Teaching data generation device, teaching data generation method, and program

Abstract

To provide a teaching data generation device that efficiently generates teaching data using processing results of a recognition device that recognizes a recognition object.SOLUTION: Recognition area information that a recognition device has identified through recognition processing on a recognition object is acquired. Based on a likelihood assigned to each unit area, a recognition object area in an image indicated by the recognition area information is segmented. An image including the segmented area is generated as teaching data for generating learning data that the recognition device uses to recognize the recognition object.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a teacher data generation device, a teacher data generation method, and a program. [Background technology]

[0002] There is a technology that automatically recognizes objects in images using a learning model generated by machine learning. Such automatic recognition technology uses machine learning to learn training data to generate a learning model. The training data is correct answer data for the object to be recognized. Furthermore, training data is generated by humans with a lot of effort. For example, if the object to be recognized is an aircraft, the training data indicates the area of ​​the aircraft that appears with high accuracy in images captured in the past.

[0003] Patent Document 1 discloses a technique for a teacher data creation device that can create images (teacher data) with more accurate labels for each pixel. Patent Document 1 also discloses a technique for image segmentation using a machine learning model when creating teacher data. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2019-101535 A Summary of the Invention [Problem to be solved by the invention]

[0005] When generating the above-mentioned training data, it is desirable to be able to create the training data efficiently.

[0006] The purpose of this disclosure is to provide a teacher data generation device, a teacher data generation method, and a program that solve the above problems. [Means for solving the problem]

[0007] According to a first aspect of this disclosure, a teacher data generation device includes an acquisition means for acquiring recognition area information identified by a recognition device through a recognition process of a recognition target, the recognition area information including at least the recognition target and an image including the recognition target and its area, and a likelihood that the recognition target is the recognition target assigned to each unit area in the image, a segmentation means for segmenting the area of ​​the recognition target in the image indicated by the recognition area information based on the likelihood assigned to each unit area, and a generation means for generating an image including the segmented area as teacher data for generating learning data to be used by the recognition device for recognizing the recognition target.

[0008] According to a second aspect of this disclosure, a teacher data generation method includes acquiring recognition area information identified by a recognition device through a recognition process of a recognition target, the recognition area information including at least the recognition target and an image including the recognition target and its area, and a likelihood that the recognition target is the recognition target assigned to each unit area in the image, segmenting the area of ​​the recognition target in the image indicated by the recognition area information based on the likelihood assigned to each unit area, and generating an image including the segmented area as teacher data for generating learning data to be used by the recognition device to recognize the recognition target.

[0009] According to a third aspect of this disclosure, the program causes a computer to function as: an acquisition means for acquiring recognition area information identified by a recognition device through a recognition process of a recognition target, the recognition area information including at least an image including the recognition target and its area, and a likelihood that the recognition target is the recognition target, assigned to each unit area in the image; a segmentation means for segmenting the area of ​​the recognition target in an image indicated by the recognition area information based on the likelihood assigned to each unit area; and a generation means for generating an image including the segmented area as teacher data for generating learning data to be used by the recognition device to recognize the recognition target. Effect of the Invention

