Supervision information creation support device, image classification device, and program

Abstract

A supervision information creation support device is provided with: a category score calculation unit for calculating, for an image comprising a plurality of pixels, a category score indicating the degree of reliability of a pixel belonging to a category, for each combination of the pixel and the category, using a segmentation model; an inference image determination unit that calculates an inference image on the basis of the category score; an adjacent category calculation unit that receives input of a category of interest and calculates an adjacent category that is easily adjacent to the category of interest on the basis of the inference image and the category of interest; and a score-to-be-corrected calculation unit that determines a score to be corrected for at least one pixel on the basis of the category of interest, the adjacent category, and the category score.

Description

Technical Field

[0001] This invention relates to a supervisory information production aid, an image classification apparatus, and a program, such as a technique for reducing the cost of producing supervisory data in an image classification system with learning capabilities. Background Technology

[0002] In creating image classification models that use deep learning for each pixel, supervised information is required, but providing such information is costly. As a countermeasure, interactive segmentation exists where the user indicates the correction location and content based on the model's inference results, allowing for learning while coordinating with the user. As a method for selecting correction locations, consider choosing locations with low inference probabilities among those inferred as belonging to the category of interest, but avoid selecting locations not inferred as belonging to the category of interest. Therefore, excluding false negatives does not improve accuracy.

[0003] Existing technical documents

[0004] Patent documents

[0005] Patent Document 1: International Publication No. 2020 / 129235

[0006] Patent Document 2: International Publication No. 2020 / 189269 Summary of the Invention

[0007] The problem that the invention aims to solve

[0008] In Patent Document 1, the tendency for the inference probability to decrease is used to set the parts with low inference probability of the model as candidates for correction parts. However, sometimes the selection of misidentified parts that are unrelated to the category of interest makes it difficult to improve the accuracy of the category of interest.

[0009] In Patent Document 2, parts with low inference probability among the parts classified as categories of interest are set as candidates for correction parts, but parts that are not inferred as categories of interest are not selected. Therefore, the accuracy is not improved by not selecting parts that are false negatives.

[0010] Therefore, the purpose of this invention is to indicate the areas to be corrected that can be expected to improve the accuracy of the category of interest, including false negative areas.

[0011] Methods for solving problems

[0012] An example of the supervisory information production aid device of the present invention includes:

[0013] The category score calculation unit, for an image composed of multiple pixels, uses a segmentation model to calculate a category score representing the reliability of a pixel belonging to that category, based on the combination of pixels and categories.

[0014] The inference image determination unit calculates the inference image based on the category score;

[0015] An adjacency category calculation unit accepts input of the category of interest and calculates, based on the inference image and the category of interest, the adjacency categories that are likely to be adjacent to the category of interest; and

[0016] The score calculation unit needs to correct the score for at least one pixel based on the attention category, the adjacent category, and the category score.

[0017] An example of the image classification device of the present invention includes:

[0018] An image category inference unit, for an image composed of multiple pixels, uses a segmentation model to calculate a category score representing the reliability that a pixel belongs to a category, based on a combination of pixel and category; and calculates an inferred image based on the category score; and

[0019] The GUI for creating monitoring information includes a section for setting interest categories.

[0020] The attention category setting unit accepts the input of attention categories.

[0021] The image classification device calculates the parts to be corrected based on the category of interest and the category score.

[0022] The monitoring information is generated by a GUI that outputs the part to be corrected, and the monitoring information is then processed by the monitoring information assignment department.

[0023] One example of the program of the present invention enables a computer to function as an auxiliary device for producing the aforementioned supervisory information.

[0024] One example of the program of the present invention enables a computer to function as the aforementioned image classification device.

[0025] Invention Effects

[0026] According to the present invention, it is possible to indicate the parts that need to be corrected in relation to improving the accuracy of false negatives for categories of interest. Attached Figure Description

[0027] Figure 1 This is a functional structure example of the monitoring information production auxiliary device of Embodiment 1 of the present invention.

