Learning device, inference device, program, learning method, and inference method

Abstract

A processing information generation device (100) is characterized in that at least one learning model is trained using a plurality of processing feature images, which are a plurality of images respectively visualizing a plurality of features for processing of a target, and correct answer data indicating the result of correctly performing classification into a plurality of processing regions in the target, so that, when the plurality of processing feature images are input, the learning model classifies a region to be processed in one step into one processing region and performs classification into the plurality of processing regions indicated by the correct answer data.

Description

Learning device, inference device, program, learning method, and inference method

[0001] The present disclosure relates to a learning device, an inference device, a program, a learning method, and an inference method.

[0002] In NC (Numerical Control) machining, machining areas to be NC machined are identified from the external shape of the target object represented by CAD (Computer Aided Design) data, and a path is identified for each machining area.

[0003] When identifying a machining area from the external shape of an object to be NC machined, for example, semantic segmentation inference may be performed using an AI (Artificial Intelligence) model. Specifically, the machining area can be identified using the technology described in Non-Patent Document 1.

[0004] Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation”, arXiv:1505.04597v1 [cs. CV], June 18, 2015

[0005] As a method for identifying a machining area by semantic segmentation, there is a method using a depth map that indicates the depth of the machining part. However, when performing semantic segmentation using a depth map, it is sometimes not possible to extract sufficient information about machining features, such as the shape of the curved surface of a corner or the curved surface of a bottom, which are important for identifying the machining area, and it is therefore not possible to identify the machining area with sufficient accuracy.

[0006] Therefore, one or more aspects of the present disclosure aim to sufficiently extract the characteristics of processing and enable classification into multiple processing regions.

[0007] A learning device according to one aspect of the present disclosure is characterized in that, by inputting a plurality of processing feature images, which are a plurality of images that respectively visualize a plurality of features in the processing of an object, and correct answer data that indicates the result of correctly classifying the object into a plurality of processing areas, the learning device trains at least one learning model to classify an area that is processed in one process into a single processing area and classify the area into a plurality of processing areas indicated by the correct answer data.

[0008] An inference device according to one aspect of the present disclosure is characterized in that by inputting at least two learning processing feature images, which are at least two images that each visualize at least two features of the learning processing, which is the processing of the object to be learned, into a learning model that has been trained to classify an area to be processed in one process into one processing area, and to classify into multiple processing areas, by inputting at least two processing feature images, which are at least two images that each visualize at least two features of the processing of the object to be processed, into the learning model.

[0009] A program according to one aspect of the present disclosure is characterized in that, by inputting a plurality of processing feature images, which are a plurality of images that respectively visualize a plurality of features in the processing of an object, and correct answer data that indicates the result of correctly classifying the object into a plurality of processing areas into a computer, the program trains at least one learning model to classify an area that is processed in one process into a single processing area and classify it into the multiple processing areas indicated by the correct answer data.

[0010] A program according to one aspect of the present disclosure is characterized in that by inputting at least two learning processing characteristic images, which are at least two images that each visualize at least two characteristics of the learning processing, which is the processing to be learned, into a computer, the computer classifies an area to be processed in one process into one processing area, and by inputting at least two processing characteristic images, which are at least two images that each visualize at least two characteristics of the processing to be performed, into a learning model that has been trained to classify into multiple processing areas, the computer classifies into multiple processing areas.

[0011] A learning method according to one aspect of the present disclosure is characterized in that, by inputting a plurality of processing feature images, which are a plurality of images that respectively visualize a plurality of features in the processing of an object, and correct answer data that indicates the result of correctly classifying the object into a plurality of processing areas, at least one learning model is trained to classify an area that is processed in one process into a single processing area and classify it into the multiple processing areas indicated by the correct answer data.

[0012] An inference method according to one aspect of the present disclosure is characterized in that by inputting at least two learning processing feature images, which are at least two images that each visualize at least two features of the learning processing, which is the processing to be learned, into a learning model that has been trained to classify an area to be processed in one process into one processing area, and to classify into multiple processing areas, by inputting at least two processing feature images, which are at least two images that each visualize at least two features of the processing to be performed, into a learning model that has been trained to classify into multiple processing areas.

[0013] According to one or more aspects of the present disclosure, it is possible to sufficiently extract features of processing and classify the processing into multiple processing regions.

[0014] 1 is a block diagram schematically showing the configuration of a processing information generating device according to a first embodiment. FIG. 2 is a schematic diagram for explaining processing in a segmentation unit according to the first embodiment. FIG. 3 is a block diagram schematically showing the configuration of a PC. FIG. 4 is a flowchart showing processing in a learning phase of the processing information generating device according to the first embodiment. FIG. 5 is a flowchart showing processing in an inference phase of the processing information generating device according to the first embodiment. FIG. 6 is a block diagram schematically showing the configuration of a processing information generating device according to a second embodiment. FIG. 7 is a schematic diagram for explaining processing in a segmentation unit according to the second embodiment. FIG. 8 is a flowchart showing processing in a learning phase of the processing information generating device according to the second embodiment. FIG. 9 is a flowchart showing processing in an inference phase of the processing information generating device according to the second embodiment. FIG. 10 is a block diagram schematically showing the configuration of a processing information generating device according to a third embodiment. FIG. 11 is a flowchart showing processing in a learning phase of the processing information generating device according to the third embodiment. FIG. 12 is a flowchart showing processing in an inference phase of the processing information generating device according to the third embodiment. FIG. 13 is a schematic diagram showing an example of a learning model constructed by a combination setting unit. FIG. 14 is a schematic diagram showing another example of a learning model constructed by the combination setting unit. FIG. 15 is a block diagram schematically showing the configuration of a processing information generating device according to a fifth embodiment. FIG. 16 is a schematic diagram for explaining processing in a segmentation unit according to the fifth embodiment. 10 is a flowchart showing a process in a learning phase of the processing information generating device according to embodiment 5. FIG 11 is a flowchart showing a process in an inference phase of the processing information generating device according to embodiment 5. FIG 12 is a flowchart showing a process in an inference phase of the processing information generating device according to embodiment 5.

[0015] 1 is a block diagram showing a schematic configuration of a processing information generating device 100 according to embodiment 1. The processing information generating device 100 includes a CAD data storage unit 101, a processing feature image generating unit 102, a model storage unit 103, a segmentation unit 104, and a learning unit 105.

[0016] Here, the processing information generation device 100 functions as a learning device that uses multiple processing feature images, which are multiple images that visualize multiple features of processing of an object, and correct answer data that indicates the results of correctly classifying the object into multiple processing areas, as input, to classify an area to be processed in one process into one processing area and classify it into the multiple processing areas indicated by the correct answer data. Here, the object during learning is also referred to as a learning object, and the processing during learning is also referred to as learning processing. Furthermore, the processing feature images generated during learning are also referred to as learning processing feature images, and the classification performed during learning is also referred to as learning classification.

[0017] The processing information generation device 100 also functions as an inference device that classifies an area to be processed in one step into a single processing area by inputting at least two learning processing feature images, which are at least two images that visualize at least two features of the learning processing that is the processing of the learning target, into a learning model that has been trained to classify the area into a plurality of processing areas, by inputting at least two learning processing feature images, which are at least two images that visualize at least two features of the processing of the object to be processed, into the learning model that has been trained to classify the area into a plurality of processing areas. This will be explained below.

[0018] The CAD data storage unit 101 stores CAD data as design data indicating the design content of a processing object to be processed. The CAD data indicates, for example, the dimensions and shape of the processing object as the design content.

[0019] The processing feature image generation unit 102 generates a plurality of processing feature images from the CAD data. For example, the processing feature image generation unit 102 generates a processing feature image that shows the features of processing of the processing object. Here, the processing feature image generation unit 102 generates a plurality of processing feature images that each show different features from the processing direction of the processing object. The generated plurality of processing feature images are provided to the segmentation unit 104.

[0020] For example, the processing characteristic image generating unit 102 may generate, as one of the processing characteristic images, a depth map, which is a depth image indicating the depth at the time of processing, from the shape of the processing object indicated by CAD data.