[0010] According to the present disclosure, training data can be efficiently generated using the processing results of a recognition device that recognizes a recognition target. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram illustrating an overview of a teacher data generation device according to an embodiment of the present disclosure. [Diagram 2] A diagram illustrating the hardware configuration of a teacher data generation device according to an embodiment of the present disclosure. [Diagram 3] A functional block diagram of a teacher data generation device according to one embodiment of the present disclosure. [Figure 4] FIG. 11 is a diagram showing a first method for generating related teacher data. [Diagram 5] FIG. 13 is a diagram showing a second method for generating related teacher data. [Figure 6] FIG. 2 is a diagram showing an overview of recognition area information according to one embodiment of the present disclosure. [Figure 7] A first diagram showing the processing flow of a teacher data generation device according to one embodiment of the present disclosure. [Figure 8] A second diagram showing the processing flow of the teacher data generation device according to one embodiment of the present disclosure. [Figure 9] FIG. 13 illustrates a first example of a complementation process according to an embodiment of the present disclosure. [Figure 10] FIG. 13 illustrates a second example of a complementation process according to an embodiment of the present disclosure. [Figure 11] A diagram showing one aspect of a teacher data generation device according to another embodiment of the present disclosure. [Figure 12] A diagram showing a processing flow of a teacher data generation device according to another embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] FIG. 1 is a diagram illustrating an overview of a teacher data generation device according to one embodiment of the present disclosure. The teacher data generation device 1 of this disclosure is connected to and communicates with the object detection device 100. The object detection device 100 is one aspect of a recognition device that performs recognition processing of a recognition target appearing in an image. The object detection device 100 detects a predetermined recognition target appearing in a SAR (Synthetic Aperture Radar) image by using a learning model generated in the past. The SAR image is one aspect of a satellite image. The teacher data generation device 1 generates teacher data by using recognition area information acquired from the object detection device 100. The teacher data generation device 1 outputs the teacher data to the object detection device 100. The object detection device 100 uses the teacher data acquired from the teacher data generation device 1 to generate a new learning model to be used for detecting a recognition target appearing in the SAR image. The generation of the new learning model may be performed by another device.

[0013] The object classification device 200 determines the classification of the recognition target detected by the object detection device 100. For example, the object detection device 100 detects aircraft, ships, etc. captured in a SAR image. In this case, the object classification device 200 determines the classification of the aircraft type (passenger plane, fighter plane, etc.) and the ship type (passenger ship, container ship, warship), etc. The size estimation device 300 estimates the size of the recognition target and outputs a determination result 400 that describes the information determined by the object detection device 100, the information determined by the object classification device 200, and the information determined by the size estimation device 300. The object detection device 100 may detect a recognition target such as an airliner, a fighter jet, a helicopter, a passenger ship, a container ship, or a warship, and the object classification device 200 may determine the type (model or name) of the passenger plane, the type (model or name) of the fighter jet, the type (model or name) of the helicopter, the type (model or name) of the passenger ship, the type (model or name) of the container ship, or the type (model or name) of the warship. The recognition targets are not limited to the above. The object detection device 100, the object classification device 200, and the size estimation device 300 may also perform respective determinations on recognition targets appearing in images other than SAR images.

[0014] Fig. 2 is a diagram showing a hardware configuration of a teacher data generation device according to an embodiment of the present disclosure. As shown in Fig. 2, the teacher data generation device 1 is a computer including various pieces of hardware such as a CPU (Central Processing Unit) 101, a ROM (Read Only Memory) 102, a RAM (Random Access Memory) 103, other storage devices 104, and a communication module 105. The object detection device 100, the object classification device 200, and the size estimation device 300 are also computers having similar hardware configurations.

[0015] FIG. 3 is a functional block diagram of a teacher data generation device according to one embodiment of the present disclosure. The teacher data generation device 1 starts a teacher data generation program stored in advance. As a result, the teacher data generation device 1 performs the functions of an acquisition unit 11, a segmentation unit 12, a complementation unit 13, a generation unit 14, a database 15, and a confirmation request unit 16.

[0016] The acquisition unit 11 acquires recognition area information identified by the object detection device 100 through a recognition process of a recognition target, and which includes at least an image including the recognition target and its area, and a likelihood assigned to each unit area in the image that it is the recognition target. The segmentation unit 12 segments the area of ​​the recognition target in the image indicated by the recognition area information, based on the likelihood assigned to each unit area. The complementation unit 13 complements the segmented region. The generation unit 14 generates an image including the segmented and complemented regions as teacher data for generating learning data that is used by the recognition device to recognize the recognition target. The database 15 stores information used for various processes. The confirmation request unit 16 outputs the segmentation result or the completion result based on a comparison of the probability information of the recognition target with a threshold value, and obtains corrected boundary information between the recognition target area and other areas by the segmentation process or completion process, or approval information of the result.

