Ai model re-training system

Abstract

This invention enables re-training of an AI model by efficiently collecting training data effective for re-training the AI model.　In an image processing system according to one embodiment of the present invention: an environment parameter generation unit 401 generates a parameter group defining a virtual 3D space; a pseudo image generation unit 402 and an annotation data generation unit 403 generate a pseudo image group and an annotation data group on the basis of the parameter group; a pseudo training data evaluation unit 404 performs AI processing using an AI model 211 for the pseudo image group, and compares the results of the AI processing to the annotation data group to determine whether the pseudo image group and the annotation data group are effective for re-training the AI model 211; and a re-training unit 303 re-trains the AI model 211 by using the pseudo image group and the annotation data group determined to be valid for re-training the AI model 211.

Description

AI model re-learning system

[0001] The present invention relates to an AI model re-learning system that re-learns an AI model used in image processing.

[0002] In recent years, image processing systems have become mainstream, using image processing techniques that use AI (Artificial Intelligence) technology. Such image processing systems require a trained AI model, and re-training of the AI ​​model is performed to prevent performance degradation due to domain compatibility and aging. Re-training requires new training data consisting of images and annotation data. Conventionally, new training data has been obtained by methods such as manually collecting images and creating annotation data, automatically generating annotation data for manually collected images, and directly creating training data using a CG model. The AI ​​model is re-trained using the training data obtained in this manner.

[0003]

[0003] Prior art in the technical field related to the present invention includes the following. For example, Patent Document 1 discloses a technology for generating training data including a synthetic image including a CG model of an object and a teacher signal of the object, and using the training data to learn a recognition function that recognizes information about the object from the synthetic image through neuro-computation. For example, Patent Document 2 discloses a technology for inferring object detection in video frames using a trained model, detecting missed video frames in which object detection was not performed by the trained model from multiple video frames inferred over time, interpolating detection results for the detected missed frames from the multiple video frames, and generating training data using the interpolated detection results.

[0004] Patent No. 7167668 International Publication No. 2022 / 202178

[0005] In conventional image processing systems, when relearning an AI model, there is a problem in that it is unclear what training data is effective for the existing AI model that has been built in. Therefore, it is necessary for humans to collect training data that is predicted to be effective for the existing AI model, which takes an enormous amount of time.

[0006] The present invention has been made in consideration of the above-described conventional circumstances, and aims to efficiently collect learning data that is effective for relearning an AI model, thereby enabling the relearning of the AI ​​model.

[0007] In order to achieve the above object, an AI model re-learning system according to one aspect of the present invention has the following technical features: That is, an AI model re-learning system that re-learns an AI model used for image processing, comprising: a data generation unit that generates a group of images and a group of annotation data that are candidates for learning data to be used for re-learning the AI ​​model based on a group of parameters that define a virtual 3D space, a data evaluation unit that performs image processing on the group of images using the AI ​​model and compares the results of the image processing with the group of annotation data to determine whether the group of images and the group of annotation data are effective for re-learning the AI ​​model, and a re-learning unit that re-learns the AI ​​model using the group of images and the group of annotation data that have been determined to be effective for re-learning the AI ​​model.

[0008] Here, the above-described AI model re-learning system may further include a parameter generation unit that generates a plurality of parameters, including one or more parameters related to a camera that captures the virtual 3D space, one or more parameters related to a 3D object to be placed in the virtual 3D space, one or more parameters related to a light source present in the virtual 3D space, one or more parameters related to a background of the virtual 3D space, and one or more parameters related to occlusion that causes a shadow to be cast on the target, and the parameter generation unit may repeatedly generate the parameter group while changing the value of at least one parameter in the parameter group until the image group and the annotation data group that are determined to be effective for re-learning the AI ​​model are found.

[0009] Furthermore, in the above-described AI model re-learning system, the parameter generation unit may fix the value of at least one parameter in the parameter group and repeat generating the parameter group so as to match the usage conditions of the camera that captures the image that is to be subjected to image processing by the AI ​​model.

[0010] In addition, in the above-mentioned AI model re-learning system, the re-learning unit may use a portion of the image group and the annotation data group that were not determined to be effective for re-learning the AI ​​model, for re-learning the AI ​​model.

[0011] According to the present invention, it becomes possible to efficiently collect learning data that is effective for relearning an AI model, and to relearn the AI ​​model.

[0012] Fig. 1 is a block diagram showing an example of the configuration of an image processing system according to an embodiment of the present invention. Fig. 2 is a block diagram showing an example of the configuration of an AI update unit in the image processing device in the image processing system of Fig. 1. Fig. 3 is a block diagram showing an example of the configuration of a pseudo learning data generation unit in the AI ​​update unit of Fig. 3. Fig. 4 is a flowchart showing an example of processing by the pseudo learning data generation unit of Fig. 4. Fig. 5 is a diagram showing example data of an AI model database in the image processing system of Fig. 1.

