Information processing device, relearning system, relearning method, and relearning program

Abstract

An information processing device (100) comprises: an acquisition unit (120) that acquires an estimation target image (10) and a learned model; an estimation unit (130) that performs estimation using the estimation target image (10) and the learned model; a generation unit (140) that generates a heat map (11) using the estimation result; an extraction unit (150) that extracts a plurality of characteristics on the basis of the heat map (11); and a relearning processing unit (180). The generation unit (140) uses the estimation target image (10) to execute changes with respect to each of the characteristics, and thereby generates a plurality of change images. The estimation unit (130) executes estimation using the change images (21-26) and the learned model. The generation unit (140) uses a plurality of the estimation results to generate heat maps (21a to 26a). If the grounds for estimation are mistaken, the relearning processing unit (180) uses learning data to generate relearning data in which the characteristic indicated by the grounds for estimation has been changed, and the relearning processing unit then relearns the learned model.

Description

Information processing device, relearning system, relearning method, and relearning program

[0001] The present disclosure relates to an information processing device, a relearning system, a relearning method, and a relearning program.

[0002] In the field of machine learning, HITL (Human-in-the-Loop) is known. HITL means that feedback from humans is included in the learning process. Technology related to HITL has been proposed (see Patent Document 1). The image recognition device in Patent Document 1 modifies an attention map based on human correction operations. This allows the attention map to be modified based on human knowledge.

[0003] Japanese Patent Application Laid-Open No. 2021-22368

[0004] However, there are cases where a trained model does not make an inference based on correct inference grounds, and using such a trained model is problematic.

[0005] The objective of the present disclosure is to enable a trained model to make estimations based on correct estimation grounds.

[0006] An information processing device according to one aspect of the present disclosure is provided. The information processing device includes an acquisition unit that acquires an estimation target image and a trained model, an estimation unit that performs estimation using the estimation target image and the trained model, a generation unit that generates a heat map indicating the basis of the estimation result using the estimation result, an extraction unit that extracts multiple features based on an area in the estimation target image that corresponds to the location of the basis of the estimation result indicated in the heat map and the area, and a re-learning processing unit. When the multiple features are extracted, the generation unit generates multiple modified images by making modifications to each of the features using the estimation target image. When the multiple modified images are generated, the estimation unit performs estimation using the multiple modified images and the trained model. When estimation is performed using the multiple modified images and the trained model, the generation unit generates multiple heat maps using multiple estimation results. The acquisition unit acquires training data if the estimation basis information is incorrect, based on the multiple modified images and the multiple heat maps, or based on the estimation target image, the heat map, the multiple modified images, and the multiple heat maps. The re-learning processing unit uses the training data to generate re-learning data in which the features indicated by the estimation grounds have been changed, and re-learns the trained model using the re-learning data.

[0007] According to the present disclosure, a trained model can perform estimation based on correct estimation grounds.

[0008] 1 is a diagram illustrating hardware included in an information processing device of embodiment 1. FIG. 2 is a block diagram illustrating functions of the information processing device of embodiment 1. FIG. 3 is a diagram illustrating a specific example of an estimation target image and a heat map of embodiment 1. FIG. 4 is a diagram illustrating an example of a generation process of embodiment 1. FIG. 5 is a diagram illustrating a specific example of output information of embodiment 1. FIG. 6 is a diagram illustrating a specific example (part 1) of a generation process of estimation basis information of embodiment 1. FIG. 7 is a diagram illustrating a specific example (part 2) of a generation process of estimation basis information of embodiment 1. FIG. 8 is a flowchart illustrating an example (part 1) of a process executed by the information processing device of embodiment 1. FIG. 9 is a flowchart illustrating an example (part 2) of a process executed by the information processing device of embodiment 1. FIG. 10 is a block diagram illustrating functions of an information processing device of embodiment 2. FIG. 11 is a diagram illustrating a specific example of a generation process of embodiment 2. FIG. 12 is a flowchart illustrating an example (part 1) of a process executed by the information processing device of embodiment 2. FIG. 13 is a flowchart illustrating an example (part 2) of a process executed by the information processing device of embodiment 2. FIG. 14 is a diagram illustrating a re-learning system of embodiment 3.

[0009] Hereinafter, embodiments will be described with reference to the drawings. The following embodiments are merely examples, and various modifications are possible within the scope of the present disclosure.