[0028] Figure 2 This is an example of the processing by the adjacency category calculation department.

[0029] Figure 3 This is an example of the processing result from the adjacency category calculation unit.

[0030] Figure 4 This is an example of a section that needs to be revised in the score calculation department.

[0031] Figure 5 This is another example of a method for calculating uncertainty.

[0032] Figure 6 This is another example of a method for calculating the score that needs to be corrected.

[0033] Figure 7 This is yet another example of a method for calculating and correcting scores.

[0034] Figure 8 This is a functional structure example of the image classification device in Embodiment 2 of the present invention.

[0035] Figure 9 This is a functional structure example of the image classification device in Embodiment 3 of the present invention.

[0036] Figure 10 This is another example of the functional structure of the image classification device in Embodiment 3 of the present invention.

[0037] Figure 11 This is an example of creating a GUI using supervisory information.

[0038] Figure 12 This is another example of creating a GUI using supervisory information. Detailed Implementation

[0039] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0040] [Example 1]

[0041] Figure 1 This section illustrates a functional structure example of the supervisory information production aid device according to Embodiment 1 of the present invention. Firstly, regarding... Figure 1 The functional structure of the supervision information production aid 100 is described in summary. It takes the input image D1, the segmentation model D2, and the interest category D4 as inputs, and has: a category score calculation unit 101, an inference image determination unit 102, an adjacent category calculation unit 103, and a score correction calculation unit 104.

[0042] The category score calculation unit 101, for an input image D1 consisting of multiple pixels, uses a segmentation model D2 to calculate a category score D3, representing the reliability of a pixel belonging to a category, based on the combination of pixels and categories. The category score D3 is a vector, per pixel, for the input image D1, containing values ​​representing the height of the probability of each category being identified.

[0043] The inference image determination unit 102 calculates the inference image D5 based on the category score D3. The inference image D5 is an image that determines the category as one pixel by inference. Specifically, it is calculated pixel by pixel with respect to the category score D3 by setting the category with the maximum value as the recognition category.

[0044] The adjacency category calculation unit 103 receives the input of the interest category D4 and calculates the adjacency category D6 based on the inference image D5 and the interest category D4. The adjacency category refers to the category that is likely to be adjacent to the interest category (i.e., the category with a high likelihood of adjacency), and specific examples will be described later.

[0045] The score calculation unit 104 determines the score to be corrected D7 based on the interest category D4, adjacent categories D6, and category score D3. For example, the score to be corrected is calculated for parts that are difficult to identify with respect to the interest category for the segmentation model D2, and a specific example will be described later.

[0046] The monitoring information production aid 100 can be configured using a computer, for example. The computer has a known computer hardware structure, such as an arithmetic unit and a storage unit. The arithmetic unit includes, for example, a processor, and the storage unit includes, for example, storage media such as semiconductor memory devices and disk drives. Some or all of the storage media can be non-transitory storage media.

[0047] In addition, computers may also have input / output units. Input / output units may include, for example, input devices such as keyboards and mice, output devices such as monitors and printers, and communication devices such as network interfaces.

[0048] The storage unit can also store programs. The processor executes the program, and the computer functions as the supervisory information creation aid 100 of this embodiment or as a device in other embodiments.

[0049] The following is about Figure 1 The details of each structural element are described.

[0050] The input image D1 is the image of the object to which the score to be corrected is calculated. For at least one pixel (pixel or group of pixels) in this image, the score to be corrected is calculated to improve the accuracy of the segmentation model D2.

[0051] Segmentation model D2 is a pre-learned model that has undergone machine learning and possesses image recognition performance. However, the pre-learning can be implemented using a dataset with features similar to the input image D1, or it can be implemented using a dataset unrelated to the input image D1.

