Teacher data generation program, etc.

Abstract

To provide a technique for generating supervised data even without a background image in actual use.SOLUTION: Supervised data generation programs 36, 36a each are a program that generates supervised data 17 for generating a trained model 10 that, when an image 11 including a target 12 corresponding to a target image 13 is input, outputs a result from identifying the target. The supervised data generation program causes a computer to: select a first target image 13a from an image group 14 in which a plurality of different target images 13 are prepared; performing transformation processing to generate a background image 15; select a second target image 13b from the image group 14; and perform synthesizing processing for synthesizing the second target image 13b on the background image 15 to generate the supervised date 17.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a technique for detecting a target to be detected from an image. [Background technology]

[0002] In recent years, object detection algorithms such as YOLO (You Only Look Once) and SSD (Single Shot Multibox Detector) have been used to detect targets in a variety of business fields. These object detection learning models require a certain amount of high-quality training data to be prepared and trained in order to correctly detect targets.

[0003] There is also a known technique called data augmentation, which artificially inflates the amount of data by applying processes such as translation, scaling, rotation, and adding noise to training data images.

[0004] For example, Patent Document 1 is a technology related to data augmentation. Patent Document 1 discloses an information processing device for selecting images suitable for machine learning. When a composite image created by arranging multiple element images, each showing a target element, on a background image includes overlapping element images, the information processing device includes: a specifying unit that specifies an occlusion degree, which indicates the degree to which a first element image arranged in the background in the overlapping state is occluded by a second element image arranged in the foreground in the overlapping state; and a selecting unit that selects the composite image as training data to be used in machine learning for generating a recognition model to detect the target element, when the occlusion degree is equal to or less than an upper limit specified according to the complexity of the first element image. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2022-26456 Summary of the Invention [Problem to be solved by the invention]

[0006] Object detection AI separates targets from images that contain a mixture of background images and targets. In machine learning, information (knowledge) about the background image is important in addition to the target's features. Ideally, the learned knowledge of the background image should match the knowledge of the background image in actual operation. However, it is sometimes impossible to obtain background images in operation.

[0007] Therefore, one aspect of the present invention is to provide a technology for generating training data even when a background image during operation is not available. [Means for solving the problem]

[0008] (1) One embodiment of the teacher data generation program of the present invention is a program that generates teacher data for generating a learning model that, when inputted, outputs results of identifying the target when image data of an image including a target corresponding to a target image is inputted, and is characterized in that the program causes a computer to select a first target image from an image group in which multiple types of target images are prepared, perform a transformation process to generate a background image, select a second target image from the image group, and perform a synthesis process to synthesize the second target image with the background image to generate teacher data.

[0009] (2) In such a teacher data generation program, it is preferable that the transformation process is performed on an image obtained by dividing the first target image.

[0010] (3) Another aspect of the teacher data generation method of the present invention is a method for generating teacher data for generating a learning model that, when inputted, outputs a result of identifying a target when image data of an image including a target corresponding to a target image is inputted, and is characterized in that a computer selects a first target image from an image group in which a plurality of types of target images are prepared, performs a transformation process to generate a background image, selects a second target image from the image group, and performs a synthesis process to synthesize the second target image with the background image to generate teacher data.

[0011] (4) Another aspect of the teacher data generation device of the present invention is a device that generates teacher data for generating a learning model that, when inputted with image data of an image including a target corresponding to a target image, outputs the result of identifying the target, and is characterized by comprising a background image generation unit that selects a first target image from an image group in which a plurality of types of target images are prepared, performs a transformation process to generate a background image, and a teacher data generation unit that selects a second target image from the image group, performs a synthesis process to synthesize the second target image with the background image, and generates teacher data.

[0012] (5) Another aspect of the learning device of the present invention is characterized by comprising a background image generation unit that selects a first target image from an image group in which multiple types of target images are prepared and performs a transformation process to generate a background image; a training data generation unit that selects a second target image from the image group and performs a synthesis process to synthesize the second target image with the background image to generate multiple training data; and a learning model generation unit that uses the multiple training data to generate a learning model that outputs a result of identifying the target when image data of an image including a target corresponding to the target image is input.