[0021] In addition, the processed characteristic image generation unit 102 may generate, as one of the processed characteristic images, a curved surface image in which the curved surface is visualized so that a predetermined color, for example, red, becomes darker as the R of the curved surface becomes smaller than a predetermined degree, and another predetermined color, for example, blue, becomes darker as the R of the curved surface becomes larger than a predetermined degree.

[0022] In addition, the processing characteristic image generation unit 102 may generate a boundary image as one of the processing characteristic images, which visualizes the boundaries of the areas that constitute the processing object in the CAD data by drawing them in a predetermined color, for example, black.

[0023] In addition, the processing feature image generation unit 102 may generate a processing area image that is visualized by coloring the area that constitutes the processing object in the CAD data and the parts other than the processing object in different colors as one of the processing feature images.

[0024] In addition, the processing characteristic image generation unit 102 may generate a surface image that is visualized by coloring the uneven surfaces and flat surfaces of the area that constitutes the processing object in the CAD data with different colors, as one of the processing characteristic images.

[0025] In addition, the processing characteristic image generation unit 102 may generate a thickness image as one of the processing characteristic images in which the thickness between adjacent regions in the region that constitutes the processing object in the CAD data is visualized with a predetermined color brightness.

[0026] The processing characteristic image generating unit 102 is not limited to the above examples and generates a plurality of processing characteristic images showing characteristics when processing the processing object. Here, since the depth map contains the most basic information when performing processing, in the first embodiment, it is desirable that the processing characteristic image generating unit 102 always generate a depth map as one of the processing characteristic images.

[0027] The model storage unit 103 stores a learning model for performing semantic segmentation. The learning model in the first embodiment is a model for inferring a region classification image, which is a result of classifying a plurality of processed feature images into a plurality of processed regions. Here, the region classification image is information indicating the regions classified for each pixel.

[0028] The segmentation unit 104 inputs a plurality of processed feature images into a learning model to classify the images into a plurality of processed regions. For example, the segmentation unit 104 inputs the generated plurality of processed feature images into the learning model to acquire a region classification image from the learning model.

[0029] 2 is a schematic diagram for explaining the processing in the segmentation unit 104. As shown in FIG. 2, the segmentation unit 104 acquires a region classification image 122 by inputting a plurality of processed feature images 120-1, 120-2, 120-3, ..., 120-n (n is a positive integer indicating the number of processed feature images) into a learning model 121.

[0030] In the inference phase, the segmentation unit 104 classifies the multiple processed feature images generated by the processed feature image generation unit 102 into multiple processed regions by inputting the multiple processed feature images into a learning model trained by the learning unit 105. Then, the segmentation unit 104 generates a region image that visualizes the regions by coloring each region indicated in the region classification image output from the learning model with a predetermined color.

[0031] Returning to FIG. 1 , the learning unit 105 performs learning of the learning model in the learning phase. For example, the learning unit 105 learns the learning model using the classification results by the segmentation unit 104 and the correct answer data. Specifically, the learning unit 105 acquires correct answer data indicating the correct area classification image to be processed in advance, and learns the learning model so that the difference between the area classification image inferred by the segmentation unit 104 and the area classification image indicated by the correct answer data is small. The learning model learned in this manner is stored in the model storage unit 103.

[0032] The processing information generating device 100 described above can be realized by, for example, a computer such as the PC 10 shown in Fig. 3. The PC 10 includes a storage 11 such as a hard disk drive (HDD) and a solid state drive (SSD), a memory 12, a processor 13 such as a central processing unit (CPU), a communication interface (I / F) 14 such as a network interface card (NIC), an input interface 15 such as a keyboard and a mouse, and a display 16.

[0033] For example, the CAD data storage unit 101 and the model storage unit 103 can be realized by the storage 11 or the memory 12. The processed feature image generation unit 102, the segmentation unit 104, and the learning unit 105 can be realized by the processor 13 executing a program.

[0034] The program may be downloaded to the storage 11 from a recording medium (not shown) via a reader / writer (not shown) or from a network via the communication I / F 14, and then loaded onto the memory 12 and executed by the processor 13. Alternatively, the program may be directly loaded onto the memory 12 from a recording medium via the reader / writer or from a network via the communication I / F 14, and then executed by the processor 13. In other words, the program may be provided by a computer program product such as a recording medium.

[0035] 4 is a flowchart showing processing in the learning phase of the processing information generating device 100 according to the first embodiment. First, the processing feature image generating unit 102 generates a plurality of processing feature images showing features when processing the processing object from the CAD data stored in the CAD data storage unit 101 (S10). The plurality of processing feature images generated here each show a different feature of the processing object. The generated plurality of processing feature images are provided to the segmentation unit 104.

[0036] Next, the segmentation unit 104 inputs the plurality of processed feature images generated in step S10 into the learning model stored in the model storage unit 103, thereby acquiring a region classification image from the learning model (S11). The acquired region classification image is provided to the learning unit 105.

[0037] The learning unit 105 calculates the difference between the area classification image acquired in step S11 and the area classification image indicated by the previously acquired correct answer data (S12). The learning unit 105 then determines whether convergence has occurred by determining whether the difference in step S12 is less than a predetermined threshold (S13). If convergence has occurred (Yes in S13), the process ends; if convergence has not occurred (No in S13), the process proceeds to step S14.

[0038] In step S14, the learning unit 105 learns the learning model so as to reduce the difference in step S12. For example, the learning unit 105 updates the weights of the learning model. Then, the process returns to step S11.

[0039] 4, convergence is determined based on the difference being less than a threshold value, but the first embodiment is not limited to this example. For example, the learning unit 105 may proceed to step S14 after the process of step S12, and determine convergence based on whether the number of times learning in step S14 has been performed exceeds a predetermined number of times after the process of step S14. In such a case, if convergence is determined, the process ends, and if convergence is not determined, the process returns to step S11.

[0040] 5 is a flowchart showing processing in the inference phase of the processing information generating device 100 according to the first embodiment. First, the processing feature image generating unit 102 generates a plurality of processing feature images showing features when processing the processing object from the CAD data stored in the CAD data storage unit 101 (S20). Each of the plurality of processing feature images generated here shows a different feature. The generated plurality of processing feature images are provided to the segmentation unit 104.

[0041] Next, the segmentation unit 104 inputs the multiple processed feature images generated in step S10 into the learning model stored in the model storage unit 103, thereby obtaining a region classification image from the learning model (S21).

[0042] Then, the segmentation unit 104 generates a region image, which is an image that visualizes the region, by coloring each region indicated in the region classification image output from the learning model in step S21 with a predetermined color (S22).

[0043] As described above, according to the first embodiment, important information regarding processing can be input into a learning model that performs semantic segmentation, thereby improving the accuracy of inference.

[0044] 6 is a block diagram showing a schematic configuration of a processing information generating device 200 according to embodiment 2. The processing information generating device 200 includes a CAD data storage unit 101, a processing feature image generating unit 102, a model storage unit 203, a segmentation unit 204, a learning unit 205, and a combination search unit 206.

[0045] The CAD data storage unit 101 and the processing characteristic image generation unit 102 of the processing information generation device 200 according to the second embodiment are similar to the CAD data storage unit 101 and the processing characteristic image generation unit 102 of the processing information generation device 100 according to the first embodiment.

[0046] The model storage unit 203 stores a learning model for performing semantic segmentation. In the second embodiment, one learning model is constructed corresponding to each of the plurality of processed feature images generated by the processed feature image generation unit 102, and the plurality of learning models corresponding to the plurality of processed feature images are stored in the model storage unit 203.

[0047] Each of the multiple learning models stored in the model memory unit 203 is a model for inferring a region classification image that is the result of receiving a corresponding processed feature image as input and classifying that processed feature image into multiple processed regions.