[0052] Segmentation model D2 calculates feature values ​​based on the pixels of the input image D1, and then calculates the reliability for class identification based on these feature values. The reliability is sometimes expressed as a probability value through pixel-wise normalization using the Softmax function, and sometimes it is expressed using the logit function before Softmax normalization; either approach is acceptable.

[0053] Category D4 represents the categories identified by the segmentation model that should be prioritized for improved accuracy. For example, in a semiconductor measurement image, it represents the category indicating the structure of the object being measured.

[0054] Figure 2 This illustrates an example of the processing by the adjacency category calculation unit. As processing step S201, the adjacency category calculation unit 103 first extracts adjacent pixel groups.

[0055] Here, firstly, the adjacency category calculation unit 103 extracts pixels belonging to the interest category from the pixels of the inference image D5. Then, for each pixel inferred to belong to the interest category, the pixels adjacent to that pixel are extracted as adjacent pixels (however, if the adjacent pixel also belongs to the interest category, that pixel is not extracted as an adjacent pixel).

[0056] Next, as processing step S202, the adjacency category calculation unit 103 determines the adjacency category. Here, for adjacent pixel groups, the adjacency category calculation unit 103 calculates the frequency of the category in the inference image D5, and determines the category with high frequency as the adjacency category for the category of interest D4. Thus, the category that is likely to be adjacent to the category of interest D4 is determined. Here, if there are multiple high-frequency categories, multiple adjacency categories are also considered for determination. Thus, the category that is likely to be adjacent to the category of interest D4 can be calculated.

[0057] Thus, in this embodiment, a category that is likely to be adjacent to category D4 of interest refers to a category that has a high frequency in the inference image D5 as the category to which pixels adjacent to pixels belonging to category D4 of interest belong. The definition of "high frequency" can be appropriately determined by those skilled in the art, for example, it may refer to a situation that exceeds a predetermined threshold.

[0058] Figure 3 This represents an example of the processing result from the adjacency category calculation unit. Figure 3 This represents an example where an inferred image is taken as input, with each pixel identified as one of categories 1, 2, or 3, and the category of interest is set to category 1.

[0059] First, the pixel group adjacent to the part inferred as category 1 in S201 is calculated as the adjacent pixel group 300. Next, in S202, the frequency of the category to which the adjacent pixel group 300 belongs is calculated, and as the category with the higher frequency, category 2 is determined as the adjacent category.

[0060] For example, if the category of the structure of the measurement object in a semiconductor measurement image represents a structure that is repeatedly the same within the image, then the categories adjacent to the specific category are defined. Therefore, the high-frequency category can be set as the adjacent category by the method described above.

[0061] Additionally, for example, when the white lines of a road image are set as the category of interest, the asphalt portion of the road, being a frequently occurring adjacent category, is registered as an adjacent category to the asphalt portion using the method described above.

[0062] Figure 4 This indicates an example of how to correct the score calculation unit 104. To correct the score calculation unit 104, in processing step S401, the unit first extracts the category score of the category of interest D4 and the category score of the adjacent category D6 for each pixel in the inference image D5 that is inferred to belong to the category of interest D4 and each pixel in the inference image D5 that is inferred to belong to the adjacent category D6.

[0063] Then, based on the extracted category scores, the score calculation unit 104 calculates uncertainty as a value representing the degree to which the segmentation model is not sufficiently learned (e.g., contains insufficiently learned features) in processing step S402. For example, the uncertainty is calculated as the size of the category score for a specified category (focus category D4 and adjacent category D6) in the extracted category grouping.

[0064] For example, when the category score is an inferred probability, 1 - inferred probability (the value obtained by subtracting the inferred probability from 1) is set as the uncertainty value. Generally, in a learned segmentation model, pixels with sufficiently learned features have a higher inferred probability, while pixels with insufficiently learned features have a lower inferred probability. Therefore, it can be said that the parts with a high 1 - inferred probability (parts with high uncertainty) are the parts that are insufficiently learned.