[0010] Embodiment 1. Fig. 1 is a diagram showing hardware included in an information processing device according to embodiment 1. The information processing device 100 is a computer. The information processing device 100 is a device that executes a relearning method. The information processing device 100 includes a processor 101, a volatile storage device 102, and a non-volatile storage device 103.

[0011] The processor 101 controls the entire information processing device 100. For example, the processor 101 is a central processing unit (CPU) or a field programmable gate array (FPGA). The processor 101 may be a multiprocessor. The information processing device 100 may also include a processing circuit.

[0012] The volatile storage device 102 is a main storage device of the information processing device 100. For example, the volatile storage device 102 is a random access memory (RAM). The nonvolatile storage device 103 is an auxiliary storage device of the information processing device 100. For example, the nonvolatile storage device 103 is a hard disk drive (HDD) or a solid state drive (SSD).

[0013] Next, a description will be given of the functions of the information processing device 100. Fig. 2 is a block diagram showing the functions of the information processing device of embodiment 1. The information processing device 100 has a storage unit 110, an acquisition unit 120, an estimation unit 130, a generation unit 140, an extraction unit 150, an output unit 160, an identification unit 170, and a relearning processing unit 180.

[0014] The storage unit 110 may be realized as a storage area secured in the volatile storage device 102 or the non-volatile storage device 103. Some or all of the acquisition unit 120, estimation unit 130, generation unit 140, extraction unit 150, output unit 160, identification unit 170, and relearning processing unit 180 may be realized by a processing circuit. Furthermore, some or all of the acquisition unit 120, estimation unit 130, generation unit 140, extraction unit 150, output unit 160, identification unit 170, and relearning processing unit 180 may be realized as program modules executed by the processor 101. For example, the program executed by the processor 101 is also referred to as a relearning program. For example, the relearning program is recorded on a recording medium.

[0015] The memory unit 110 stores various information. The acquisition unit 120 acquires an estimation target image. For example, the acquisition unit 120 acquires the estimation target image from the memory unit 110. Also, for example, the acquisition unit 120 acquires the estimation target image from an external device. For example, the external device is a cloud server. Note that the external device is omitted from the illustration. Also, for example, the acquisition unit 120 acquires an estimation target image obtained by a user's input operation. Note that the estimation target image is an image of the estimation target. The acquisition unit 120 acquires a trained model. For example, the acquisition unit 120 acquires the trained model from the memory unit 110 or an external device.

[0016] The estimation unit 130 performs estimation using the estimation target image and the trained model. In detail, the estimation unit 130 inputs the estimation target image to the trained model, and the trained model outputs an estimation result.

[0017] The generation unit 140 uses the estimation result to generate a heat map showing the basis of the estimation result. Specifically, the generation unit 140 generates the heat map using XAI (Explainable Artificial Intelligence) and the estimation result. For example, the generation unit 140 generates the heat map using Gradient-weighted Class Activation Mapping (Grad-CAM). Here, specific examples of the estimation target image and the generated heat map are shown.

[0018] 3 is a diagram showing a specific example of an estimation target image and a heat map according to embodiment 1. Fig. 3 shows an estimation target image 10 and a heat map 11. Furthermore, a range 11a indicates the location of the basis for the estimation result.

[0019] The extraction unit 150 extracts multiple features based on a region 10a in the estimation target image 10 that corresponds to the location on the basis of the estimation result. For example, the extraction unit 150 extracts a light orange color based on the region 10a. The extraction unit 150 extracts gray, which is the color of the surrounding area within the region 10a. In other words, the extraction unit 150 extracts gray, which is the color of the area other than the fingers (i.e., the background) within the region 10a. Based on the region 10a, the extraction unit 150 extracts that the brightness of the hand area is 75%. The extraction unit 150 extracts that the brightness of the surrounding area within the region 10a is 50%. The extraction unit 150 extracts high-frequency components based on the region 10a. For example, the high-frequency components are fine lines such as wrinkles. The extraction unit 150 extracts low-frequency components based on the region 10a. For example, the low-frequency components are thick lines such as fingers and hands.

[0020] Furthermore, for example, the feature may be a mid-frequency component. The extracting unit 150 can extract high-frequency components, mid-frequency components, and low-frequency components by using multi-resolution analysis. In this way, the extracting unit 150 may extract frequency components as features based on the region 10 a.