[0048] The segmentation unit 204 performs multiple classifications into multiple processing regions by inputting each of the multiple processing feature images generated by the processing feature image generation unit 102 into a corresponding one of the multiple learning models. For example, the segmentation unit 204 inputs each of the multiple generated processing feature images into the corresponding learning model, thereby acquiring multiple region classification images from the multiple learning models. Then, in the learning phase, the segmentation unit 204 provides the acquired multiple region classification images to the learning unit 205. Furthermore, the segmentation unit 204 provides the acquired multiple region classification images after learning in the learning phase by the learning unit 205 is completed to the combination search unit 206.

[0049] In the inference phase, the segmentation unit 204 performs two or more classifications into a plurality of processing regions by inputting two or more processing feature images generated by the processing feature image generation unit 102 into two or more corresponding learning models, and performs one classification into a plurality of processing regions by aggregating the results of the two or more classifications according to a predetermined rule. Here, the two or more processing feature images are two or more images that are determined to be effective when performing one classification into a plurality of processing regions indicated by the correct answer data, by inputting multiple learning processing feature images into multiple corresponding learning models, and aggregating the results of multiple classifications into a plurality of processing regions according to a predetermined rule.

[0050] For example, in the inference phase, the segmentation unit 204 acquires two or more region classification images by inputting two or more processed feature images corresponding to the combination identified by the combination search unit 206 into a corresponding learning model. Then, the segmentation unit 204 generates one region classification image using the acquired two or more region classification images. Here, the segmentation unit 204 may generate one region classification image by, for example, determining for each pixel by majority vote whether or not the pixel belongs to one of a plurality of processed regions.

[0051] 7 is a schematic diagram for explaining the processing in the segmentation unit 204. As shown in FIG. 7, the segmentation unit 204 inputs each of a plurality of processed feature images 120-1, 120-2, 120-3, ..., 120-n (n is a positive integer indicating the number of processed feature images) into corresponding learning models 221-1, 221-2, 221-3, ..., 221-n, thereby acquiring a plurality of region classification images 222-1, 222-2, 222-3, ..., 222-n.

[0052] Then, in the inference phase, the segmentation unit 204 generates one region classification image according to the combination GR found by the combination search unit 206. In Fig. 7, for example, it is assumed that, out of multiple region classification images 222-1, 222-2, 222-3, ..., 222-n, the combination of region classification images 222-1, 222-2, and 222-3 is judged to be "valid."

[0053] In such a case, the segmentation unit 204 identifies the regions shown in the region classification images 222-1, 222-2, and 222-3 for each pixel as one region by majority vote and generates one region classification image to indicate the identified region.The segmentation unit 204 then generates a region image that visualizes the region by coloring each region shown in the identified region classification image with a predetermined color.

[0054] 1 , in the learning phase, the learning unit 205 executes learning of a plurality of learning models corresponding to each of a plurality of processing feature images. For example, the learning unit 105 acquires in advance correct answer data indicating the correct area classification image of the processing target, and learns the corresponding learning model so that the difference between each of the plurality of area classification images inferred by the segmentation unit 204 and the area classification image indicated by the correct answer data is small. The plurality of learning models thus learned are stored in the model storage unit 203.

[0055] The combination search unit 206 is a search unit that searches for two or more processing feature images that are effective when performing a single classification from multiple processing feature images into multiple processing areas indicated by the correct answer data by aggregating the results of multiple classifications by the segmentation unit 204 according to predetermined rules.

[0056] Here, the combination search unit 206 searches for an effective combination that can be used to generate one region classification image from among the multiple region classification images 222-1, 222-2, 222-3, ..., 222-n during the learning phase. For example, for each combination of the multiple region classification images 222-1, 222-2, 222-3, ..., 222-n, the combination search unit 206 identifies a region by majority vote for each pixel, and compares the identified region with the correct region classification image indicated by the correct answer data, thereby searching for a combination that maximizes accuracy. In other words, here, the counting is performed by majority vote as a predetermined rule.

[0057] Here, the combination search unit 206 identifies two or more effective combinations from the plurality of region classification images 222-1, 222-2, 222-3, ..., 222-n for each processing region indicated by the correct answer data. Then, the combination search unit 206 notifies the segmentation unit 204 of all region classification images included in the two or more combinations identified for each processing region as effective combinations. Here, the combination search unit 206 searches for effective region classification images, but because the region classification images are generated for each processing feature image, searching for effective region classification images is synonymous with searching for effective processing feature images.

[0058] The processing information generation device 200 described above can also be realized by a computer such as the PC 10 shown in Fig. 3. For example, the combination search unit 206 can also be realized by the processor 13 executing a program.

[0059] 8 is a flowchart showing the processing in the learning phase of the processing information generating device 200 according to the second embodiment. First, the processing feature image generating unit 102 generates a plurality of processing feature images showing features when processing the processing object from the CAD data stored in the CAD data storage unit 101 (S30). Each of the plurality of processing feature images generated here shows a different feature. The generated plurality of processing feature images are provided to the segmentation unit 204.

[0060] Next, the segmentation unit 204 inputs each of the plurality of processed feature images generated in step S30 into the corresponding learning model stored in the model storage unit 203, thereby acquiring a plurality of region classification images from the plurality of learning models (S31). The acquired plurality of region classification images are provided to the learning unit 205.

[0061] The learning unit 205 calculates the difference between each of the multiple region classification images acquired in step S31 and the region classification image indicated by the previously acquired correct answer data (S32).The learning unit 205 then determines whether convergence has occurred for each learning model by determining whether the difference in step S32 is less than a predetermined threshold for each learning model (S33).If convergence has occurred for all learning models (Yes in S33), the process proceeds to step S35.If there is a learning model that has not converged (No in S33), the process proceeds to step S34.

[0062] In step S34, the learning unit 205 learns the learning model determined not to have converged so that the difference in step S32 is reduced. For example, the learning unit 205 updates the weight of the learning model. Then, the process returns to step S31, and the process is repeated for the learning model determined not to have converged.

[0063] In step S35, the combination search unit 206 searches for an effective combination that can be used to generate one region classification image from among the multiple region classification images. The effective combination found here is notified to the segmentation unit 204.

[0064] In the flowchart shown in FIG. 8 , convergence is determined based on the difference being less than a threshold value, but the second embodiment is not limited to this example. For example, the learning unit 205 may proceed to step S34 after the process of step S32, and after the process of step S34, determine whether convergence has occurred based on whether the number of times learning in step S34 has been performed exceeds a predetermined number. In such a case, the process of step S35 may be performed after it is determined that convergence has occurred. If it is not determined that convergence has occurred, the process returns to step S31, and the process is repeated for the learning model determined not to have converged.

[0065] 9 is a flowchart showing the processing in the inference phase of the processing information generating device 200 according to the second embodiment. First, the processing feature image generating unit 102 generates a plurality of processing feature images showing features when processing the processing object from the CAD data stored in the CAD data storage unit 101 (S40). Each of the plurality of processing feature images generated here shows a different feature. The generated plurality of processing feature images are provided to the segmentation unit 204.

[0066] Next, the segmentation unit 204 identifies, from the plurality of processed characteristic images, two or more processed characteristic images that correspond to the combination notified by the combination search unit 206 (S41).

[0067] Then, the segmentation unit 204 inputs the two or more processed feature images identified in step S41 into two or more corresponding learning models stored in the model storage unit 203, thereby acquiring two or more region classification images from the two or more learning models (S42).The segmentation unit 204 generates one region classification image by majority vote using the two or more acquired region classification images.

[0068] Then, the segmentation unit 204 generates a region image, which is an image that visualizes the region, by coloring each region indicated in one region classification image generated in step S42 with a predetermined color (S43).

[0069] 9, two or more processed feature images corresponding to valid combinations are identified in step S41 from the plurality of processed feature images generated in step S40, but the second embodiment is not limited to this example. For example, the processed feature image generating unit 102 may generate only two or more processed feature images corresponding to valid combinations.

[0070] As described above, in the second embodiment, important information regarding processing can be input into the learning model that performs semantic segmentation, thereby improving the accuracy of inference.