[0065] By inputting the category of interest and the adjacent category as specified categories, the uncertainty of the parts inferred to be in the category of interest and the uncertainty of the parts inferred to be in the adjacent category are obtained. Then, based on these uncertainties, the parts with high uncertainty that need to be corrected are identified in processing step S403, and the correction score of the identified parts to be corrected is calculated.

[0066] The area to be corrected can be, for example, defined as the area where the uncertainty exceeds a specified threshold. The specific calculation method for the corrected score can be appropriately designed by those skilled in the art; for example, it could be the sum of the uncertainty of the category of interest and the uncertainty of adjacent categories.

[0067] Regions with high uncertainty values ​​indicate regions that are difficult for the segmentation model to identify. By prioritizing these regions in the learning process, the accuracy of the segmentation model can be expected to improve. Therefore, based on regions of interest that are inferred to be likely to be false positives, the score to be corrected can be calculated using candidates from neighboring regions that are inferred to be likely to be false negatives.

[0068] Figure 5 Another example of a method for calculating uncertainty. For example... Figure 5 As shown, as a method for calculating uncertainty related to adjacent categories, it is also considered to set the size of the category score for the category of interest as uncertain in the inferred part belonging to the adjacent category.

[0069] exist Figure 5 In the example, in processing step S501, the score calculation unit 104 extracts category groups for each pixel inferred to belong to an adjacent category in the inference image. Then, in processing step S502, uncertainty (second uncertainty) is calculated based on the category score related to the category of interest in the extracted category groups. The second uncertainty can be, for example, set as the magnitude of the category score related to the category of interest.

[0070] Furthermore, regarding the parts that are inferred to belong to the category of interest, they can be set as... Figure 4 The same calculation method applies. That is, the score calculation unit 104 calculates the uncertainty (first uncertainty) based on the category score related to the category of interest for each pixel in the inference image that is inferred to belong to the category of interest.

[0071] In this case, the score calculation unit 104 calculates the score to be corrected based on the first uncertainty and the second uncertainty.

[0072] In areas of false negatives, the input image sometimes contains features of the category of interest; that is, in such areas, the category of interest may have a high score even if the category with the highest score is an adjacent category. Therefore, within adjacent categories, the score to be corrected can be calculated for areas where false negatives are more likely than those of the category of interest.

[0073] In Example 1, by using the method described above, in addition to the false positives, it is also possible to calculate the correction score that is expected to improve the accuracy associated with false negatives.

[0074] Figure 6 This is another example illustrating the method for calculating the corrected score. Figure 6 In the example, in processing step S601, the score calculation unit 104 smooths the score D7 to be corrected for each pixel within a specified range containing that pixel, and calculates the part with the larger value of the smoothed score D8 to be corrected as the part to be corrected.

[0075] When a user makes corrections, it is more efficient to correct a large area at once than to correct only a single isolated pixel. Therefore, by calculating the areas with high correction scores within a specified range, efficient correction can be achieved. The specific method for determining the specified range and the specific operations used for smoothing within the specified range can be appropriately designed by those skilled in the art; for example, two-dimensional convolution with a Gaussian function can be used, or a simple averaging operation can be used.

[0076] Figure 7 This is another example illustrating the method for calculating and correcting scores. Figure 7 In the example, the score calculation unit 104 determines the area with a large score to be corrected as the area D9 to be corrected in processing step S701. The definition of the area with a large score to be corrected can be appropriately determined by those skilled in the art. For example, it can be set as the area with a score above a threshold to be corrected, or it can be set as the area with a large deviation value of the score to be corrected.

[0077] By deciding which parts to correct, users can efficiently identify the areas that receive the most corrections.

[0078] [Example 2]

[0079] In Example 2, a GUI for creating supervisory information is proposed using the score to be corrected calculated in Example 1. Descriptions of parts common to Example 1 are sometimes omitted.