[0013] (6) Another aspect of the data structure of teacher data of the present invention is a data structure of teacher data having image data of a teacher image for generating a learning model that outputs a result of identifying a target when image data of an image including a target corresponding to a target image is input, characterized in that the teacher image consists of a second target image selected from a group of multiple target images corresponding to the targets to be mutually identified by the learning model, and a background image arranged around the second target image, and the background image is formed using a deformed image obtained by deforming a first target image selected from the group of images.

[0014] (7) It is preferable that the data structure of such training data further includes position information of the second target image relative to the background image. [Effects of the Invention]

[0015] It is possible to efficiently generate training data, etc. [Brief explanation of the drawings]

[0016] [Figure 1] FIG. 1 is a functional block diagram illustrating an embodiment of a learning model generation device. [Figure 2] 10 is a flow chart illustrating an embodiment of a process flow for identifying a target. [Figure 3] FIG. 2 is a schematic diagram illustrating an example of a hardware configuration. [Figure 4] 10 is a flow chart illustrating an embodiment of a process flow for generating a target image. [Figure 5] 10 is a flowchart illustrating an embodiment of a process flow for generating a background image. [Figure 6] FIG. 2 is a schematic diagram illustrating an embodiment of a data structure of training data. [Figure 7] FIG. 1 is a schematic diagram illustrating an example of the data structure of a training data set database. [Figure 8] FIG. 10 is a schematic diagram illustrating an example of a target detection result using a learning model. [Figure 9] FIG. 1 is a schematic diagram showing an image of Comparative Example 1. [Figure 10] FIG. 10 is a schematic diagram showing images of Comparative Examples 2 and 3. [Figure 11] FIG. 10 is a schematic diagram showing another embodiment of a teacher image of the teacher data. [Figure 12] FIG. 10 is a schematic diagram showing another embodiment of a teacher image of the teacher data. [Figure 13] FIG. 10 is a schematic diagram showing another embodiment of a teacher image of the teacher data. DETAILED DESCRIPTION OF THE INVENTION

[0017] [1. Brief description] (Introduction) The training data generated in this embodiment (see FIG. 6) has a distinctive background image. The background image is generated from an image obtained by transforming a target image, which is the object (target) of recognition. Because the background image is a confusingly similar image to the target image, it provides the learning model with rich background information, improving learning efficiency.

[0018] For example, there are multiple types of buttons / icons displayed in an app. From the perspective of UI (User Interface) / UX (User Experience), the style (color, pattern) of multiple types of buttons / icons tends to be similar to the style of the entire background of the display including the buttons / icons. For this reason, the background image generated with multiple types of icons will not match the characteristics of the app's background (color distribution, pattern, etc.). Tar It is easy to reflect the changes in the students' learning patterns, making learning more efficient.

[0019] (Learning Model 10) First, a learning model generated by machine learning training data of this embodiment will be described with reference to Fig. 2. The learning model 10 shown in the figure receives image data of an image 11 as input and outputs the result of identifying a target 12. As the learning model 10, for example, an object detection algorithm such as YOLO (You Only Look Once) or SSD (Single Shot MultiBox Detector) is used.

[0020] (Image 11) The image 11 is an image containing a target 12 to be recognized by the learning model 10. It may be a still image or a video. In the case of a video, it may be an image taken at a predetermined time interval. For example, it may be a capture of an image displayed on a screen, or an image taken by a camera, an in-vehicle camera, a surveillance camera, or the like.

[0021] (target) The target is an object to be detected that is included in the image 11. When the target is detected by the learning model, it is displayed, for example, by surrounding it with a frame (not shown). Alternatively, it may be displayed with an arrow pointing at it or by adding color so that it can be distinguished from other parts.

[0022] (Teacher data generation device 1, learning device 4) The teacher data generation device will be described with reference to Fig. 1. The teacher data generation device 1 shown in the figure comprises a background image generation unit 2 and a teacher data generation unit 3. The learning device 4 includes a teacher data generating device 1 and a learning model generating unit 5.

[0023] [2.Each configuration] (Background image generation part 2) The background image generating unit 2 selects a first target image 13a from an image group 14 in which a plurality of types of target images 13 are prepared, and generates a background image 15 by performing transformation processing including enlargement, reduction, rotation, or inversion.