[0017] FIG. 4 is a diagram showing a first method for generating related teacher data. FIG. 5 is a diagram showing a second method for generating related teacher data. In the first method of generating teacher data shown in FIG. 4, a rectangular area 41 including an aircraft as a recognition target is set as teacher data. In this case, areas other than the aircraft are also included. When a recognition target is recognized by a learning model obtained by machine learning using such teacher data, information of a range other than the recognition target is also learned, so there is a concern that the recognition accuracy of the recognition target may decrease. In the method of generating teacher data shown in FIG. 5, the outline of the aircraft is specified by a person, and the area surrounded by the outline 51 is set as teacher data. Depending on the state of the recognition target shown in the image, the boundary between the recognition target area and other areas may be difficult to distinguish, and such a method of generating teacher data requires a lot of effort and causes variation in the designation of the outline depending on the person. Therefore, the teacher data generating device 1 of the present disclosure uses recognition area information output by the object detection device 100, which is a recognition device, and includes at least the recognition target and an image including the area, and the likelihood that the recognition target is the recognition target assigned to each unit area in the image.

[0018] FIG. 6 is a diagram showing an overview of recognition area information according to one embodiment of the present disclosure. As shown in FIG. 6, the recognition area information output by the object detection device 100 includes an image of an area including the recognition target of the SAR image, to which the likelihood of being a recognition target is assigned for each pixel or for each pixel area of ​​an arbitrary size, such as 10 pixels vertically and horizontally. This image includes a surrounding area of ​​the recognition target. The recognition area information may include a label (identification information such as an aircraft or a ship) of the recognition target detected in the SAR image by the object detection device 100, and the probability that the recognition target is the recognition target indicated by the label. FIG. 6 is a diagram showing whether the likelihood is high or low depending on the density of a pixel. A pixel or a pixel area is an example of a unit area. The recognition area information is a result of the object detection device 100 performing a recognition process of a recognition target appearing in the SAR image using a learning model generated in the past. The learning model may be obtained by learning teacher data generated by the generation method shown in FIG. 4 or FIG. 5. The calculation of the likelihood that an object is a recognition target, which is assigned by object detection device 100 to each pixel or pixel region in the recognition area information, may be a value calculated by object detection device 100 using a known technique during the object recognition process, and the likelihood may correspond to the ship index value described in paragraph 0026 of Japanese Patent No. 7056751, for example. The calculation of the probability of the label of the recognition target, which is assigned by object detection device 100 to the recognition area information, may also be a value calculated by object detection device 100 using a known technique during the object recognition process.

[0019] FIG. 7 is a first diagram showing a processing flow of a teacher data generation device according to one embodiment of the present disclosure. Next, the processing flow of the teacher data generating device 1 will be explained step by step. The acquisition unit 11 of the teacher data generating device 1 acquires the recognition area information from the object detection device 100 (step S101). The segmentation unit 12 reads the likelihood of each unit area in the image included in the recognition area information. Based on the likelihood of each unit area and the segmentation threshold, the segmentation unit 12 identifies the boundary between the unit area showing the likelihood equal to or greater than the segmentation threshold and the unit area showing the likelihood less than the segmentation threshold. The segmentation unit 12 identifies a group of unit areas showing the likelihood equal to or greater than the segmentation threshold as a recognition target candidate area (step S102). This completes the segmentation process. The recognition target candidate area identified by segmentation may be a state in which unit areas showing the likelihood equal to or greater than the segmentation threshold are gathered together, or a plurality of unit areas showing the likelihood equal to or greater than the segmentation threshold are present apart from each other in the image. The segmentation unit 12 outputs the segmentation result to the complementation unit 13.

[0020] The complementing unit 13 acquires the segmentation result. The complementing unit 13 judges whether the likelihood of an area determined not to be a recognition target candidate area because the likelihood is less than the segmentation threshold in the segmentation result is equal to or greater than the complementing threshold. When the likelihood of an area determined not to be a recognition target candidate area because the likelihood is less than the segmentation threshold is equal to or greater than the complementing threshold (for example, segmentation threshold>complementing threshold), the complementing unit 13 judges that the area is included in the recognition target candidate area (step S103). This is an example of a complementing process (closing process). The complementing unit 13 may perform the closing process by other known methods. The complementing unit 13 may compare the area with the correct answer pattern of the recognition target and perform a complementing process (closing process) to include an area determined not to be a recognition target candidate area because the likelihood is less than the segmentation threshold in the recognition target candidate area. The complementing unit 13 may perform a known opening process as a complementing process. In the opening process, some of the regions determined to be candidates for the recognition target by comparison with the correct pattern of the recognition target may be changed to regions other than the recognition target. The complementation unit 13 outputs the regions of the recognition target candidates identified as a result of the complementation process to the generation unit 14.