[0080] Figure 8 This illustrates a functional structure example of the image classification apparatus of Embodiment 2 of the present invention. The image classification apparatus 800 also functions as a supervisory information production aid, similar to that in Embodiment 1. For example, the image classification apparatus 800 may also have... Figure 1 It has the same structure as the monitoring information production auxiliary device 100.

[0081] First of all, Figure 8 The functional structure is described in summary. The image classification device 800 takes the input image D801 and the segmentation model D802 as inputs and has an image category inference unit 801 and a supervision information production GUI 802.

[0082] The image category inference unit 801, for an input image D801 composed of multiple pixels, uses a segmentation model D802 to calculate a category score D804, representing the reliability of a pixel belonging to a category, based on the combination of pixels and categories. Based on the category score D804, it calculates the inference image D803. The inference image D803 and the category score D804 are respectively compared with... Figure 1 The inference image D5 and the category score D3 are calculated in the same way.

[0083] The Supervision Information Creation GUI 802 is a GUI designed to generate supervision information D805 for each pixel using the input image D801, the inference image D803, and the category score D804. The Supervision Information Creation GUI 802 has the following functions: simultaneously displaying the input image D801 and the inference image D803, making it easy to identify errors in the inference results; or allowing the machine to determine the areas requiring correction based on the category score D804, making it easy to select the areas to be corrected.

[0084] Figure 11 This is an example of a supervisory information creation GUI 802. The supervisory information creation GUI includes: an input image display unit that displays an input image D801; an inference image display unit that displays an inference image D803; a supervisory information assignment unit that has the function of assigning supervisory information creation; a focus category setting unit that accepts input of focus categories; and a correction area display unit that outputs (e.g., displays) information based on the input focus category, input image, and category score. Figure 1 The supervision information production aid 100 shown is based on the calculated part to be corrected; the uncertainty display unit displays the calculated uncertainty; and the model learning start unit accepts the operation for starting the learning of the segmentation model.

[0085] The uncertainty display section and the correction area display section can be displayed as reference information when assigning supervision information, or they can be left undisplayed.

[0086] Alternatively, it may have an automatic correction setting unit that accepts operations for automatically calculating correction areas.

[0087] The image classification device 800 can calculate the part to be corrected based on the category of interest and the category score, for example, the same as the supervision information production assistance device 100 in Embodiment 1.

[0088] The GUI 802, which generates supervisory information, receives and assigns supervisory information via the supervisory information assignment unit. This unit can use computer click processing or a graphics tablet to set the category for each pixel. Figure 11 In the example, the input image is displayed as the background, and the assigned supervision information is overlaid on the foreground of the input image. However, the input image can also be displayed without a background. Additionally, it can have the following functionality: correcting multiple areas to be corrected to the same category in a single operation.

[0089] The uncertainty display unit displays information that is set as a reference when assigning supervisory information. Similar to the supervisory information creation aid 100 in Embodiment 1, the image classification device 800 calculates the uncertainty representing the degree of inadequacy in the learning of the segmentation model, and the uncertainty display unit displays this uncertainty. Figure 11 In the example, the white area represents the area with high uncertainty.

[0090] Areas not selected for correction but with high uncertainty are mostly areas where the inference is incorrect. Furthermore, even if the inference is correct, the segmentation model's learning is often insufficient. Therefore, by using the uncertainty display as a reference, users can assign supervisory information to areas with relatively high uncertainty, thereby improving the accuracy of the segmentation model. However, when automatically selecting areas for correction, the uncertainty display can be omitted.

[0091] In addition, the uncertainty display section can also display separately. Figure 4 The diagram shows uncertainty within the category of interest (first uncertainty) and uncertainty within adjacent categories (second uncertainty). Additionally, in the case of multiple adjacent categories, the second uncertainty can be divided into multiple regions for display.

[0092] The part display that needs correction will show parts with high uncertainty as the parts that need correction. Figure 11 In the example, the area to be corrected is represented by a white area. Users can refer to this area when assigning supervision information. However, if the area to be corrected is automatically selected, it may not be displayed.