[0021] Furthermore, for example, the feature may be a line, a shape, a specific shape, etc. For example, the shape may be a circle. Furthermore, for example, the specific shape may be the shape of a finger, the shape of a hand, etc. Note that the extraction unit 150 can extract the line, shape, specific shape, etc. using a trained model, image analysis technology, etc.

[0022] When multiple features are extracted, the generation unit 140 uses the estimation target image to modify each feature, thereby generating multiple modified images. Note that a modified image is an image in which the feature has been modified or erased. The generation process will be explained using the drawings.

[0023] 4 is a diagram showing an example of the generation process of the first embodiment. The generation unit 140 uses the estimation target image 10 to generate an image of a color opposite in hue to the characteristic color. Specifically, the generation unit 140 generates a modified image 21 in which light orange (i.e., the color of the hand) is changed to light blue. As a result, the modified image 21 including the light blue hand is generated. Note that by generating an image of the opposite color, the information processing device 100 can eliminate the influence of the characteristic.

[0024] The generation unit 140 uses the estimation target image 10 to generate an image having the same color as the surrounding color in the region 10a. Specifically, the generation unit 140 generates a modified image 22 having the same color as the surrounding color in the region 10a (i.e., gray). In this way, a modified image 22 including a gray hand is generated. Note that, as will be described later, by generating the modified image 22, the information processing device 100 can confirm whether the surrounding color in the region 10a is the basis for estimation.

[0025] The generation unit 140 uses the estimation target image 10 to generate an image with a brightness opposite to the brightness of the characteristic. Specifically, the generation unit 140 generates a modified image 23 with a brightness (25% = 100 - 75) opposite to the brightness of the characteristic (i.e., 75%). This generates a modified image 23 with a brightness of 25% for the hand. Note that by generating an image with the opposite brightness, the information processing device 100 can eliminate the influence of the characteristic.

[0026] The generation unit 140 uses the estimation target image 10 to generate an image with the same brightness as the brightness of the surrounding area within the region 10a. Specifically, the generation unit 140 generates a modified image 24 with the same brightness as the brightness of the surrounding area within the region 10a (i.e., 50%). This generates a modified image 24 with a brightness of 50% for the hand. Note that, as will be described later, by generating the modified image 24, the information processing device 100 can confirm whether the brightness of the surrounding area within the region 10a is the basis for estimation.

[0027] The generation unit 140 uses the estimation target image 10 to generate a modified image 25 from which characteristic high-frequency components have been removed. The generation unit 140 uses the estimation target image 10 to generate a modified image 26 from which characteristic low-frequency components have been removed.

[0028] Here, "removal" includes the meaning of weakening frequency components. As described below, the modified image is input to a trained model. If the trained model does not read frequency components, the frequency components do not need to be removed. Therefore, the generation unit 140 may generate a modified image in which the frequency components are weakened.

[0029] The estimation unit 130 performs estimation using a plurality of modified images and a trained model. Specifically, the estimation unit 130 inputs each of the plurality of modified images to the trained model, and the trained model outputs a plurality of estimation results.

[0030] The generating unit 140 generates a plurality of heat maps using a plurality of estimation results. In particular, the generating unit 140 generates the plurality of heat maps using a plurality of estimation results and XAI.

[0031] The output unit 160 outputs the estimation target image 10, the heat map 11, the multiple modified images, and the multiple heat maps. For example, the output unit 160 outputs the estimation target image 10, etc. to a display of the information processing device 100. Also, for example, the output unit 160 outputs the estimation target image 10, etc. to an external device. Here, examples of the estimation target image 10, the heat map 11, the multiple modified images, and the multiple heat maps are shown.

[0032] Fig. 5 is a diagram showing a specific example of output information according to embodiment 1. Fig. 5 shows an estimation target image 10, a heat map 11, modified images 21 to 26, and heat maps 21a to 26a.

[0033] For example, by looking at the output information displayed on the display, the user can analyze that the high-frequency components (i.e., wrinkles) are the basis for the estimation. That is, among the heat maps 21a to 26a, only the heat map 25a does not show the location of the basis for the estimation result. The high-frequency components have been removed from the modified image 25. Therefore, the heat map 25a does not show the location of the basis for the estimation result. In other words, since the modified images 21 to 24 and 26 contain high-frequency components, the heat maps 21a to 24a and 26a show the location of the basis for the estimation result. Therefore, the user can analyze that the high-frequency components are the basis for the estimation.