[0021] The generation unit 14 generates teacher data including an image area consisting of pixels only of the area determined by the segmentation unit 12 and the completion unit 13 to be a recognition target candidate, and a label (step S104). The generation unit 14 transmits the teacher data to the object detection device 100 (step S105). The teacher data generation device 1 similarly repeatedly generates teacher data for the same or different recognition targets, and transmits it to the object detection device 100. As a result, the object detection device 100 performs machine learning using the multiple pieces of teacher data acquired from the teacher data generation device 1, and generates a learning model for recognizing each recognition target. The learning model may be generated by another device. The object detection device 100 performs the next recognition process for the recognition target using the newly generated learning model.

[0022] According to the above-mentioned process, the efficiency of generating teacher data can be improved because the teacher data generation device 1 can generate teacher data without human intervention by using the recognition area information output by the object detection device 100. Furthermore, according to the above-mentioned process, the teacher data generation device 1 can automatically generate teacher data including only the area of ​​the recognition target by the segmentation process and the complementation process, so that a learning model for more accurately recognizing the recognition target can be generated by generating a learning model using the teacher data.

[0023] FIG. 8 is a second diagram showing the processing flow of the teacher data generation device according to one embodiment of the present disclosure. Here, after the completion of the complementation process in step S103, the confirmation request unit 16 of the teacher data generation device 1 may determine whether the probability of the label of the recognition target included in the recognition area information is equal to or greater than a threshold. If the probability of the label is less than the threshold, the confirmation request unit 16 requests the user to confirm the result of the complementation process. For example, the confirmation request unit 16 acquires the recognition area information from the acquisition unit 11. The confirmation request unit 16 compares the probability of the label included in the recognition area information with the label threshold (step S201). If the probability of the label included in the recognition area information is less than the label threshold, the possibility that the recognition target recognized in the SAR image matches the recognition target indicated by the label is low. In such a case, the confirmation request unit 16 receives confirmation of the result of the complementation process from a user such as an administrator.

[0024] When the probability of the label included in the recognition area information is less than the label threshold, the confirmation request unit 16 outputs image information indicating the area determined to be a recognition target candidate as a result of the complementation process to a monitor or the like (step S202). The user may confirm the area determined to be a recognition target candidate using an input device such as a mouse, and may input information to correct the area determined to be a recognition target candidate. Alternatively, the user may input information indicating approval of the area determined to be a recognition target candidate using an input device such as a mouse or a keyboard. The input may be performed using a web interface function of the teacher data generation device 1. That is, the confirmation request unit 16 may output to the user's terminal an image information web page indicating an image area consisting of pixels of only the area determined to be a recognition target candidate as a result of the complementation process. The web page may be provided with an approval button or an editing function that allows the area of ​​the recognition target candidate to be changed. The confirmation request unit 16 acquires an image indicating the area of ​​the recognition target candidate based on the boundary information of the corrected area, and approval information based on the user's input (step S203). When the confirmation request unit 16 acquires the approval information, it outputs the complementation result by the complementation unit 13 to the generation unit 14. The generation unit 14 generates training data including an image area consisting of pixels only of the area determined to be the approved recognition target candidate, and a label (step S104).

[0025] When the confirmation request unit 16 receives an image showing the area of ​​the recognition-target candidate based on the boundary information of the corrected area, the confirmation request unit 16 outputs the area of ​​the recognition-target candidate shown in the image to the generation unit 14. In this case, the generation unit 14 generates training data including an image area made up of pixels of only the area of ​​the recognition-target candidate after the correction, and a label.

[0026] The processing of the confirmation request unit 16 is an example of processing for outputting the result of the complementation processing based on a comparison of the probability information of the recognition target with a threshold value, and obtaining corrected boundary information between the area of ​​the recognition target and other areas by the complementation processing, or approval information of the result.