[0071] In the second embodiment, the combination search unit 206 determines whether the region classification image is valid or invalid, but the second embodiment is not limited to such an example. For example, the combination search unit 206 may search for weights that indicate the validity of the region classification image. In this case, since the number of weights to be searched becomes enormous, the combination search unit 206 may efficiently search for combinations and weights using an optimization method such as Bayesian optimization. In other words, the combination search unit 206 may function as a search unit that aggregates the results of multiple classifications by the segmentation unit 204 according to a predetermined rule, and then searches for weights to be assigned to each of the results of the multiple classifications when performing one classification into multiple processing regions.

[0072] In this case, in the inference phase, the segmentation unit 204 performs multiple classifications into multiple processing areas by inputting the multiple processing feature images generated by the processing feature image generation unit 102 into multiple corresponding learning models, and then performs one classification into multiple processing areas by adding a predetermined weight to each of the results of the multiple classifications and then aggregating them according to a predetermined rule.

[0073] 10 is a block diagram showing a schematic configuration of a processing information generating device 300 according to embodiment 3. The processing information generating device 300 includes a CAD data storage unit 101, a processing feature image generating unit 102, a model storage unit 303, a segmentation unit 304, a learning unit 305, a combination searching unit 206, and a combination setting unit 307.

[0074] The CAD data storage unit 101 and the processing feature image generation unit 102 of the processing information generation device 300 according to the third embodiment are the same as the CAD data storage unit 101 and the processing feature image generation unit 102 of the processing information generation device 100 according to the first embodiment. Furthermore, the combination search unit 206 of the processing information generation device 300 according to the third embodiment is the same as the combination search unit 206 of the processing information generation device 200 according to the second embodiment. Therefore, also in the third embodiment, the combination search unit 206 functions as a search unit that searches for two or more processing feature images that are effective for classifying a plurality of processing feature images into a plurality of processing regions indicated by the correct answer data. However, in the third embodiment, the combination search unit 206 also notifies the combination setting unit 307 of the effective combinations that have been found.

[0075] The model storage unit 303 stores a learning model for performing semantic segmentation. In the third embodiment, one learning model is configured corresponding to each of the plurality of processed feature images generated by the processed feature image generation unit 102, and the model storage unit 303 stores a plurality of learning models corresponding to each of the plurality of processed feature images.

[0076] Each of the multiple learning models stored in the model memory unit 303 is a model for inferring an area classification image that is the result of receiving a corresponding processed feature image as input and classifying that processed feature image into multiple processed areas.

[0077] In addition, in the third embodiment, the model storage unit 303 also stores a learning model that infers an area classification image by inputting two or more processed feature images corresponding to effective combinations that are searched for by the combination search unit 206, set by the combination setting unit 307, and learned by the learning unit 305, as will be described later.

[0078] In the learning phase, the segmentation unit 304 inputs each of the multiple processed feature images generated by the processed feature image generation unit 102 into a corresponding learning model, thereby acquiring multiple region classification images from the multiple learning models. Then, the segmentation unit 304 provides the acquired multiple region classification images to the learning unit 305. Furthermore, the segmentation unit 304 provides the multiple region classification images acquired after learning by the learning unit 305 is completed in the learning phase to the combination search unit 206.

[0079] Then, in the learning phase, the segmentation unit 304 inputs two or more processed feature images corresponding to two or more area classification images that have been determined to be an effective combination by the combination search unit 206 into a learning model constructed by the combination setting unit 307, and provides the output area classification image to the learning unit 305.

[0080] In the inference phase, the segmentation unit 304 performs one classification into a plurality of processing regions and obtains one region classification image by inputting two or more processed feature images corresponding to two or more region classification images determined to be an effective combination by the combination search unit 206 into a newly trained learning model.The segmentation unit 304 then generates a region image that visualizes the regions by coloring each region indicated in that one region classification image with a predetermined color.

[0081] In the learning phase, the learning unit 305 first executes learning of a plurality of learning models corresponding to each of the plurality of processing feature images. For example, the learning unit 305 acquires in advance correct answer data indicating the correct region classification image of the processing target, and learns the corresponding learning model so that the difference between each of the plurality of region classification images inferred by the segmentation unit 304 and the region classification image indicated by the correct answer data is small. The plurality of learning models learned in this manner are stored in the model storage unit 303.

[0082] Furthermore, when a new learning model is constructed by the combination setting unit 307, the learning unit 305 further trains the new learning model by inputting the two or more processed feature images generated by the processed feature image generation unit 102 and the correct answer data, so as to classify an area to be processed in one process into one processed area and classify the area into multiple processed areas indicated by the correct answer data. The two or more processed feature images here are two or more processed feature images corresponding to two or more area classification images determined to be valid by the combination search unit 206.

[0083] For example, when a new learning model is constructed by the combination setting unit 307, the learning unit 305 trains the learning model so as to reduce the difference between one region classification image inferred by the segmentation unit 304 and the region classification image indicated by the correct answer data. The one learning model trained in this way is stored in the model storage unit 303.

[0084] The combination setting unit 307 functions as a construction unit that inputs two or more processed feature images corresponding to two or more area classification images that have been determined to be an effective combination by the combination search unit 206, and constructs a new learning model for inferring the area classification image.

[0085] The processing information generating device 300 described above can also be realized by a computer such as the PC 10 shown in Fig. 3. For example, the combination setting unit 307 can also be realized by the processor 13 executing a program.

[0086] 11 is a flowchart showing processing in the learning phase of the processing information generating device 300 according to the third embodiment. First, the processing feature image generating unit 102 generates a plurality of processing feature images showing features when processing the processing object from the CAD data stored in the CAD data storage unit 101 (S50). Each of the generated processing feature images shows a different feature. The generated processing feature images are provided to the segmentation unit 304.

[0087] Next, the segmentation unit 304 inputs each of the plurality of processed feature images generated in step S50 into the corresponding learning model stored in the model storage unit 303, thereby acquiring a plurality of region classification images from the plurality of learning models (S51). The acquired plurality of region classification images are provided to the learning unit 305.

[0088] The learning unit 305 calculates the difference between each of the multiple region classification images acquired in step S51 and the region classification image indicated by the previously acquired correct answer data (S52).The learning unit 305 then determines whether convergence has occurred for each learning model by determining for each learning model whether the difference in step S52 is less than a predetermined threshold (S53).If convergence has occurred for all learning models (Yes in S53), the process proceeds to step S55.If there is a learning model that has not converged (No in S53), the process proceeds to step S54.

[0089] In step S54, the learning unit 305 learns the learning model determined not to have converged so that the difference in step S52 is reduced. For example, the learning unit 305 updates the weight of the learning model. Then, the process returns to step S51, and the process is repeated for the learning model determined not to have converged.

[0090] In step S55, the combination search unit 206 searches for an effective combination that can be used to generate one region classification image from among the multiple region classification images. The effective combination found here is notified to the segmentation unit 304 and the combination setting unit 307.

[0091] Next, the combination setting unit 307 inputs two or more processed feature images corresponding to two or more area classification images determined to be an effective combination by the combination search unit 206, and constructs a learning model for inferring the area classification image (S56).

[0092] Next, the segmentation unit 304 inputs two or more processed feature images corresponding to the two or more region classification images determined to be a valid combination in step S55 into the learning model constructed in step S56, thereby acquiring region classification images from the learning model (S57). The acquired region classification images are provided to the learning unit 305.

[0093] The learning unit 305 calculates the difference between the area classification image acquired in step S57 and the area classification image indicated by the previously acquired correct answer data (S58). The learning unit 305 then determines whether convergence has occurred by determining whether the difference acquired in step S58 is less than a predetermined threshold (S59). If convergence has occurred (Yes in S59), the process ends; if convergence has not occurred (No in S59), the process proceeds to step S60.

[0094] In step S60, the learning unit 305 learns the learning model so as to reduce the difference in step S58. For example, the learning unit 305 updates the weights of the learning model. Then, the process returns to step S57.