[0013] An embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing an example of the configuration of an image processing system according to an embodiment of the present invention. The image processing system in the figure is realized by a video capture device 101, an image processing device 102, a user 103, and a display output device 104. The video capture device 101 can capture real-time or past video and still images. The image processing device 102 can perform various processes, such as image processing, on the video and still images captured by the video capture device 101. The user 103 can issue various instructions, such as re-learning an AI model incorporated in the image processing device 102. The display output device 104 can display and output the processing results of the image processing device 102 to an external device in various formats.

[0014] The video capture device 101 may be an industrial camera or surveillance camera capable of capturing real-time or past video or still images, a USB (Universal Serial Bus) camera, a smartphone, a recorder or player capable of playing a CD (Compact Disc), an HDD (Hard Disc Drive), or a Blu-ray (registered trademark) Disc, or any other camera or video recording device.

[0015] The image processing device 102 is configured as hardware including arithmetic units such as a CPU (Central Processing Unit), a DSP (Digital Signal Processor), an FPGA (Field-Programmable Gate Array), a GPU (Graphics Processing Unit), a memory, a recording device, etc., and may also include other electronic devices and hardware. In FIG. 1, the image processing device 102 is configured by one device, but the image processing device 102 may also be configured by multiple devices that can operate in cooperation with each other.

[0016] The display output device 104 is a device that can display, in some form, data received from the image processing device 102, more specifically, data of the AI ​​processing results received from the information transmission unit 204 described below. The display output device 104 is realized, for example, by a display, a smartphone, a mobile phone, a business radio, an alarm device such as an alarm, or the like, but may also be realized by other devices.

[0017] As shown in the configuration example of Figure 2, the image processing device 102 has a still image acquisition unit 201, an AI processing unit 202, an AI update unit 203, an information transmission unit 204, an AI model 211, and an AI model database 212.

[0018] The still image acquisition unit 201 acquires an input still image from the video acquisition device 101. The AI ​​processing unit 202 performs image processing (hereinafter referred to as "AI processing") on the input still image using an AI model 211 and outputs the AI ​​processing result. The AI ​​update unit 203 updates the AI ​​model 211 in cooperation with the AI ​​model database 212 and the AI ​​model 211 when a re-learning instruction is received from the user 103 or when it is determined that the AI ​​processing result by the AI ​​processing unit 202 needs to be updated. The information transmission unit 204 transmits the AI ​​processing result output from the AI ​​processing unit 202 to the outside in the form of an image, text, audio, compressed data thereof, or other information format. The configurations and operations of the still image acquisition unit 201, the AI ​​processing unit 202, the AI ​​update unit 203, and the information transmission unit 204 will be described in detail below.

[0019] When video (moving images) is acquired by the video acquisition device 101, the still image acquisition unit 201 divides the video into single frames and acquires each frame as an input still image. Furthermore, when still images are acquired by the video acquisition device 101, the still image acquisition unit 201 acquires the images as input still images as they are. When dividing the video into single frames, the video may be thinned out every few frames before division. Here, the color space of the input still image is expressed in RGB (Red, Green, Blue), HSV (Hue, Saturation, Value), HLS (Hue, Luminance, Saturation), or other color spaces. Furthermore, the still image acquisition unit 201 may perform preprocessing such as a smoothing filter, an edge enhancement filter, brightness conversion, or histogram equalization to reduce the effects of noise, flicker, and the like. In addition, the still image acquisition unit 201 may perform enlargement or reduction processing on the input still image to a predetermined size in order to improve the performance of the AI ​​processing unit 202 and reduce processing costs.

[0020] The AI ​​processing unit 202 performs AI processing using the AI ​​model 211 on the input still image acquired by the still image acquisition unit 201, and outputs the AI ​​processing result. Here, the AI ​​processing may be object recognition, object detection, object tracking, segmentation, anomaly detection, style transfer, super-resolution, image restoration, or other AI processing. The AI ​​processing result may also include the degree of degradation of the AI ​​model 211. The degree of degradation is an index indicating that the AI ​​model 211 is more degraded as the degree of degradation increases, and is expressed, for example, as a value ranging from 0 to 1. The degree of degradation may be calculated from an evaluation value for test data prepared in advance, or, if the AI ​​model 211 has a function to directly calculate its own degree of degradation, that value may be used. Evaluation values ​​include the AI's accuracy rate, precision, recall, F-measure, IoU (Intersection over Union), MOTA (Multi-Object Tracking Accuracy), SSIM (Structural Similarity), etc., and can be selected according to the AI ​​processing task.