[0027] According to this type of processing, even if the label determined for the recognition target by the object detection device 100 is unlikely to match the recognition target, it is possible to create training data to be used for generating a learning model for recognizing the recognition target with greater accuracy after user confirmation.

[0028] FIG. 9 is a diagram illustrating a first example of the complementation process according to an embodiment of the present disclosure. 9, the likelihood of unit area 81 and unit area 82 is equal to or less than the segmentation threshold, so they are not determined as recognition target candidate areas in segmentation. However, the completion unit 13 determines that unit area 81 and unit area 82 are recognition target candidate areas because their likelihood is equal to or greater than the completion threshold.

[0029] FIG. 10 is a diagram illustrating a second example of the complementation process according to an embodiment of the present disclosure. 10, the likelihoods of unit area 81, unit area 82, and unit area 91 are below the segmentation threshold, so they are not determined as recognition target candidate areas in segmentation. The completion unit 13 determines that unit area 81 and unit area 82 are recognition target candidate areas because their likelihoods are above the completion threshold. The completion unit 13 may also determine that unit area 91 is included in the recognition target candidate areas based on a comparison with the correct answer data.

[0030] FIG. 11 is a diagram showing one aspect of a teacher data generation device according to another embodiment of the present disclosure. FIG. 12 is a diagram showing a processing flow of a teacher data generation device according to another embodiment of the present disclosure. The teacher data generation device 1 includes an acquisition unit 11, a segmentation unit 12, and a generation unit 14. The acquisition unit 11 acquires recognition area information identified by the recognition device through a recognition process of a recognition target, the recognition area information including at least the recognition target and an image including the recognition target and its area, and a likelihood that each unit area in the image is the recognition target (step S301). The segmentation unit 12 segments the area of ​​the recognition target in the image indicated by the recognition area information, based on the likelihood assigned to each unit area (step S302). The generating unit 14 generates an image including only the segmented region as teacher data for generating learning data to be used by the recognition device for recognizing the recognition target (step S303).

[0031] In the above description, the teacher data generation device 1 generates teacher data used by the object detection device 100 to generate a learning model for object detection. However, the teacher data generation device 1 may generate teacher data used by the object classification device 200 to generate a learning model for object classification by a similar process. In this case, the teacher data generation device 1 acquires recognition area information that is specified by the object classification device 200 through a recognition process of the recognition target, and that includes at least an image including the recognition target and its area, and a likelihood that the recognition target is the recognition target assigned by the object detection device 100 for each unit area in the image. Then, the teacher data generation device 1 segments the area of ​​the recognition target in the image indicated by the recognition area information based on the likelihood assigned to each unit area, and generates an image including the segmented area as teacher data for generating learning data used by the object classification device 200 to recognize the recognition target.

[0032] Alternatively, the teacher data generation device 1 may generate teacher data used by the size estimation device 300 to generate a learning model for size estimation by a similar process. In this case, the teacher data generation device 1 acquires recognition area information identified by the size estimation device 300 in the process of recognizing the size of the recognition target, the recognition area information including at least an image including the recognition target and its area, and a likelihood that the recognition target is the recognition target assigned to each unit area in the image by the object detection device 100. The teacher data generation device 1 then segments the area of ​​the recognition target in the image indicated by the recognition area information based on the likelihood assigned to each unit area, and generates an image including the segmented area as teacher data for generating learning data used by the size estimation device 300 to recognize the size of the recognition target.

[0033] The program executed by the teacher data generation device 1 may be a program for implementing some of the above-mentioned functions. Furthermore, it may be a so-called difference file (difference program) that can implement the above-mentioned functions in combination with a program already recorded in the computer system.

[0034] Although the teacher data generation device 1 of this disclosure has been described above, this disclosure is not limited to the above-mentioned embodiment. Various modifications that can be understood by a person skilled in the art can be made to the configuration and details of this disclosure within the scope of this disclosure. Furthermore, each embodiment can be appropriately combined with other embodiments.

[0035] Note that some or all of the above-described embodiments can be described as, but are not limited to, the following supplementary notes.