[0024] (Target images 13, 13a) The target image 13 is an image corresponding to the target included in the image 11 . The first target image 13 a consists of one or more target images 13 that are used to generate the background image 15 .

[0025] (Image set 14) The image group 14 consists of a plurality of target images 13 to be classified. The target images 13 in the image group 14 consist of images to be classified by the learning model 10. The image group 14 is stored in a storage unit 31, which will be described later. Alternatively, the image group 14 may be stored in an external server.

[0026] (Background image 15) The background image 15 is generated by arranging the deformed images 16 in the background frame 15a. In this embodiment, the deformed images 16 are arranged randomly until the background frame 15a is filled. When arranging, the deformed images 16 may overlap each other or may not overlap. The rectangle indicated by the reference numeral 16 in the background frame 15a indicates a part of how the deformed images 16 are arranged.

[0027] (deformed image 16, deformation processing) The deformed image 16 is generated by performing deformation processing such as enlarging, reducing, rotating, flipping upside down, or flipping left and right on the first target image 13a selected from the image group 14. Other processing such as adding noise or projective transformation may also be performed. Alternatively, for example, a general data augmentation technique may be used. The transformation process corresponds to steps R2 and R3 in FIG. 5, which will be described later.

[0028] (Teacher Data Generation Unit 3) The training data generation unit 3 selects a second target image 13 from the image group 14. b is selected, and a synthesis process is performed to synthesize the second target image 13b onto the background image 15, thereby generating training data 17 (see FIG. 6).

[0029] (Second target image 13b) The second target image 13 b consists of one or more target images 13 that are used to composite onto the background image 15 .

[0030] (Teacher data 17) The training data 17 is data used for supervised learning in machine learning. In this embodiment, the training data 17 includes a background image 15, a target image 13, and position information 18. The training data 17 will be described in detail in the description of FIG. 6.

[0031] [3. Hardware configuration] Next, the hardware configuration of the teacher data generation device 1 and the learning device 4 will be described with reference to FIG.

[0032] (Hardware configuration of the teacher data generation device 1 and the learning device 4) We will now explain the hardware configuration of the teacher data generation device 1. The hardware configuration of the learning device 4 is almost the same as the hardware configuration of the teacher data generation device 1, so a description thereof will be omitted.

[0033] 3, the teacher data generation device 1 of this embodiment is a computer equipped with a CPU (or GPU) 30. The CPU 30 is connected via a bus line 35 to, for example, a memory (hereinafter referred to as a storage unit) 31, a connection port 33 for connecting / reading a storage device 32, and a communication circuit 34 for communicating with the outside via a network.

[0034] The storage unit 31 stores programs 36 and 36a for generating training data and generating a learning model, respectively. A browser program 37 and an OS (operating system) 38 may also be stored. The programs 36 and 36a are installed, for example, by the storage device 32. In this embodiment, the storage unit 31 stores the target image 13 (image group 14), training data 17, and learning model 10, but these may also be stored in an external server device.

[0035] In this embodiment, the programs 36 and 36a may operate in cooperation with each other by utilizing the functions of the OS 38 and the browser program 37. Note that the programs 36 and 36a may also operate independently without utilizing the browser program 37 and the OS 38, respectively.

[0036] In the hardware configuration of the programs 36 and 36a described above, the functions shown in the functional block diagram of FIG. 1 are realized, for example, using a CPU 30 and the programs 36 and 36a, but some or all of them may be sequence-controlled using a logic circuit such as a microcomputer or a PLC (programmable logic controller).

[0037] [4. Program] (Flowchart showing the process of generating training data) FIG. 4 is a flowchart showing an embodiment of the processing of the programs 36 and 36a used in the teacher data generation device 1. The following flowchart is for generating one piece of teacher data 17. In practice, the following flowchart is used to generate a large amount of different types of teacher data 17. When generating, multiple pieces of teacher data 17 may be generated simultaneously by parallel processing.

[0038] (S0: Pre-treatment process) In a pre-processing step S0 (not shown), the file format of the image data of the target image 13 to be processed by the program 36 is adjusted to a predetermined format.

[0039] (S1: Generate background image) The CPU 30 of the teacher data generation device 1 (see FIG. 3) generates a background image 15.