[0093] Figure 12 This is another example of using supervised information to create GUI802. This example is... Figure 11 This is an example of a GUI being used for defect detection in the semiconductor manufacturing process. (And...) Figure 11 The three categories are displayed in the same way. In particular, the categories "Construction 1" and "Construction 2" represent the proper construction of the semiconductor structure, and the category "Defect" represents the construction defects.

[0094] With this GUI, users can efficiently create supervisory information that can be used for defect detection in semiconductor fabrication.

[0095] [Example 3]

[0096] In Example 3, an image classification apparatus is proposed for the case of interactive segmentation in which segmentation models are learned from each other and user annotations are performed, and the segmentation model is being learned midway. Descriptions of parts common to Examples 1 or 2 are sometimes omitted.

[0097] Figure 9 This illustrates a functional structure example of an image classification apparatus according to Embodiment 3 of the present invention. The image classification apparatus 900 first learns a segmentation model D802 using the supervision information D805 calculated by the method shown in Embodiment 2 and the input image.

[0098] The model learning unit 901 uses the input image D801 and the supervision information D805 to learn the segmentation model D802.

[0099] The segmentation model D802 uses a model that has been learned in advance during the input phase. This learning can utilize limited supervision information from the input image D801, or it can employ publicly available datasets.

[0100] In the image classification apparatus 900, the model learning unit 901 learns a segmentation model D802. Then, it generates supervision information D805 through processing by the image category inference unit 801, etc., and uses the supervision information D805 to train the segmentation model D802. This process is repeated until there are no misidentifications in the inference image D803, thus obtaining a segmentation model D802 with good recognition accuracy.

[0101] Figure 10 This illustrates another example of the functional structure of the image classification apparatus of Embodiment 3 of the present invention. The image classification apparatus 1000 calculates uncertainty in the same manner as in Embodiments 1 and 2, using the uncertainty of each pixel for emphasis learning.

[0102] Regions with high uncertainty are those that are not adequately learned by the segmentation model. The model learning unit 1001 teaches the segmentation model based on the uncertainty information D1001 so that the loss of pixels with high uncertainty is less than the loss of pixels with low uncertainty.

[0103] Through such learning, accuracy is expected to improve. For example, by increasing the loss per pixel proportionally to the uncertainty, and learning according to this loss during model learning, it becomes possible to learn features of areas with high uncertainty more effectively.

[0104] Symbol Explanation

[0105] 100…Supervision Information Production Auxiliary Device

[0106] 101… Category Score Calculation Department

[0107] 102…Inference of the image determination part

[0108] 103…Adjacency Class Calculation Department

[0109] 104… The score calculation section needs to be corrected.

[0110] 300…adjacent pixel group

[0111] 800…Image Classification Device

[0112] 801…Image Category Inference Department

[0113] 802… Supervision Information Creation GUI

[0114] 900…Image Classification Device

[0115] 901…Model Learning Department

[0116] 1000… Image classification device

[0117] 1001…Model Learning Department

[0118] D1… Input Image

[0119] D2…segmentation model

[0120] D3…Category Score

[0121] D4…Categories of Interest

[0122] D5…Inference Image

[0123] D6…Adjacency Class

[0124] D7…Score needs to be corrected.

[0125] D8…Score needs to be corrected.

[0126] D9… Area to be corrected

[0127] D801… Input Image

[0128] D802…Segmentation Model

[0129] D803…Inference Image

[0130] D804…Category Score

[0131] D805…Supervision Information

[0132] D1001…Uncertainty Information

[0133] S201…Processing Steps

[0134] S202…Processing Steps

[0135] S401…Processing Steps

[0136] S402…Processing Steps

[0137] S403…Processing Steps

[0138] S501…Processing Steps

[0139] S502…Processing Steps

[0140] S601…Processing Steps

[0141] S701… Processing steps.