[0036] (Appendix 1) An acquisition means for acquiring recognition area information identified by a recognition device through a recognition process of a recognition target, the recognition area information including at least an image including the recognition target and its area, and a likelihood that each unit area in the image is the recognition target; a segmentation means for segmenting an area of ​​the recognition target in the image indicated by the recognition area information based on the likelihood assigned to each unit area; A generating means for generating an image including the segmented region as teacher data for generating learning data to be used by the recognition device for recognizing the recognition target; A teacher data generating device comprising:

[0037] (Appendix 2) a complementation means for complementing the segmented region; Equipped with The generating means generates an image including the segmented and complemented region as teacher data for generating learning data to be used by the recognition device for recognizing the recognition target. A teacher data generation device as described in Appendix 1.

[0038] (Appendix 3) The recognition area information further includes label information of the recognition target recognized by the recognition process and probability information indicating a probability that the recognition target is the recognition target indicated by the label information, a confirmation request means for outputting the result of the complement based on a comparison between the probability information of the recognition target and a threshold value, and for obtaining corrected boundary information between the complemented region of the recognition target and other regions or approval information of the result; The generating means generates the training data indicating an image of the specified area based on the corrected boundary information or the approval information of the result. A teacher data generation device as described in Appendix 2.

[0039] (Appendix 4) The segmentation means segments an area of ​​the recognition target in an image indicated by the recognition area information identified by a recognition process of the recognition target included in a satellite image, The complementing means complements the segmented region based on the correct answer data of the recognition target and the result of the segmentation. A teacher data generation device as described in Appendix 3.

[0040] (Appendix 5) The segmentation means segments an area of ​​the recognition target in an image indicated by the recognition area information identified by a recognition process of the recognition target included in a satellite image, The complementing means complements the regions of the recognition target indicated by the result of the segmentation by a closing process or an opening process. A teacher data generation device as described in Appendix 3.

[0041] (Appendix 6) acquiring recognition area information identified by a recognition process of a recognition target by a recognition device, the recognition area information including at least an image including the recognition target and its area, and a likelihood that each unit area in the image is the recognition target; Segmenting the area of ​​the recognition target in the image indicated by the recognition area information based on the likelihood assigned to each unit area; An image including the segmented region is generated as training data for generating learning data to be used by the recognition device for recognizing the recognition target. How to generate training data.

[0042] (Appendix 7) Completing the segmented region; An image including the segmented and completed region is generated as training data for generating learning data to be used by the recognition device for recognizing the recognition target. A method for generating teacher data as described in Appendix 6.

[0043] (Appendix 8) The recognition area information further includes label information of the recognition target recognized by the recognition process and probability information indicating a probability that the recognition target is the recognition target indicated by the label information, outputting a result of the complement based on a comparison of probability information of the recognition target with a threshold value, and obtaining corrected boundary information between the complemented recognition target region and other regions or approval information of the result; Generate the training data indicating an image of the specified area based on the corrected boundary information or the approval information of the result. The teacher data generation method according to claim 7.

[0044] (Appendix 9) The recognition device segments an area of ​​the recognition target in an image indicated by the recognition area information identified by a recognition process of the recognition target included in a satellite image; The segmented region is complemented based on the correct answer data of the recognition target and the result of the segmentation. A method for generating teacher data as described in Appendix 8.

[0045] (Appendix 10) The recognition device segments an area of ​​the recognition target in an image indicated by the recognition area information identified by a recognition process of the recognition target included in a satellite image; The regions of the plurality of recognition objects indicated by the segmentation results are complemented by a closing process or an opening process. A method for generating teacher data as described in Appendix 8.

[0046] (Appendix 11) Computer, an acquisition means for acquiring recognition area information specified by a recognition device through a recognition process of a recognition target, the recognition area information including at least an image including the recognition target and its area, and a likelihood that each unit area in the image is the recognition target; a segmentation means for segmenting an area of ​​the recognition target in the image indicated by the recognition area information based on the likelihood assigned to each unit area; A generating means for generating an image including the segmented region as training data for generating learning data to be used by the recognition device for recognizing the recognition target; A program that functions as a

[0047] (Appendix 12) functioning as a complementing means for complementing the segmented region; The generating means generates an image including the segmented and complemented region as training data for generating learning data to be used by the recognition device for recognizing the recognition target. 12. The program according to claim 11.