[0095] Note that, in step S53 of the flowchart shown in FIG. 11 , convergence is determined based on the difference being less than a threshold value, but the third embodiment is not limited to this example. For example, the learning unit 305 may proceed to step S54 after the process of step S52, and after the process of step S54, determine whether convergence has occurred based on whether the number of times learning in step S54 has been performed exceeds a predetermined number. In such a case, the process of step S55 may be performed after it is determined that convergence has occurred. Note that, if convergence has not been determined, the process returns to step S51, and the process is repeated for the learning model determined not to have converged.

[0096] 11 , in step S59 of the flowchart, convergence is determined based on the difference being less than a threshold value, but the third embodiment is not limited to this example. For example, the learning unit 305 may proceed to step S60 after the process of step S58, and after the process of step S60, determine whether convergence has occurred based on whether the number of times learning in step S60 has been performed exceeds a predetermined number. In such a case, if convergence is determined, the process ends, and if convergence is not determined, the process returns to step S57.

[0097] 12 is a flowchart showing processing in the inference phase of the processing information generating device 300 according to the third embodiment. First, the processing feature image generating unit 102 generates a plurality of processing feature images showing features when processing the processing object from the CAD data stored in the CAD data storage unit 101 (S70). The plurality of processing feature images generated here each show a different feature. The generated plurality of processing feature images are provided to the segmentation unit 304.

[0098] Next, the segmentation unit 304 identifies, from the plurality of processed characteristic images, two or more processed characteristic images that correspond to the combination notified by the combination search unit 206 (S71).

[0099] Then, the segmentation unit 304 obtains one area classification image by inputting the two or more processed feature images identified in step S71 into a learning model stored in the model memory unit 303, constructed by the combination setting unit 307, and learned by the learning unit 305 (S72).

[0100] Then, the segmentation unit 304 generates a region image, which is an image that visualizes the region, by coloring each region indicated in one region classification image obtained in step S72 with a predetermined color (S73).

[0101] 12, two or more processed feature images corresponding to valid combinations are identified in step S71 from the plurality of processed feature images generated in step S70, but the third embodiment is not limited to this example. For example, the processed feature image generating unit 102 may generate only two or more processed feature images corresponding to valid combinations.

[0102] As described above, in the third embodiment, important information regarding processing can be input into the learning model that performs semantic segmentation, thereby improving the accuracy of inference.

[0103] 10 , a processing information generating device 400 according to a fourth embodiment includes a CAD data storage unit 101, a processing feature image generating unit 102, a model storage unit 403, a segmentation unit 404, a learning unit 405, a combination searching unit 206, and a combination setting unit 407.

[0104] The CAD data storage unit 101 and the processing feature image generation unit 102 of the processing information generation device 400 according to embodiment 4 are similar to the CAD data storage unit 101 and the processing feature image generation unit 102 of the processing information generation device 100 according to embodiment 1. Furthermore, the combination search unit 206 of the processing information generation device 400 according to embodiment 4 is similar to the combination search unit 206 of the processing information generation device 200 according to embodiment 2. However, in embodiment 4, the combination search unit 206 also notifies the combination setting unit 407 of the found effective combination.

[0105] The model storage unit 403 stores a learning model for performing semantic segmentation. In the fourth embodiment, one learning model is constructed corresponding to each of the plurality of processed feature images generated by the processed feature image generation unit 102, and the plurality of learning models corresponding to each of the plurality of processed feature images are stored in the model storage unit 403.

[0106] Each of the multiple learning models stored in the model memory unit 403 is a model for inferring a region classification image that is the result of receiving a corresponding processed feature image as input and classifying that processed feature image into multiple processed regions.

[0107] In addition, in the fourth embodiment, as will be described later, two or more processed feature images corresponding to effective combinations searched for by the combination search unit 206, set by the combination setting unit 407, and learned by the learning unit 405 are encoded by the respective encoders in two or more corresponding learning models, and the results are combined and decoded by one decoder, thereby also storing a learning model that infers a region classification image.

[0108] In the learning phase, the segmentation unit 404 inputs each of the multiple processed feature images generated by the processed feature image generation unit 102 into a corresponding learning model, thereby acquiring multiple region classification images from the multiple learning models. Then, the segmentation unit 404 provides the acquired multiple region classification images to the learning unit 405. Furthermore, the segmentation unit 404 provides the multiple region classification images acquired after learning by the learning unit 405 is completed in the learning phase to the combination search unit 206.

[0109] Then, in the learning phase, the segmentation unit 404 inputs two or more processed feature images corresponding to two or more area classification images that have been determined to be an effective combination by the combination search unit 206 into a learning model constructed by the combination setting unit 407, and provides the output area classification image to the learning unit 405.

[0110] In the inference phase, the segmentation unit 404 obtains one region classification image by inputting two or more processed feature images corresponding to the combination identified by the combination search unit 206 into a newly trained learning model. Then, the segmentation unit 404 generates a region image that visualizes the regions by coloring each region indicated in the one region classification image with a predetermined color.

[0111] In the learning phase, the learning unit 405 first executes learning of a plurality of learning models corresponding to each of the plurality of processing feature images. For example, the learning unit 405 acquires in advance correct answer data indicating the correct region classification image of the processing target, and learns the corresponding learning model so that the difference between each of the plurality of region classification images inferred by the segmentation unit 404 and the region classification image indicated by the correct answer data is small. The plurality of learning models learned in this manner are stored in the model storage unit 403.

[0112] Furthermore, when a new learning model is constructed by the combination setting unit 407, the learning unit 405 learns the learning model so as to reduce the difference between one region classification image inferred by the segmentation unit 404 and the region classification image indicated by the correct answer data. The one learning model learned in this way is stored in the model storage unit 403.

[0113] The combination setting unit 407 functions as a construction unit that constructs a new learning model for inferring region classification images by inputting two or more processed feature images corresponding to two or more region classification images determined to be an effective combination by the combination search unit 206. For example, the combination setting unit 407 extracts two or more encoders from each of two or more learning models that input two or more processed feature images corresponding to two or more region classification images determined to be effective by the combination search unit 206, among the multiple learning models. Then, the combination setting unit 407 constructs a new learning model using the two or more encoders, a concatenation layer that concatenates and outputs the outputs of the two or more encoders, and a decoder that decodes the output of the concatenation layer.

[0114] Fig. 13 is a schematic diagram showing an example of a learning model constructed by the combination setting unit 407. The multiple processed feature images 420-1, 420-2, ..., 420-m shown in Fig. 13 are processed feature images corresponding to the multiple area classification images determined to be effective by the combination searching unit 206. Here, m is a positive integer indicating the number of area classification images determined to be effective by the combination searching unit 206.

[0115] The combination setting unit 407 extracts encoders 4211-1, 4211-2, ..., 4211-m from a plurality of learning models for inputting each of a plurality of processed feature images 420-1, 420-2, ..., 420-m from the model storage unit 403. Then, the combination setting unit 407 provides a concatenate layer 4212 as a concatenated layer that connects outputs from the encoders 4211-1, 4211-2, ..., 4211-m, and decodes the output from the concatenate layer 4212 with a decoder 4213. As described above, the combination setting unit 407 constructs a learning model 421 using the encoders 4211-1, 4211-2, ..., 4211-m, the concatenate layer 4212, and the decoder 4213.

[0116] Here, the encoders 4211-1, 4211-2, ..., 4211-m are convolution layers that perform convolution, and the decoder 4213 can be a transposed convolution layer. As a result, the region classification image 422 is output from the decoder 4213.

[0117] As described above, in the fourth embodiment, important information regarding processing can be input into the learning model that performs semantic segmentation, thereby improving the accuracy of inference.

[0118] In the fourth embodiment described above, for example, as shown in FIG. 13, the outputs from the encoders 4211-1, 4211-2, ..., 4211-m are directly concatenated in the concatenate layer 4212, but the fourth embodiment is not limited to such an example.