[0034] The output unit 160 may also output estimation basis information. The estimation basis information is generated based on the estimation target image, a heat map corresponding to the estimation target image, multiple modified images, and multiple heat maps. The output of the estimation basis information will be described below.

[0035] 6 is a diagram showing a specific example (part 1) of the process of generating estimation basis information according to the first embodiment. Fig. 6 shows a modified image 30 and a heat map 30a corresponding to the modified image 30. The identifying unit 170 identifies, from among a plurality of heat maps corresponding to a plurality of modified images, the heat map that is most different from the heat map 11 corresponding to the image to be estimated 10. This identifies the heat map 30a.

[0036] The generation unit 140 generates an image 40 indicating the estimation basis based on the difference between the estimation target image 10 and the modified image 30 corresponding to the heat map 30a. The image 40 is also referred to as estimation basis information. The output unit 160 outputs the estimation target image, the heat map corresponding to the estimation target image, the multiple modified images, the multiple heat maps corresponding to the multiple modified images, and the image 40. Alternatively, the output unit 160 may output only the image 40.

[0037] The information processing device 100 outputs the image 40. Therefore, the user can easily identify the basis for the estimation by visually checking the image 40.

[0038] Furthermore, the information processing device 100 may output the estimation basis information by the following process, which will be described with reference to the drawings.

[0039] 7 is a diagram showing a specific example (part 2) of the process for generating estimation basis information according to the first embodiment. FIG. 7 shows a modified image 51 and a heat map 51a corresponding to the modified image 51. FIG. 7 also shows a modified image 52 and a heat map 52a corresponding to the modified image 52. The identification unit 170 identifies, from among the heat maps corresponding to the modified images, multiple heat maps that differ from the heat map 11 corresponding to the estimation target image 10. This identifies the heat maps 51a and 52a. The multiple heat maps are also referred to as multiple specified heat maps.

[0040] The generation unit 140 generates a plurality of images based on the difference between the estimation target image 10 and a plurality of modified images corresponding to the identified plurality of heat maps. For example, the generation unit 140 generates an image 61 indicating the estimation basis based on the difference between the estimation target image 10 and the modified image 51 corresponding to the heat map 51a. Furthermore, for example, the generation unit 140 generates an image 62 indicating the estimation basis based on the difference between the estimation target image 10 and the modified image 52 corresponding to the heat map 52a. The images 61 and 62 are also referred to as estimation basis information.

[0041] Furthermore, the generation unit 140 generates a plurality of images based on the difference between the heat map 11 corresponding to the estimation target image 10 and the identified plurality of heat maps. For example, the generation unit 140 generates an image 61a based on the difference between the heat map 11 and the heat map 51a. For example, the generation unit 140 generates an image 62a based on the difference between the heat map 11 and the heat map 52a. The plurality of images are also referred to as a plurality of estimation basis degree images.

[0042] The output unit 160 outputs the estimation basis information. For example, the output unit 160 outputs images 61 and 62. By visually checking the images 61 and 62, the user can recognize that there are multiple estimation bases.

[0043] Furthermore, the output unit 160 outputs a plurality of estimation basis degree images. For example, the output unit 160 outputs images 61a and 62a. The user can recognize the degree of estimation basis by visually checking the images 61a and 62a. For example, region 61a1 of image 61a is a region indicating the estimation basis. Region 62a1 of image 62a is a region indicating the estimation basis. The color of region 62a1 is darker than the color of region 61a1. Therefore, the user can recognize that the estimation basis indicated by image 62 is stronger than the estimation basis indicated by image 61.

[0044] Here, the user can recognize the basis for the estimation by visually checking the output information of the information processing device 100. For example, as shown in FIG. 5, the user can recognize the basis for the estimation by looking at the output information displayed on the display. Also, for example, as shown in FIGS. 6 and 7, the user can recognize the basis for the estimation by looking at the output information displayed on the display. In this way, the user can recognize the basis for the estimation by visually checking the output information. Therefore, the user can determine whether the basis for the estimation is incorrect.