[0040] Next, the generation of a background image will be described with reference to Fig. 5. Fig. 5 is a flowchart showing an embodiment of the processing of the program 36a used in the teacher data generation device 1.

[0041] (R1: Select target image) A first target image 13a is randomly selected from the image set 14 (see FIG. 1).

[0042] (R2: Cut) For the selected first target image 13a, either the vertical or horizontal direction is randomly selected. The vertical or horizontal direction is determined for each target image 13. Either 3, 4, or 5 divisions is randomly selected. The image is divided into the selected number of divisions parallel to the selected direction. One image is randomly selected from the 3 to 5 images formed by the division.

[0043] (R3: Transformation processing) The selected divided image is transformed to generate a transformed image 16 (see FIG. 6). A plurality of transformation processes are registered in advance, and a transformation process is selected at random from among them. In this embodiment, one transformation is selected at random from the registered transformation processes of rotation / inversion only, enlargement / reduction only, and enlargement / reduction + rotation / inversion. The rectangle indicated by the reference numeral 16 in the background frame 15a indicates a part of how the transformed image 16 is arranged.

[0044] (R4: Place in background frame) The deformed images 16 are arranged in the background frame 15a. In this embodiment, they are arranged randomly in the background frame 15a. In this embodiment, the deformed images 16 may overlap each other. Furthermore, the deformed image 16 arranged later is displayed on top.

[0045] (R5: Is the background frame filled?) If the background frame 15a is filled with the deformed image 16, the process proceeds to the next step, and if it is not filled, the process returns to step R1. Note that the background frame 15a does not have to be completely filled, and the process may proceed to the next step as long as a predetermined amount is filled. The predetermined amount is, for example, 90 to 95% of the area within the background frame 15a.

[0046] (R6: Background completed) Background image 15 is completed.

[0047] (S2: Select the second target image) 4, the second target image 13b is selected at random, and may include all or part of the image selected as the first target image 13a. When selecting, if a specific second target image 13b has already been selected many times in other training data 17, the probability of selection is adjusted to decrease. On the other hand, if only a small number of images have been selected, the probability of selection is adjusted to increase. Adjustments are made so that the number of selected second target images 13b in the training data set 20 is the same / approximately the same. In this embodiment, for example, a roulette selection method is adopted to adjust the number of second target images 13b to be selected, but other known methods may also be adopted. Furthermore, when the generation of the training data 17 is processed in parallel, the number of specific second target images 13b in each process is added up, and in the next selection of each process, the probability of selection is increased / decreased, and dynamically adjusted so that the probability of each second target image 13b being selected is balanced.

[0048] (S3: Composite) The second target image 13b is randomly placed on the completed background image 15. The second target image 13b is composited so as to be displayed on top of the background image 15. Also, position information 18 of the second target image 13b relative to the background frame 15a is stored. The position serving as the starting point of the second target image 13b corresponding to the position information 18 is set to the center, center of gravity, or a corner of a rectangle. As information about the second target image 13b, for example, the shape such as square, rectangle, circle, etc., and the size such as the length of a side and the diameter may be stored together with the position information 18.

[0049] (S4: Is the target image displayed appropriately?) It is determined whether the synthesized second target image 13b is properly displayed on the background image 15. For example, if none of the following (a, b, c) applies, it is determined to be proper, and the process proceeds to the next step. (a) A plurality of second target images 13b overlap each other. (b) The second target image 13b is larger than the background frame 15a (c) The second target image 13b extends beyond the background frame 15a If any of the above (a, b, c) applies, the synthesis in step S3 is canceled and the process returns to step S2.

[0050] (S5: Link location information) The position information 18 of the second target image 13b combined with the background image 15 is linked and stored in the storage unit 31.

[0051] (S6: Is the number of target images appropriate?) If the number of second target images 13b does not reach the predetermined number, the process returns to step S2, and if the number reaches the predetermined number, the process ends. In this embodiment, the number of second target images 13b is determined randomly within a range of 10 to 20. The number of second target images 13b is determined between step S1 and step S2. However, the number may be determined in another step.

[0052] (Other processing) In the above-described processes, some or all of the processes that are randomly selected or arranged may be selected or arranged in a predetermined order. Furthermore, the cropping step R2 may be omitted. In this case, it may be difficult to distinguish between the second target image 13b to be identified and the transformed image 16 of the background image, so that the transformation process of only enlarging / reducing or only rotating / flipping may not be performed. In generating the background image, the deformed images 16 may be arranged so as not to overlap each other.