[0119] For example, as shown in FIG. 14, the combination setting unit 407 assigns weights w １ , w ２ ,...,w ｍBy adding a weighting layer 4214 that adds

[0120] Weight w １ , w ２ ,...,w ｍ The weight w may be set by the user via an input unit such as a keyboard (not shown). １ , w ２ ,...,w ｍ In this case, since the number of combinations and weights to be searched becomes enormous, the combination search unit 206 may use an optimization method such as Bayesian optimization to efficiently search for combinations and weights. In addition, the weights w calculated using an activation function such as a Global Average Pooling layer or a sigmoid function for the final layer of the encoder may be used. １ , w ２ ,...,w ｍ The weights of the combinations may be learned using a network that multiplies the weights before entering the concatenate layer.

[0121] In other words, the combination setting unit 407 may function as a construction unit that extracts two or more encoders from each of two or more learning models that input two or more processed feature images among multiple learning models, and constructs a new learning model using the two or more encoders, a weighting layer that adds weights to the outputs of the two or more encoders and outputs them, a concatenation layer that concatenates and outputs the outputs of the weighting layer, and a decoder that decodes the output of the concatenation layer.

[0122] Furthermore, the learning unit 405 sets the weight w １ , w ２ ,...,w ｍmay be learned. In this case, a known attention mechanism may be used. After the combination search unit 206 searches for a weight as a first weight, the learning unit 405 may also learn the weight as a second weight. In this case, the learning unit 405 may perform learning so that the difference between the first weight and the second weight becomes small. Alternatively, the learning unit 405 may perform learning so that the weight set by the user becomes the first weight and the second weight calculated by the learning unit 405 approaches the first weight.

[0123] 15 is a block diagram showing a schematic configuration of a processing information generating device 500 according to embodiment 5. The processing information generating device 500 includes a CAD data storage unit 101, a processing feature image generating unit 102, a model storage unit 503, a segmentation unit 504, a learning unit 505, a combination search unit 506, and a mask image conversion unit 508.

[0124] The CAD data storage unit 101 and the processing characteristic image generation unit 102 of the processing information generation device 500 according to embodiment 5 are similar to the CAD data storage unit 101 and the processing characteristic image generation unit 102 of the processing information generation device 100 according to embodiment 1.

[0125] The model storage unit 503 stores a learning model for performing semantic segmentation. In the fifth embodiment, one learning model is constructed corresponding to each of the plurality of processed feature images generated by the processed feature image generation unit 102, and the model storage unit 503 stores a plurality of learning models corresponding to each of the plurality of processed feature images.

[0126] Each of the multiple learning models stored in the model memory unit 503 is a model for inferring a region classification image that is the result of receiving a corresponding processed feature image as input and classifying that processed feature image into multiple processed regions.

[0127] In addition, in the fifth embodiment, as will be described later, two or more processed feature images corresponding to valid combinations searched for by the combination search unit 506 and learned by the learning unit 505 are input, and a learning model is also stored that performs mask processing using a mask image created from one of the searched processed feature images to infer a region classification image.

[0128] The segmentation unit 504 selects one processing feature image from the multiple processing feature images generated by the processing feature image generation unit 102 as a mask processing feature image, and selects two or more processing feature images from the multiple processing feature images excluding the mask processing feature image as two or more input processing feature images. The segmentation unit 504 then inputs the two or more input processing feature images into a learning model that decodes the results of mask processing performed on the output from the encoder using a mask image generated from the mask processing feature image, thereby classifying the area to be processed in one process into one processing area. The process of classifying the multiple processing areas is performed by changing the combination of the one processing feature image selected as the mask processing feature image and the two or more processing feature images selected as the two or more input processing feature images. The mask processing feature image in the learning phase is also referred to as a learning mask processing feature image, the input processing feature image in the learning phase is also referred to as a learning input processing feature image, and the mask image in the learning phase is also referred to as a learning mask image.

[0129] For example, in the learning phase, the segmentation unit 504 selects one processed feature image as a mask processed feature image from the multiple processed feature images generated by the processed feature image generation unit 102. The segmentation unit 504 also provides the mask processed feature image to the mask image conversion unit 508 and receives a mask image converted from the mask processed feature image.

[0130] The segmentation unit 504 then inputs two or more processed feature images selected from the plurality of processed feature images excluding the one processed feature image into the learning model, performs mask processing on the features encoded in the learning model using the mask image, and decodes the output of the mask processing to obtain a region classification image. The segmentation unit 504 provides the obtained region classification image, information identifying the mask processed feature image and the two or more input processed feature images to the learning unit 505.

[0131] Furthermore, in the inference phase, the segmentation unit 504 inputs two or more processed feature images corresponding to the two or more input processed feature images identified by the combination search unit 506 into the learning model, and acquires one region classification image by performing mask processing using a mask image converted from the processed feature image corresponding to the masking processed feature image identified by the combination search unit 506. Then, the segmentation unit 504 generates a region image that visualizes the regions by coloring each region indicated in the one region classification image with a predetermined color.

[0132] 16 is a schematic diagram for explaining the processing in the segmentation unit 504. The segmentation unit 504 acquires a region classification image 522 by inputting two or more processed feature images 520-1, 520-2, ..., 520-p selected from the remaining processed feature images generated by the processed feature image generation unit 102 after excluding the mask processed feature images, as two or more input processed feature images to a learning model 521.

[0133] The learning model 521 includes an encoder 5211 and a decoder 5212, and performs mask processing on the feature map output from the encoder 5211 using a mask image, and the output after the mask processing is input to the decoder 5212. The mask image here is generated by the mask image conversion unit 508, as will be described later. Note that the mask processing here is performed by multiplying each pixel value of the feature map by the corresponding pixel value of the mask image.

[0134] The learning unit 505 performs learning of the learning model in the learning phase. For example, the learning unit 505 acquires correct answer data indicating the correct area classification image to be processed in advance, and learns a corresponding learning model so that the difference between the area classification image inferred by the segmentation unit 504 and the area classification image indicated by the correct answer data is small. Then, upon completion of learning, the learning unit 505 provides the combination search unit 506 with information identifying the mask processing feature image and two or more input processing feature images used in learning, as well as the difference at the time learning was completed. Note that the multiple learning models learned in this manner are stored in the model storage unit 503.

[0135] The mask image conversion unit 508 generates a mask image from the processed feature image for masking provided by the segmentation unit 504. For example, the mask image conversion unit 508 generates a mask image by converting each pixel value of the processed feature image for masking into a pixel value between 0 and 1.0 according to a predetermined rule.

[0136] Specifically, when the processing feature image for masking is a surface image visualized by coloring the concave and convex surfaces and flat surfaces of the area constituting the processing target with different colors, since processing of concave surfaces is important in NC processing, by setting concave surfaces to "1.0", convex surfaces to a value between "1.0" and "0", for example, "0.5", and flat surfaces to "0", it becomes possible to perform learning and inference that is focused on effective features. From the above, it is assumed that the mask image conversion unit 508 has predetermined rules for converting the processing feature image for masking to a mask image such that the higher the degree of effectiveness of the features shown in the processing feature image for processing, the higher the value.

[0137] The mask image conversion unit 508 may reduce the resolution of the mask image so that it is equal to the resolution of the feature map output from the encoder of the learning model.

[0138] The combination search unit 506 functions as a search unit that searches for the optimal combination for classification from among combinations of a mask processing feature image and two or more input area classification images based on the learning results of the learning model.

[0139] For example, in the learning phase, the combination search unit 506 identifies a combination of a mask feature image and two or more input feature images with the smallest difference based on information identifying the mask feature image and two or more input feature images used in learning, which information is provided by the learning unit 505, and the difference at the time when learning is completed.The combination search unit 506 then notifies the segmentation unit 504 of the identified combination of the mask feature image and two or more input feature images.

[0140] The processing information generation device 500 described above can also be realized by a computer such as the PC 10 shown in Fig. 3. For example, the mask image conversion unit 508 can also be realized by the processor 13 executing a program.

[0141] 17 is a flowchart showing the processing in the learning phase of the processing information generating device 500 according to the fifth embodiment. First, the processing feature image generating unit 102 generates a plurality of processing feature images showing features when processing the processing object from the CAD data stored in the CAD data storage unit 101 (S80). The plurality of processing feature images generated here each show a different feature. The generated plurality of processing feature images are provided to the segmentation unit 504.

