normal region, and creates a similar normal region image set according to the similar normal region designation parameter, and the similar normal region designation parameter defines the size of the normal region and defines a threshold used to determine whether the surrounding features are similar to the detection region.

6. The learning support device according to claim 4, characterized in that the detection region image creation unit creates the detection region image set and the similar normal region image set by applying at least one of the following as preprocessing to the detection region image and the similar normal region image: color jittering, grayscale conversion, Gaussian blur, horizontal flipping, vertical flipping, random cropping, and resizing.

7. The learning support device according to claim 6, wherein the detection region image creation unit acquires an image enhancement parameter that defines the probability of applying the preprocessing, and the detection region image creation unit applies one of the preprocessing steps according to the probability defined by the image enhancement parameter.

8. The learning support device according to claim 4, characterized in that the image similarity calculation unit extracts feature vectors from the detected region image and the similar normal region image, and calculates the similarity between the feature vectors to calculate the similarity between the detected region image and the similar normal region image.

9. The learning support device according to claim 8, characterized in that the image similarity calculation unit selects one of the detection regions from which the feature vectors have been extracted as a first detection region, the image similarity calculation unit extracts two feature vectors derived from the image including the first detection region as a first feature vector and a second feature vector, the image similarity calculation unit calculates a first similarity between the first feature vector and the second feature vector, and the image similarity calculation unit updates the feature extraction model so that the first similarity is increased.

10. The learning support device according to claim 8, characterized in that the image similarity calculation unit selects one of the detection regions from which the feature vectors have been extracted as a first detection region, and extracts the feature vectors derived from the image including the first detection region as a third feature vector; the image similarity calculation unit selects one of the detection regions from which the feature vectors have been extracted that is different from the first detection region as a second detection region, and extracts the feature vectors derived from the image including the second detection region as a fourth feature vector; the image similarity calculation unit calculates a second similarity between the third feature vector and the fourth feature vector; and the image similarity calculation unit updates the feature extraction model so that the second similarity becomes lower.

11. The learning support device according to claim 8, characterized in that the image similarity calculation unit selects one of the detection regions from which the feature vectors have been extracted as a first detection region, extracts the feature vectors derived from the image including the first detection region as a fifth feature vector, the image similarity calculation unit extracts the feature vectors derived from the image including the normal region corresponding to the first detection region as a sixth feature vector, the image similarity calculation unit calculates a third similarity between the fifth feature vector and the sixth feature vector, and the image similarity calculation unit updates the feature extraction model so that the third similarity becomes lower.

12. The learning support device according to claim 8, characterized in that the image similarity calculation unit selects one of the detection regions from which the feature vectors have been extracted as a first detection region, and selects one of the regions different from the first detection region as a second detection region; the image similarity calculation unit extracts the feature vector derived from the image including the normal region corresponding to the first detection region as a seventh feature vector; the image similarity calculation unit extracts the feature vector derived from the image including the normal region corresponding to the second detection region as an eighth feature vector; the image similarity calculation unit calculates a fourth similarity between the seventh feature vector and the eighth feature vector; and the image similarity calculation unit updates the feature extraction model so that the fourth similarity is increased.

13. The learning support device according to claim 1, wherein the image distribution display unit provides a user interface for displaying the distribution, and when the user inputs on the user interface that the distribution indicates that the machine learning is being performed well, the image distribution display unit presents the detected region image corresponding to the distribution as the annotation candidate region on the user interface.

14. The learning support device according to claim 13, characterized in that the image distribution display unit presents images of the detected region on the user interface as annotation candidate regions such that the similarity between the detected region and similar regions is low.

15. The learning support device according to claim 13, wherein the user interface displays a distribution map showing the feature quantities of partial images of the learning image data for each partial image, and the image distribution display unit presents selected feature quantities from the feature quantities on the user interface as annotation candidate regions on the user interface.

Images