[0045] If the user determines that the estimation basis is incorrect, the user inputs an instruction to change the feature indicated by the estimation basis in the training data and an instruction to re-learn to the information processing device 100. As a result, the acquisition unit 120 acquires the change instruction and the re-learning instruction. The acquisition unit 120 acquires the training data. For example, the acquisition unit 120 acquires the training data from the storage unit 110 or an external device. Note that, for example, the training data is data used when generating the trained model. The relearning processing unit 180 uses the training data to generate relearning data in which the feature indicated by the estimation basis has been changed. The relearning processing unit 180 re-trains the trained model using the relearning data. As a result, the location of the estimation basis in the trained model is changed to the correct location.

[0046] The above process will be described using a specific example. Assume that the feature indicated by the estimation basis is a high-frequency component. The user determines that the estimation basis is not a high-frequency component. The user inputs an instruction to change the high-frequency component in the training data and an instruction for re-learning to the information processing device 100. The re-learning processing unit 180 uses the training data to generate re-learned data in which the high-frequency component has been changed. For example, the re-learning processing unit 180 generates re-learned data obtained by removing the high-frequency component from the training data. Also, for example, the re-learning processing unit 180 generates re-learned data obtained by weakening the high-frequency component from the training data. The re-learning processing unit 180 re-learns the trained model using the re-learned data. As a result, the trained model no longer performs estimation based on the high-frequency component. Therefore, the location of the estimation basis from the trained model is changed to the correct location.

[0047] In this way, the acquisition unit 120 acquires training data when the estimation basis is incorrect based on the multiple modified images and multiple heat maps, or the estimation basis information. The relearning processing unit 180 uses the training data to generate retrained data. The relearning processing unit 180 uses the retrained data to retrain the trained model.

[0048] Next, the processing executed by the information processing device 100 will be described using a flowchart. FIG. 8 is a flowchart showing an example (part 1) of the processing executed by the information processing device of embodiment 1. (Step S11) The acquisition unit 120 acquires an estimation target image and a trained model. (Step S12) The estimation unit 130 performs estimation using the estimation target image and the trained model. (Step S13) The generation unit 140 generates a heat map using the estimation result. (Step S14) The extraction unit 150 extracts multiple features based on an area corresponding to the location of the basis of the estimation result indicated by the heat map. Note that this area is an area within the estimation target image.

[0049] (Step S15) The generation unit 140 generates multiple modified images based on multiple features. (Step S16) The estimation unit 130 performs estimation using the multiple modified images and the trained model. (Step S17) The generation unit 140 generates multiple heat maps using multiple estimation results. (Step S18) The output unit 160 outputs the estimation target image, a heat map corresponding to the estimation target image, multiple modified images, and multiple heat maps corresponding to the multiple modified images. Note that the output unit 160 does not necessarily need to output the estimation target image and the heat map corresponding to the estimation target image. Then, the processing proceeds to step S21.

[0050] 9 is a flowchart showing an example (part 2) of processing executed by the information processing device of embodiment 1. (Step S21) The acquisition unit 120 acquires a change instruction and a re-learning instruction for the feature indicated by the estimation basis in the training data. (Step S22) The acquisition unit 120 acquires the training data. (Step S23) The relearning processing unit 180 uses the training data to generate re-learned data in which the feature indicated by the estimation basis has been changed. (Step S24) The relearning processing unit 180 uses the relearning data to re-learn the trained model.

[0051] According to embodiment 1, when the estimation basis is incorrect, the information processing device 100 generates re-learned data in which the features indicated by the estimation basis have been changed. Then, the information processing device 100 re-learns the trained model using the re-learned data. As a result, the trained model no longer performs estimation based on the features indicated by the estimation basis. In this way, the information processing device 100 re-learns the trained model so that the trained model performs estimation based on the correct estimation basis. Therefore, the information processing device 100 can cause the trained model to perform estimation based on the correct estimation basis.

[0052] Second Embodiment Next, a second embodiment will be described. In the second embodiment, differences from the first embodiment will be mainly described. Furthermore, in the second embodiment, descriptions of the commonalities between the first embodiment and the second embodiment will be omitted.