[0048] (Appendix 13) The recognition area information further includes label information of the recognition target recognized by the recognition process and probability information indicating a probability that the recognition target is the recognition target indicated by the label information, a function as a confirmation request means for outputting the result of the complement based on a comparison of the probability information of the recognition target with a threshold value, and obtaining corrected boundary information between the complemented recognition target area and other areas or approval information of the result; The generating means generates the training data indicating an image of the specified area based on the corrected boundary information or the approval information of the result. 13. The program according to claim 12.

[0049] (Appendix 14) The segmentation means segments an area of ​​the recognition target in an image indicated by the recognition area information identified by a recognition process of the recognition target included in a satellite image, The complementing means complements the segmented region based on the correct answer data of the recognition target and the result of the segmentation. 14. The program according to claim 13.

[0050] (Appendix 15) The segmentation means segments an area of ​​the recognition target in an image indicated by the recognition area information identified by a recognition process of the recognition target included in a satellite image, The complementing means complements the regions of the recognition target indicated by the result of the segmentation by a closing process or an opening process. 14. The program according to claim 13. [Explanation of symbols]

[0051] 1. Teacher data generator 11...Acquisition part 12. Segmentation section 13 Complementary section 14...Generation section 15. Database 16. Confirmation Request Section 100 Object detection device 200...Object classification device 300 Size estimation unit

Claims

1. An acquisition means for acquiring recognition area information identified by a recognition device through a recognition process of a recognition target, the recognition area information including at least an image including the recognition target and its area, and a likelihood that each unit area in the image is the recognition target; a segmentation means for segmenting an area of ​​the recognition target in the image indicated by the recognition area information based on the likelihood assigned to each unit area; A generating means for generating an image including the segmented region as teacher data for generating learning data to be used by the recognition device for recognizing the recognition target; A teacher data generating device comprising:

2. a complementation means for complementing the segmented region; Equipped with The generating means generates an image including the segmented and complemented region as teacher data for generating learning data to be used by the recognition device for recognizing the recognition target. The teacher data generating device according to claim 1 .

3. The recognition area information further includes label information of the recognition target recognized by the recognition process and probability information indicating a probability that the recognition target is the recognition target indicated by the label information, a confirmation request means for outputting the result of the complement based on a comparison between the probability information of the recognition target and a threshold value, and for obtaining corrected boundary information between the complemented region of the recognition target and other regions or approval information of the result; The generating means generates the training data indicating an image of the specified area based on the corrected boundary information or the approval information of the result. The teacher data generating device according to claim 2 .

4. The segmentation means segments an area of ​​the recognition target in an image indicated by the recognition area information identified by a recognition process of the recognition target included in a satellite image, The complementing means complements the segmented region based on the correct answer data of the recognition target and the result of the segmentation. The teacher data generating device according to claim 3 .

5. The segmentation means segments an area of ​​the recognition target in an image indicated by the recognition area information identified by a recognition process of the recognition target included in a satellite image, The complementing means complements the regions of the recognition target indicated by the result of the segmentation by a closing process or an opening process. The teacher data generating device according to claim 3 .

6. acquiring recognition area information identified by a recognition process of a recognition target by a recognition device, the recognition area information including at least an image including the recognition target and its area, and a likelihood that each unit area in the image is the recognition target; Segmenting the area of ​​the recognition target in the image indicated by the recognition area information based on the likelihood assigned to each unit area; An image including the segmented region is generated as training data for generating learning data to be used by the recognition device for recognizing the recognition target. How to generate training data.

7. Computer, an acquisition means for acquiring recognition area information specified by a recognition device through a recognition process of a recognition target, the recognition area information including at least an image including the recognition target and its area, and a likelihood that each unit area in the image is the recognition target; a segmentation means for segmenting an area of ​​the recognition target in the image indicated by the recognition area information based on the likelihood assigned to each unit area; A generating means for generating an image including the segmented region as training data for generating learning data to be used by the recognition device for recognizing the recognition target; A program that functions as a

Images