[0142] The segmentation unit 504 selects one of the plurality of processing feature images as a processing feature image for masking, provides the processing feature image for masking to the mask image conversion unit 508, and receives a mask image from the mask image conversion unit 508 in response (S81). The mask image conversion unit 508 generates a mask image from the processing feature image for masking.

[0143] Next, the segmentation unit 504 identifies two or more input processing feature images from the remaining processing feature images generated in step S80, excluding the mask processing feature image (S82).

[0144] The segmentation unit 504 then inputs the two or more input processing feature images identified in step S82 into the learning model stored in the model storage unit 503, and performs masking using the mask image acquired in step S81 to acquire a region classification image (S83). The acquired region classification image and information identifying the mask processing feature image and the two or more input processing feature images are provided to the learning unit 505.

[0145] The learning unit 505 calculates the difference between the region classification image acquired in step S54 and the region classification image indicated by the previously acquired correct answer data (S84).The learning unit 505 then determines whether the difference acquired in step S84 is less than a predetermined threshold, thereby determining whether the learning of the learning model has converged for each of the mask image and two or more processed feature images (S85).If the learning has converged (Yes in S84), the process proceeds to step S87.If the learning has not converged (No in S84), the process proceeds to step S86.

[0146] In step S86, the learning unit 505 learns the learning model so as to reduce the difference in step S84. For example, the learning unit 505 updates the weights of the learning model. Then, the process returns to step S83, and the process is repeated for the combination of the mask image and two or more input processed feature images that are determined not to have converged.

[0147] In step S87, the learning unit 505 notifies the combination search unit 506 of the combination of the mask processed feature image and two or more input processed feature images for which learning has converged, and the difference at which learning has converged, and the combination search unit 506 determines whether the search for combinations has been completed. For example, the combination search unit 506 can determine that the search has been completed when the number of times learning has converged by the learning unit 505 is greater than a predetermined threshold, or when the difference notified by the learning unit 505 is smaller than a predetermined threshold.

[0148] If the search is not complete (No in S87), the process returns to step S81, and the segmentation unit 504 performs learning on the combination of the mask processing feature image and two or more input processing feature images that have not yet been learned. If the search is complete (Yes in S87), the process proceeds to step S88.

[0149] In step S88, the combination of the mask processed feature image used in the learning and two or more input processed feature images with the smallest difference notified by the learning unit 505 is identified as the optimal combination. The optimal combination thus identified is notified to the segmentation unit 504 and used in the inference phase.

[0150] Note that, in step S85 of the flowchart shown in FIG. 17 , convergence is determined based on the difference being less than a threshold value, but the fifth embodiment is not limited to this example. For example, the learning unit 505 may proceed to step S86 after processing step S84, and after processing step S86, determine convergence based on whether the number of times learning in step S86 has been performed exceeds a predetermined number. In such a case, the processing of step S87 may be performed after it is determined that convergence has occurred. Note that, if convergence has not been determined, the processing returns to step S83, and the processing is repeated for the combination of the mask image and two or more input processing feature images that are determined not to have converged.

[0151] 18 is a flowchart showing the processing in the inference phase of the processing information generating device 500 according to the fifth embodiment. First, the processing feature image generating unit 102 generates a plurality of processing feature images showing features when processing the processing object from the CAD data stored in the CAD data storage unit 101 (S90). The plurality of processing feature images generated here each show a different feature. The generated plurality of processing feature images are provided to the segmentation unit 504.

[0152] Next, the segmentation unit 504 identifies, from the plurality of processing feature images, a processing feature image corresponding to the processing feature image for masking in the optimal combination notified by the combination search unit 506, and provides the processing feature image to the mask image conversion unit 508. In response, the mask image conversion unit 508 receives a mask image (S91). The mask image conversion unit 508 generates a mask image from the processing feature image for masking.

[0153] Next, the segmentation unit 504 identifies, from the plurality of processed feature images, two or more processed feature images that correspond to the two or more input processed feature images in the optimal combination notified by the combination search unit 506 (S92).

[0154] Then, the segmentation unit 504 inputs the two or more processed feature images identified in step S92 into the learning model stored in the model memory unit 503, and performs masking processing using the mask image acquired in step S91 to acquire a region classification image (S93).

[0155] Then, the segmentation unit 504 generates a region image, which is an image that visualizes the region, by coloring each region indicated in the region classification image acquired in step S93 with a predetermined color (S94).

[0156] 18, from the plurality of processed feature images generated in step S90, processed feature images corresponding to valid combinations are identified in steps S91 and S92, but the third embodiment is not limited to this example. For example, the processed feature image generating unit 102 may generate only processed feature images corresponding to valid combinations.

[0157] As described above, in the fifth embodiment, important information regarding processing can be input into the learning model that performs semantic segmentation, thereby improving the accuracy of inference.

[0158] 100, 200, 300, 400, 500 Machining information generation device, 101 CAD data storage unit, 102 Machining feature image generation unit, 103, 203, 303, 403, 503 Model storage unit, 104, 204, 304, 404, 504 Segmentation unit, 105, 205, 305, 405, 505 Learning unit, 206, 506 Combination search unit, 307, 407 Combination setting unit, 508 Mask image conversion unit.

Claims

1. A learning device characterized by inputting a plurality of processing feature images, which are a plurality of images that respectively visualize a plurality of features in the processing of an object, and correct answer data that indicates the result of correctly classifying the object into a plurality of processing areas, and training at least one learning model to classify an area that is processed in one process into a single processing area and classify it into the multiple processing areas indicated by the correct answer data.

2. The learning device according to claim 1, characterized in that it comprises: a processing feature image generation unit that generates the plurality of processing feature images from design data indicating the design content of the target; a segmentation unit that classifies the plurality of processing feature images into a plurality of processing regions by inputting the plurality of processing feature images into the learning model; and a learning unit that learns the learning model using the results of the classification and the correct answer data.

3. The learning device according to claim 1, further comprising: a processing feature image generation unit that generates the plurality of processing feature images from design data indicating the design content of the target; a segmentation unit that performs a plurality of classifications into a plurality of processing areas by inputting each of the plurality of processing feature images into a corresponding one of the plurality of learning models; a learning unit that trains the plurality of learning models using the results of the plurality of classifications and the supervised data; and a search unit that searches for two or more processing feature images that are effective in performing a single classification into a plurality of processing areas indicated by the supervised data from the plurality of processing feature images by aggregating the results of the plurality of classifications according to a predetermined rule.

4. The learning device according to claim 1, further comprising: a processing feature image generation unit that generates the plurality of processing feature images from design data indicating the design contents of the target; a segmentation unit that performs a plurality of classifications into a plurality of processing areas by inputting each of the plurality of processing feature images into a corresponding one of the plurality of learning models; a learning unit that trains the plurality of learning models using the results of the plurality of classifications and the correct answer data; and a search unit that searches for weights to be assigned to each of the results of the plurality of classifications when performing a single classification into a plurality of processing areas by aggregating the results of the plurality of classifications according to a predetermined rule.

5. The learning device according to claim 1, comprising: a processing feature image generation unit that generates the plurality of processing feature images from design data indicating the design content of the target; a segmentation unit that performs a plurality of classifications into a plurality of processing regions by inputting each of the plurality of processing feature images into a corresponding one of the plurality of learning models; a learning unit that trains the plurality of learning models using the results of the plurality of classifications and the supervised data; and a search unit that searches for two or more processing feature images from the plurality of processing feature images that are effective for classification into the plurality of processing regions indicated by the supervised data, wherein the learning unit further trains a new learning model using the two or more processing feature images and the supervised data by inputting the two or more processing feature images to classify an area to be processed in one process into one processing region and classify into the plurality of processing regions indicated by the supervised data.

6. The learning device described in claim 5, further comprising a construction unit that extracts two or more encoders from each of two or more learning models that input the two or more processed feature images out of the plurality of learning models, and constructs the new learning model using the two or more encoders, a concatenation layer that concatenates and outputs the outputs of the two or more encoders, and a decoder that decodes the output of the concatenation layer.