[0053] FIG. 10 is a block diagram showing the functions of an information processing device according to a second embodiment. The information processing device 100a has a generation unit 140a, an identification unit 170a, and a relearning processing unit 180a. The functions of the generation unit 140a, the identification unit 170a, and the relearning processing unit 180a will be described later. Next, the generation process performed by the information processing device 100a will be described using the drawings. FIG. 11 is a diagram showing a specific example of the generation process according to the second embodiment. FIG. 11 shows a modified image 70 and a heat map 70a corresponding to the modified image 70.

[0054] The identification unit 170a identifies, from among a plurality of heat maps corresponding to a plurality of modified images, the heat map that is most different from the heat map 11 corresponding to the estimation target image 10. In this way, the heat map 70a is identified.

[0055] The generator 140a generates, as character information 80, the features that were changed when the modified image 70 corresponding to the heat map 70a was generated. For example, the modified image 70 is an image from which high-frequency components, which are features, have been removed. Therefore, the generator 140a generates character information 80 that indicates the high-frequency components. The character information 80 is also referred to as estimation basis information.

[0056] The output unit 160 may output the text information 80. This allows the user to intuitively recognize the basis for the estimation.

[0057] Next, a process related to re-learning in the second embodiment will be described. The acquisition unit 120 acquires correct answer information indicating a correct estimation basis. For example, the acquisition unit 120 acquires the correct answer information from the storage unit 110 or an external device. The relearning processing unit 180a determines whether the estimation basis is incorrect based on the estimation basis information and the correct answer information. If the estimation basis is incorrect, the acquisition unit 120 acquires training data. The relearning processing unit 180a uses the training data to generate re-learning data in which the features indicated by the estimation basis information have been changed. For example, the relearning processing unit 180a uses the training data to generate re-learning data in which the features (e.g., high-frequency components) indicated by the estimation basis information (e.g., character information 80) have been changed. The relearning processing unit 180a uses the re-learning data to re-learn the trained model.

[0058] Next, the processing executed by the information processing device 100a will be described using a flowchart. Fig. 12 is a flowchart showing an example (part 1) of the processing executed by the information processing device of embodiment 2. The processing in Fig. 12 differs from the processing in Fig. 8 in that steps S18a and 18b are executed. Therefore, steps S18a and 18b will be described in Fig. 12. Further, a description of the processing other than steps S18a and 18b will be omitted.

[0059] (Step S18a) The identification unit 170a identifies, from among the heat maps corresponding to the multiple modified images, the heat map that is most different from the heat map 11 corresponding to the image to be estimated 10. (Step S18b) The generation unit 140a generates, as text information, the features that were changed when the modified image corresponding to the identified heat map was generated. This generates estimation basis information.

[0060] Fig. 13 is a flowchart showing an example (part 2) of the process executed by the information processing device of embodiment 2. The process of Fig. 13 differs from the process of Fig. 9 in that steps S21a, 21b, and 23a are executed. Therefore, steps S21a, 21b, and 23a will be described in Fig. 13. Description of the process other than steps S21a, 21b, and 23a will be omitted.

[0061] (Step S21a) The acquisition unit 120 acquires correct answer information. (Step S21b) The relearning processing unit 180a determines whether the estimation basis is incorrect based on the estimation basis information and the correct answer information. If the estimation basis is incorrect, the process proceeds to step S22. If the estimation basis is correct, the process ends. (Step S23a) The relearning processing unit 180a uses the learning data to generate relearning data in which the features indicated by the estimation basis information have been changed.

[0062] According to the second embodiment, when the estimation basis is incorrect, the information processing device 100a generates re-learned data in which the features indicated by the estimation basis have been changed. Then, the information processing device 100a re-learns the trained model using the re-learned data. As a result, the trained model no longer performs estimation based on the features indicated by the estimation basis. In this way, the information processing device 100a re-learns the trained model so that the trained model performs estimation based on the correct estimation basis. Therefore, the information processing device 100a can cause the trained model to perform estimation based on the correct estimation basis.

[0063] The above describes a case where one estimation basis is identified. When multiple estimation basis are identified, the information processing device 100a may generate multiple pieces of character information. Specifically, the generator 140a generates multiple pieces of character information, which are estimation basis information, based on multiple features that were changed when multiple modified images corresponding to multiple heat maps were generated. The process will be specifically described with reference to FIG. 7 . The generator 140a generates multiple pieces of character information, which are based on multiple features that were changed when modified images 51 and 52 corresponding to heat maps 51a and 52a were generated. The output unit 160 may output the multiple pieces of character information as estimation basis information. This allows the user to intuitively recognize the multiple estimation basis.