[0053] (Teacher data 17) 6 is a schematic diagram showing an example of the data structure of the teacher data. The teacher data 17 shown in the figure includes the same components as those described in the teacher data generation device 1, so the same components are given the same reference numerals and their description will be omitted. The training data 17 shown in the figure is used for machine learning to generate a learning model 10 that, when inputted with image data of an image 11 including a target 12 corresponding to a target image 13, outputs a result of identifying the target 12. The training data 17 consists of a training image 19 and position information 18.

[0054] (Teacher image 19) The teacher image 19 consists of a second target image 13b selected from an image group 14 of multiple target images 13 corresponding to targets 12 to be mutually discriminated by the learning model 10, and a background image 15 arranged around the second target image 13b. The background image 15 is formed from a transformed image 16 obtained by transforming a first target image 13a selected from the image group 14.

[0055] (training dataset) 7 is a schematic diagram showing an example of the data structure of a training dataset. The training dataset 20 shown in the figure consists of training data 17 and a target image 13. For example, training data T1 is used to learn a target image L1 through machine learning.

[0056] (Machine Learning) Returning to FIG. 1 , the learning model generation unit 5 generates the learning model 10 by repeated learning based on the teacher dataset 20 stored in the storage unit 31. Various methods can be used to generate the learning model 10. The learning model generation unit 5 stores the generated learning model 10 in the storage unit 31.

[0057] (Detection result 21) Figure 8 is a schematic diagram showing an example of the detection results of a target 12 using a learning model. The detection results 21 shown in the figure include the detection number, backgrounds 1 to 4, the number of generated images, the number of epochs, the learning time, the number of correctly detected targets, the number of undetected targets, and the number of falsely detected targets. The total number of targets to be detected is 20.

[0058] No. in the table is a number that identifies the training dataset 20. The number of generated images is the number of training data in the training dataset 20. The number of epochs is the number of repeated learning. The training time is the time required to generate a training model. The number of correctly detected targets is the number of targets that were correctly detected. The number of undetected targets is the number of targets that could not be detected. The number of falsely detected targets is the number of targets that were incorrectly detected when they were not targets.

[0059] (Background 1) Background 1 is a background image 15 (see FIG. 6) generated by the method of this embodiment.

[0060] (Background 2: Comparative Example 1) Figure 9 is a schematic diagram showing an image of Comparative Example 1. The background image 21a of the teacher image 21 shown in the figure is an image with a lot of information, such as an image of open data. It corresponds to background 2 in Figure 8. Reference numeral 13 is a target image.

[0061] (Background 3: Comparative Example 2, Background 4: Comparative Example 3) FIG. 10 is a schematic diagram showing images of Comparative Examples 2 and 3. The background image 22a of the teacher image 22 of Comparative Example 2 shown in the figure is a noise image. This corresponds to background 3 in FIG. 8. The background image 23a of the teacher image 23 of Comparative Example 3 shown in the figure is a single light green color. This corresponds to background 4 in FIG. 8. Reference numeral 13 denotes a target image. Background 4 For example, a method for generating a background is to first select several targets to be placed in the background, statistically calculate the color distribution of the selected targets, and then generate a background with a single color that is the average or median of the color distribution.

[0062] Returning to Figure 8, for Detection No. 1, 200 pieces of training data for each of the background images of Backgrounds 1, 2, 3, and 4 were prepared, and a total of 800 pieces of training data were used for learning. In addition, for Detection No. 2, 800 pieces of training data for background images of Background 1 were prepared and trained. As with Detection No. 2, Detection Nos. 3, 4, and 5 were trained with 800 pieces of training data for background images of Backgrounds 2, 3, and 4, respectively. In Detection No. 6, 400 training data sets for background images 1 and 2 were prepared, for a total of 800 training data sets. o8 For each of the models 1 and 9, 400 training data sets for background 1 and 2 were prepared, and a total of 800 training data sets were used for training.

[0063] According to the data of detection Nos. 2 and 3, the number of correctly detected targets is good when background 1 or background 2 is used as training data. That is, the background image 15 of this embodiment can obtain results comparable to those when open data is used as the background image.