Claims

1. A supervising information production assisting apparatus characterized by comprising: has: a category score calculation section that, for an image composed of a plurality of pixels, uses a segmentation model to calculate, for each combination of a pixel and a category, a category score that represents the reliability of the pixel belonging to the category; an inference image determination section that calculates an inference image from the category scores; a contiguous category calculation section that accepts input of a category of interest and calculates, from the inference image and the category of interest, a contiguous category that is likely to be contiguous with the category of interest; and a score-to-be-corrected calculation section that calculates, for at least one pixel, a score-to-be-corrected from the category of interest, the contiguous category, and the category scores.

2. The supervised information production assistance apparatus according to claim 1, wherein the score-to-be-corrected calculation section extracts, for each pixel inferred to belong to the category of interest in the inference image and each pixel inferred to belong to the contiguous category in the inference image, the category score of the category of interest and the category score of the contiguous category, calculates, from each of the extracted category scores, an uncertainty as a value that represents the degree of insufficiency of learning of the segmentation model, and calculates the score-to-be-corrected from the uncertainty.

3. The supervised information production assistance apparatus according to claim 1, wherein the score-to-be-corrected calculation section calculates, for each pixel inferred to belong to the category of interest in the inference image, a first uncertainty from the category score of the category of interest, calculates, for each pixel inferred to belong to the contiguous category in the inference image, a second uncertainty from the category score of the category of interest, and calculates the score-to-be-corrected from the first uncertainty and the second uncertainty.

4. The supervised information production assistance apparatus according to claim 1, wherein the score-to-be-corrected calculation section determines, as a portion to be corrected, a portion in which the score-to-be-corrected is large.

5. The supervised information production assistance apparatus according to claim 1, wherein the score-to-be-corrected calculation section, for each pixel, smoothes the score-to-be-corrected in a prescribed range that includes the pixel, and calculates, as a portion to be corrected, a portion in which the smoothed value is large. has:

6. An image classification apparatus characterized by comprising: an image category inference section that, for an image composed of a plurality of pixels, uses a segmentation model to calculate, for each combination of a pixel and a category, a category score that represents the reliability of the pixel belonging to the category, and calculates an inference image from the category scores; and a supervised information production GUI that has a category of interest setting section, the category of interest setting section accepts input of a category of interest, the image classification apparatus calculates a portion to be corrected from the category of interest and the category scores, the supervised information production GUI outputs the portion to be corrected and accepts, via a supervised information imparting section, imparting of supervised information.

7. The image classification apparatus according to claim 6, wherein the image classification apparatus has a model learning section that learns a segmentation model using the supervised information.

8. The image classification apparatus according to claim 7, wherein the image classification apparatus calculates an uncertainty that represents the degree of insufficiency of learning of the segmentation model, ​ ​ The model learning unit causes the segmentation model to learn so that a loss of a pixel with large uncertainty is smaller than a loss of a pixel with small uncertainty.

9. The image classification apparatus according to claim 6, wherein The image classification apparatus calculates uncertainty indicating a degree of insufficiency of learning of the segmentation model, and displays the uncertainty.

10. The image classification apparatus according to claim 6, wherein The image classification apparatus calculates, based on the inference image and the attention category, a contiguous category that is likely to be contiguous to the attention category, calculates, for each pixel inferred to belong to the attention category in the inference image, a first uncertainty based on a category score of the attention category, and calculates, for each pixel inferred to belong to the contiguous category in the inference image, a second uncertainty based on a category score of the contiguous category, The supervision information creation GUI displays the first uncertainty and the second uncertainty.

11. The image classification apparatus according to claim 8, wherein The supervision information creation GUI displays, as the site to be corrected, a site with large uncertainty.

12. A program characterized by causing a computer to function as the supervision information creation assistance apparatus according to claim 1.

13. A program characterized by causing a computer to function as the image classification apparatus according to claim 6. ​ ​

Images