[0064] When the estimation basis information is generated, the relearning processing unit 180a determines whether or not there is an error in the estimation basis based on the estimation basis information and the correct answer information. If there is an error in the estimation basis, the acquisition unit 120 acquires training data. The relearning processing unit 180a uses the training data to generate relearning data in which, of the multiple features indicated by the multiple pieces of character information, the features whose estimation basis is incorrect have been changed. The relearning processing unit 180a uses the relearning data to retrain the trained model.

[0065] As a result, the information processing device 100a retrains the trained model using retrained data in which the feature with the erroneous estimation basis among the multiple features has been changed. Therefore, the retrained trained model no longer performs estimation based on the erroneous estimation basis. Therefore, the information processing device 100a can cause the trained model to perform estimation based on the correct estimation basis.

[0066] Embodiment 3 In the first and second embodiments, the case where the information processing device is realized by one information processing device has been described. The first and second embodiments may also be realized by a relearning system. An example of a relearning system is shown. FIG. 14 is a diagram showing a relearning system according to the third embodiment. The relearning system includes a plurality of information processing devices. For example, the relearning system includes information processing devices 200 to 205. The information processing devices 200 to 205 communicate via a network 300. The network 300 is a wired network or a wireless network. The functions of the information processing device 100 or the information processing device 100a may also be realized by a plurality of information processing devices. For example, the functions of the information processing device 100 or the information processing device 100a may also be realized by the information processing devices 200 to 205. Thus, the relearning system can realize the first and second embodiments.

[0067] The features of the above-described embodiments can be combined with each other as appropriate.

[0068] 10 Estimation target image, 10a Region, 11 Heat map, 11a Range, 21 to 26 Modified image, 21a to 26a Heat map, 30 Modified image, 30a Heat map, 40 Image, 51, 52 Modified image, 51a, 52a Heat map, 61, 62 Image, 61a, 62a Image, 61a1 Region, 62a1 Region, 70 Modified image, 70a Heat map, 80 Character information, 100 Information processing device, 100a Information processing device, 101 Processor, 102 Volatile storage device, 103 Non-volatile storage device, 110 Storage unit, 120 Acquisition unit, 130 Estimation unit, 140 Generation unit, 140a Generation unit, 150 Extraction unit, 160 Output unit, 170 Identification unit, 170a Identification unit, 180 Re-learning processing unit, 180a Re-learning processing unit, 200 to 205 Information processing device, 300 Network.

Claims

1. A computer-implemented method and apparatus for detecting an image to be estimated and a trained model, comprising: an acquisition unit that acquires an image to be estimated and a trained model; an estimation unit that performs estimation using the image to be estimated and the trained model; a generation unit that generates a heat map indicating a basis for the estimation result using an estimation result; an extraction unit that extracts a plurality of features based on an area in the image to be estimated that corresponds to a location of the basis for the estimation result indicated by the heat map and that is an area in the image to be estimated; and a re-learning processing unit, wherein when the plurality of features are extracted, the generation unit generates a plurality of changed images by making a change for each of the features using the image to be estimated; when the plurality of changed images are generated, the estimation unit performs estimation using the plurality of changed images and the trained model; when estimation is performed using the plurality of changed images and the trained model, the generation unit generates a plurality of heat maps using a plurality of estimation results; and when the estimation basis information is incorrect based on the plurality of changed images and the plurality of heat maps, or the image to be estimated, the heat map, the plurality of changed images, and the plurality of heat maps, the acquisition unit acquires learning data; the re-learning processing unit uses the learning data to generate re-learned data in which a feature indicated by the estimation ground is changed, and re-learns the trained model using the re-learned data.

2. An information processing device as described in claim 1, further comprising an output unit that outputs the multiple modified images and the multiple heat maps, or the estimated basis information, and wherein the acquisition unit acquires the learning data when an instruction to modify the features indicated by the estimated basis and an instruction to re-learn are acquired after the multiple modified images and the multiple heat maps, or the estimated basis information are output.