[0064] Furthermore, according to the data of detection Nos. 6-9, good results were obtained when background 1 and background 2 were used as training data. In other words, if background 1 and background 2 are used as training data, the number of correct target detections increases.

[0065] (Other background images) In the training data 17, the target images 13 that form the background image 15 are randomly selected and subjected to random deformation processes such as enlargement. As a result, there is a range in the number of deformed images 16 used for the background image 15 placed within the background frame 15a. Figures 11, 12, and 13 show examples with few, average, and many deformed images 16, respectively.

[0066] 11 is a schematic diagram showing another embodiment of the training data 17. The background image 24a of the training image 24 shown in the figure has a smaller number of deformed images 16 arranged within the background frame 15a than the average in the training data set 20.

[0067] 12 is a schematic diagram showing another embodiment of the teacher image 15 of the teacher data 17. The background image 25a of the teacher image 25 shown in the figure has an average number of deformed images 16 arranged within the background frame 15a in the teacher data set 20.

[0068] 13 is a schematic diagram showing another embodiment of the training image 15 of the training data 17. The background image 26a of the training image 26 shown in the figure has a higher number of deformed images 16 arranged within the background frame 15a than the average in the training data set 20.

[0069] 5. Other Embodiments Next, another embodiment will be described. The modified examples described below are almost the same as the teacher data 17 described above, so the same parts are given the same reference numerals and the description thereof will be omitted.

[0070] (Variation) In the modified example, the background image uses the target image 13 as well as an image other than the target image.

[0071] [6. Other] The items described as "other" in the above-described embodiment can be used in appropriate combination.

[0072] [7. Summary] (1) The teacher data generation program 36, 36a (teacher data generation method 27, teacher data generation device 1) is a program that generates teacher data 17 for generating a learning model 10 that outputs a result of identifying a target 12 when image data of an image 11 including a target 12 corresponding to a target image 13 is input, and causes a computer to select a first target image 13a from an image group 14 in which multiple types of target images 13 are prepared, perform a transformation process to generate a background image 15, select a second target image 13b from the image group 14, and perform a synthesis process to synthesize the second target image 13b with the background image 15 to generate the teacher data 17. thing It is characterized by:

[0073] The background image 15 is a confusingly similar image to the target image 13, providing the learning model with rich background information and improving learning efficiency. In addition, the background image 15 can be generated from the target image 13 without preparing open data.

[0074] (2) In such teacher data generation programs 36 and 36a, the first target image 13a is divided and the transformation process is performed on the divided images. teacher Data 17 can be easily increased.

[0075] (3) Teaching Shide The data structure of the training data 17 is a data structure of training data 17 having image data of training images 19 for generating a learning model 10 that, when inputted, outputs a result of identifying the target 12 when image data of an image 11 including a target 12 corresponding to the target image 13 is inputted, and the training image 19 is characterized in that it comprises a second target image 13b selected from an image group 14 of a plurality of target images 13 corresponding to the targets 12 to be mutually identified by the learning model 10, and a background image 15 arranged around the second target image 13b, and the background image 15 is formed using a transformed image 16 obtained by transforming a first target image 13a selected from the image group 14.

[0076] A learning model generated from training data 17 using background image 15 generated from target image 13 can achieve the same level of detection ability for target 12 as a learning model generated from training data using background image 21a generated from open data.

[0077] (4) Such a data structure of the training data can be used as training data 17 because it further includes position information 18 of the second target image 13b relative to the background image 15.

[0078] (5) The learning model 10 is characterized in that a computer selects a first target image 13a from an image group 14 in which multiple types of target images 13 are prepared, performs a transformation process to generate a background image 15, selects a second target image 13b from the image group 14, performs a synthesis process to synthesize the second target image 13b with the background image 15, generates multiple training data 17, and generates the learning model 10 using the multiple training data 17, when image data of an image 11 including a target 12 corresponding to the target image 13 is input, the learning model 10 outputs a result of identifying the target 12.