7. The learning device according to claim 5, further comprising a construction unit that extracts two or more encoders from each of two or more learning models that input the two or more processed feature images among the plurality of learning models, and constructs the new learning model using the two or more encoders, a weighting layer that adds weights to the outputs of the two or more encoders and outputs them, a concatenation layer that concatenates the outputs of the weighting layer and outputs them, and a decoder that decodes the output of the concatenation layer.

8. The learning device according to claim 1, further comprising: a processing feature image generation unit that generates the plurality of processing feature images from design data indicating the design contents of the target; a segmentation unit that selects one processing feature image from the plurality of processing feature images as a processing feature image for masking, selects two or more processing feature images from the plurality of processing feature images excluding the processing feature image for masking as two or more input processing feature images, and decodes the result of masking using the mask image generated from the processing feature image for masking as output from an encoder, thereby classifying an area to be processed in one process into one processing area, and performs a process of classifying into a plurality of processing areas by changing the combination of the one processing feature image selected as the processing feature image for masking and the two or more processing feature images selected as the two or more input processing feature images; a learning unit that trains the learning model using the result of the classification and the correct answer data; and a search unit that searches for the combination optimal for the classification from the result of learning of the learning model.

9. An inference device characterized by inputting at least two learning processing feature images, which are at least two images that each visualize at least two features of the learning processing that is the processing to be learned, into a learning model that has been trained to classify an area that is processed in one process into one processing area, and to classify into multiple processing areas, by inputting at least two processing feature images, which are at least two images that each visualize at least two features of the processing that is the processing to be learned.

10. An inference device as described in claim 9, comprising: a processing feature image generation unit that generates a plurality of processing feature images, which are a plurality of images that visualize a plurality of features in the processing of the object, from design data that indicates the design content of the object; and a segmentation unit that inputs the plurality of processing feature images into the learning model to classify the processing feature images into a plurality of processing areas, wherein the learning model is a model trained using a plurality of learning processing feature images, which are a plurality of images that visualize a plurality of features in the learning processing, and correct answer data that indicates the result of correctly classifying the learning object into a plurality of processing areas.

11. An inference device according to claim 9, comprising: a processing feature image generation unit that generates two or more processing feature images, which are two or more images that respectively visualize two or more features in the processing of the object, from design data that indicates the design contents of the object; and a segmentation unit that performs two or more classifications into a plurality of processing regions by inputting the two or more processing feature images into two or more corresponding learning models, and aggregating the results of the two or more classifications according to a predetermined rule, thereby performing one classification into a plurality of processing regions, wherein the two or more learning models are two or more models trained using a plurality of learning processing feature images, which are multiple images that respectively visualize a plurality of features in the learning processing, and correct answer data that indicates the result of correctly classifying the learning object into a plurality of processing regions, and the two or more processing feature images are two or more images that are determined to be effective in performing one classification into a plurality of processing regions indicated by the correct answer data, by inputting the plurality of learning processing feature images into a plurality of corresponding learning models, respectively, and aggregating the results of multiple classifications into a plurality of processing regions according to the predetermined rule.

12. An inference device as described in claim 9, comprising: a processing feature image generation unit that generates a plurality of processing feature images, which are a plurality of images that visualize a plurality of features in the processing of the object, from design data that indicates the design content of the object; and a segmentation unit that performs a plurality of classifications into a plurality of processing areas by inputting the plurality of processing feature images into a plurality of corresponding learning models, and performs a single classification into a plurality of processing areas by adding a predetermined weight to each of the results of the plurality of classifications and then aggregating them according to a predetermined rule, wherein the plurality of learning models are a plurality of models trained using a plurality of learning processing feature images, which are a plurality of images that visualize a plurality of features in the learning processing, and correct answer data that indicates the result of correctly classifying the learning object into a plurality of processing areas.

13. The inference device according to claim 9, comprising: a processing feature image generation unit that generates two or more processing feature images, which are two or more images that respectively visualize two or more features in processing of the object, from design data that indicates the design content of the object; and a segmentation unit that inputs the two or more processing feature images into the learning model to classify the object into a plurality of processing areas, wherein the learning model inputs each of a plurality of learning processing feature images, which are a plurality of images that respectively visualize a plurality of features in the learning processing, into a corresponding one of a plurality of learning models to classify an area that is processed in one process into one processing area and perform a plurality of classifications into a plurality of processing areas, trains the plurality of learning models using the results of the plurality of classifications and supervised data that indicates the result of correctly classifying the object into a plurality of processing areas, searches from the plurality of learning processing feature images for two or more learning processing feature images that are effective for classifying the object into a plurality of processing areas indicated by the supervised data, and is a newly trained model using the two or more learning processing feature images found and the supervised data.

14. The inference device described in claim 13, characterized in that the learning model is constructed using two or more encoders extracted from each of two or more learning models among the plurality of learning models that input the two or more searched processed feature images, a concatenation layer that concatenates and outputs the outputs of the two or more encoders, and a decoder that decodes the output of the concatenation layer.

15. The inference device described in claim 13, characterized in that the learning model is constructed from two or more encoders extracted from two or more learning models among the plurality of learning models that input the two or more searched processed feature images, a weighting layer that adds weights to the outputs of the two or more encoders and outputs them, a concatenation layer that concatenates and outputs the outputs of the weighting layer, and a decoder that decodes the output of the concatenation layer.

16. A processing feature image generation unit that generates, from design data indicating the design content of the object, one processing feature image for masking and two or more processing feature images for input that are two or more processing feature images, as multiple images that respectively visualize multiple features in processing of the object; and a segmentation unit that classifies the two or more input processing feature images into multiple processing regions by inputting the two or more input processing feature images into the learning model that decodes the results of masking an output from an encoder using a mask image generated from the processing feature image for masking, The learning model is a model trained using the results of the learning classification and correct answer data indicating the results of correct classification of the learning object into the multiple processing regions, by selecting one learning processed feature image from multiple learning processed feature images, which are multiple images that visualize multiple features in the learning processing, as a processed feature image for a learning mask, selecting two or more learning processed feature images from the multiple learning processed feature images excluding the processed feature image for the learning mask as two or more processed feature images for learning input, inputting the two or more processed feature images for learning input to the encoder, and decoding the results of mask processing performed on the output from the encoder using a learning mask image generated from the processed feature image for the learning mask, and The inference device described in claim 9, characterized in that the processed feature image for masking and the two or more processed feature images for input are images that visualize the features visualized by the processed feature image for learning masking and the processed feature image for learning input, which are searched for as the optimal combination for the learning classification among the combinations from the results of learning of the learning model.

17. A program characterized by inputting a plurality of processing feature images, which are a plurality of images that respectively visualize a plurality of features in the processing of an object, and correct answer data that indicates the results of correctly classifying the object into a plurality of processing areas, into a computer, and training at least one learning model to classify an area that is processed in one process into a single processing area and classify it into the multiple processing areas indicated by the correct answer data.

18. A program characterized by inputting at least two learning processing feature images, which are at least two images that each visualize at least two features of the learning processing that is the processing to be learned, into a computer, thereby classifying an area to be processed in one process into one processing area, and having a learning model that has been trained to classify into multiple processing areas by inputting at least two learning processing feature images, which are at least two images that each visualize at least two features of the processing to be performed, into a computer.

19. A learning method characterized by inputting a plurality of processing feature images, which are a plurality of images that respectively visualize a plurality of features in the processing of an object, and correct answer data that indicates the result of correctly classifying the object into a plurality of processing areas, and training at least one learning model to classify an area that is processed in one process into a single processing area and classify it into a plurality of processing areas indicated by the correct answer data.

20. An inference method characterized by inputting at least two learning processing feature images, which are at least two images that each visualize at least two features of the learning processing that is the processing to be learned, into a learning model that has been trained to classify an area to be processed in one process into one processing area, and to classify into multiple processing areas, by inputting at least two processing feature images, which are at least two images that each visualize at least two features of the processing to be performed, into a learning model that has been trained to classify into multiple processing areas.

Images