3. An information processing device as described in claim 1, further comprising an identification unit that identifies, from the multiple heat maps, a heat map that is most different from the heat map corresponding to the image to be estimated, wherein the generation unit generates the estimation basis information, which is text information, from features that were changed when a modified image corresponding to the identified heat map was generated, wherein the acquisition unit, when the estimation basis information is generated, acquires correct answer information indicating a correct estimation basis, wherein the re-learning processing unit determines whether the estimation basis is incorrect based on the estimation basis information and the correct answer information, wherein the acquisition unit acquires the learning data when the estimation basis is incorrect, and wherein the re-learning processing unit uses the learning data to generate the re-learning data in which the features indicated by the estimation basis information have been changed.

4. The information processing device of claim 1, further comprising an identification unit that identifies, from the multiple heat maps, multiple heat maps that differ from the heat map corresponding to the estimation target image as multiple specific heat maps, wherein the generation unit generates, as the estimation basis information which is multiple character information, multiple features that were changed when multiple changed images corresponding to the multiple specific heat maps were generated, wherein the acquisition unit, when the estimation basis information is generated, acquires correct answer information indicating a correct estimation basis, wherein the re-learning processing unit determines whether or not there is an error in the estimation basis based on the estimation basis information and the correct answer information, wherein the acquisition unit acquires the learning data if there is an error in the estimation basis, and wherein the re-learning processing unit uses the learning data to generate the relearning data in which a feature having an erroneous estimation basis among the multiple features indicated by the multiple character information is changed.

5. A re-learning system including a plurality of information processing devices, comprising: an acquisition unit that acquires an estimation target image and a trained model; an estimation unit that performs estimation using the estimation target image and the trained model; a generation unit that generates a heat map indicating a basis for the estimation result using an estimation result; an extraction unit that extracts a plurality of features based on an area in the estimation target image that corresponds to a location of the basis for the estimation result indicated by the heat map and that corresponds to the area in the estimation target image; and a re-learning processing unit, wherein when the plurality of features are extracted, the generation unit generates a plurality of changed images by making changes for each of the features using the estimation target image; when the plurality of changed images are generated, the estimation unit performs estimation using the plurality of changed images and the trained model; and when estimation is performed using the plurality of changed images and the trained model, the generation unit generates a plurality of heat maps using a plurality of estimation results; a re-learning system, wherein the acquisition unit acquires learning data when the estimation basis is incorrect based on the multiple modified images and the multiple heat maps, or on estimation basis information generated based on the estimation target image, the heat map, the multiple modified images, and the multiple heat maps; and the re-learning processing unit uses the learning data to generate re-learning data in which features indicated by the estimation basis have been modified, and re-learns the trained model using the re-learning data.

6. A re-learning method comprising: an information processing device; acquiring an estimation target image and a trained model; performing an estimation using the estimation target image and the trained model; using the estimation result to generate a heat map indicating a basis for the estimation result; extracting a plurality of features based on an area in the estimation target image that corresponds to a location of the basis for the estimation result indicated by the heat map; using the estimation target image to generate a plurality of modified images by making modifications for each of the features; performing an estimation using the plurality of modified images and the trained model; generating a plurality of heat maps using the plurality of estimation results; acquiring training data if the estimation basis is incorrect based on the plurality of modified images and the plurality of heat maps, or estimation basis information generated based on the estimation target image, the heat map, the plurality of modified images, and the plurality of heat maps; using the training data to generate re-learned data in which the features indicated by the estimation basis have been modified; and re-learning the trained model using the re-learned data.

7. A re-learning program that causes a computer to execute processes of: acquiring an estimation target image and a trained model; performing an estimation using the estimation target image and the trained model; generating a heat map indicating a basis for the estimation result using the estimation result; extracting a plurality of features based on an area in the estimation target image that corresponds to a location of the basis for the estimation result indicated by the heat map; generating a plurality of modified images by making modifications to each of the features using the estimation target image; performing an estimation using the plurality of modified images and the trained model; generating a plurality of heat maps using the plurality of estimation results; acquiring training data if the estimation basis is incorrect based on the plurality of modified images and the plurality of heat maps, or estimation basis information generated based on the estimation target image, the heat map, the plurality of modified images, and the plurality of heat maps; using the training data to generate re-learned data in which the features indicated by the estimation basis have been modified; and re-learning the trained model using the re-learned data.

Images