[0079] The learning model 10 can obtain the same level of detection capability for the target 12 as a learning model generated from training data using the background image 21a generated from open data. [Explanation of symbols]

[0080] 1. Teacher data generation device 2. Background image generation section 3. Training data generation unit 4 Learning Device 5 Learning model generation unit 10 Learning Model 11 Captured images 12 Target 13 Target Image 13a First target image 13b Second target image 14 Image Group 15 background images 15a Background frame 16 Deformed Images 17 Training data 18 Location information 19 Teacher Images 20 Training Dataset 21 Teacher Images 21a Background Image 22 Teacher Images 22a Background image 23 Teacher Images 23a Background image 24 Teacher Images 24a background image 25 Teacher Images 25a Background image 26 Teacher Images 26a Background image 27 Training data generation method 30 CPU 31 memory 32 Recording Devices 33 connection ports 34 Communication Circuit 35 Bus Line 36, 36a Program 37 Browser Program 38 OS

Claims

1. A program for generating training data for generating a learning model that receives image data of an image including a target and outputs a result of identifying the target, the program comprising: On the computer, Selecting one or more first target images randomly or in a preset order from an image group in which target images corresponding to a plurality of types of targets to be identified by the learning model are prepared, dividing the first target image and performing a deformation process on the resulting divided images to obtain deformed images, and arranging the deformed images randomly or in a preset order until the background frame is filled or a predetermined amount is filled to generate a background image, selecting a second target image from the image group, and synthesizing the second target image with the background image to generate a teacher image; generating the teacher image and position information indicating the position of the second target image relative to the teacher image as a set of teacher data; Teacher data generation program.

2. A method for generating training data for generating a learning model that receives image data of an image including a target and outputs a result of identifying the target, comprising: The computer Selecting one or more first target images randomly or in a predetermined order from an image group in which target images corresponding to a plurality of types of targets to be identified by the learning model are prepared, dividing the first target image and performing a deformation process on the resulting divided images to obtain deformed images, and arranging the deformed images randomly or in a predetermined order until the background frame is filled or a predetermined amount is filled to generate a background image; selecting a second target image from the group of images, and synthesizing the second target image with the background image to generate a teacher image; A teacher data generation method for generating the teacher image and position information indicating the position of the second target image relative to the teacher image as a set of teacher data.

3. An apparatus for generating training data for generating a learning model that receives image data of an image including a target and outputs a result of identifying the target, the apparatus comprising: a background image generating unit that randomly or in a predetermined order selects one or more first target images from an image group in which target images corresponding to a plurality of types of targets to be identified by the learning model are prepared, divides the first target image, performs a deformation process on the resulting divided images to obtain deformed images, and generates a background image by arranging the deformed images randomly or in a predetermined order until the background frame is filled or a predetermined amount is filled; A teacher data generation device comprising: a teacher data generation unit that selects a second target image from the group of images, performs a synthesis process to synthesize the second target image with the background image to generate a teacher image, and generates a set of teacher data consisting of the teacher image and positional information indicating the position of the second target image relative to the teacher image.

4. a background image generating unit that randomly or in a predetermined order selects one or more first target images from an image group in which target images corresponding to a plurality of types of targets are prepared, divides the first target image, performs a transformation process on the resulting divided images to obtain transformed images, and generates a background image by arranging the transformed images randomly or in a predetermined order until the background frame is filled or a predetermined amount is filled; a teacher data generation unit that selects a second target image from the image group, performs a synthesis process to synthesize the second target image with the background image to generate a teacher image, and generates a set of teacher data including the teacher image and position information indicating the position of the second target image relative to the teacher image; a learning model generation unit that uses a plurality of the training data to generate a learning model that outputs a result of identifying the target when image data of an image including the target is input.

5. A method for producing training data for generating a learning model that receives image data of an image including a target and outputs a result of identifying the target, comprising: The computer executes a step of randomly or in a predetermined order selecting one or more first target images from an image group in which target images corresponding to a plurality of types of targets to be identified by the learning model are prepared, dividing the first target image and performing a deformation process on the resulting divided images to obtain deformed images, and arranging the deformed images randomly or in a predetermined order until the background frame is filled or a predetermined amount is filled to generate a background image; A teacher data production method comprising the steps of selecting a second target image from the group of images, synthesizing the second target image with the background image to generate a teacher image, and generating the teacher image and positional information indicating the position of the second target image relative to the teacher image as a set of teacher data.

Images