[0021] The information transmission unit 204 transmits the AI ​​processing results output from the AI ​​processing unit 202 in a form compatible with the display output device 104 using images, text, audio, compressed data of these, or other information forms.

[0022] When the AI ​​model 211 used in the AI ​​processing in the AI ​​processing unit 202 needs to be updated, the AI ​​update unit 203 performs re-learning using pseudo learning data to update the AI ​​model 211. As shown in an example configuration in FIG. 3 , the AI ​​update unit 203 has an AI update determination unit 301, a pseudo learning data generation unit 302, a re-learning unit 303, and an update unit 304.

[0023] The AI ​​update determination unit 301 determines the need to update the AI ​​model 211 based on instructions from the user 103 or the contents of the AI ​​processing results output from the AI ​​processing unit 202. The pseudo learning data generation unit 302 generates pseudo learning data for re-learning. The relearning unit 303 re-learns the AI ​​model using the pseudo learning data generated by the pseudo learning data generation unit 302, and generates a new learning model. The update unit 304 updates the AI ​​model 211 used in AI processing by the AI ​​processing unit 202 with the new AI model generated by relearning in the relearning unit 303. The configurations and operations of the AI ​​update determination unit 301, pseudo learning data generation unit 302, relearning unit 303, and update unit 304 will be described in detail below.

[0024] The AI ​​update determination unit 301 determines whether the AI ​​model 211 needs to be updated based on an instruction from the user 103 or the degree of degradation of the AI ​​model included in the AI ​​processing result output from the AI ​​processing unit 202. For example, when command information for updating the AI ​​model is received from the user 103, the AI ​​update determination unit 301 determines whether the AI ​​model 211 needs to be updated. The AI ​​update determination unit 301 also determines whether the AI ​​model 211 needs to be updated when the degree of degradation of the AI ​​model included in the AI ​​processing result output from the AI ​​processing unit 202 is greater than a preset threshold T1. If it is determined that the AI ​​model 211 needs to be updated, the AI ​​update determination unit 301 operates the pseudo learning data generation unit 302, the relearning unit 303, and the update unit 304 to perform processing to update the AI ​​model 211. If it is determined that the AI ​​model 211 does not need to be updated, the AI ​​model 211 is not updated. If the instruction from the user 103 is to update the AI ​​model 211 using an AI model stored in the AI ​​model database 212, the update unit 304 operates to perform processing to update the AI ​​model 211.

[0025] The pseudo learning data generation unit 302 generates pseudo learning data for relearning the AI ​​model. The pseudo learning data is data that combines a group of pseudo images and a group of annotation data, which will be described later. As shown in FIG. 4 , the pseudo learning data generation unit 302 includes an environmental parameter generation unit 401, a pseudo image generation unit 402, an annotation data generation unit 403, and a pseudo learning data evaluation unit 404.

[0026] The environmental parameter generation unit 401 generates a group of parameters that define a virtual 3D space. The pseudo image generation unit 402 uses the group of parameters generated by the environmental parameter generation unit 401 to generate a group of pseudo images that serve as candidates for learning data to be used in relearning the AI ​​model. The annotation data generation unit 403 uses the group of parameters generated by the environmental parameter generation unit 401 to generate a group of annotation data that represents the correct answer to the AI ​​processing result for the group of pseudo images. The pseudo learning data evaluation unit 404 evaluates the group of pseudo images by comparing the results of processing the group of pseudo images generated by the pseudo image generation unit 402 using an existing AI model 211 in the same manner as the AI ​​processing unit 202 with the group of annotation data generated by the annotation data generation unit 403, and outputs pseudo learning data that is effective for relearning the AI ​​model. The configurations and operations of the environmental parameter generation unit 401, the pseudo image generation unit 402, the annotation data generation unit 403, and the pseudo learning data evaluation unit 404 are described in detail below.

[0027] The environmental parameter generating unit 401 generates a group of parameters that define a virtual 3D space that serves as a basis for the group of pseudo images. The group of parameters generated by the environmental parameter generating unit 401 is composed of a plurality of parameters, including, for example, one or more parameters (such as the position, direction, angle of view, and image size of the camera) related to a camera installed in the virtual 3D space to generate pseudo images by photographing the virtual 3D space, one or more parameters related to 3D objects to be placed in the virtual 3D space (such as the number of 3D objects and the attributes, position, direction, and size of each 3D object), one or more parameters (such as the type, position, and intensity of the light source) related to light sources (such as lighting) present in the virtual 3D space, one or more parameters related to the background of the virtual 3D space, and one or more parameters (such as the type of occlusion and the degree to which the 3D object is hidden by the occlusion) related to occlusions (such as trees, buildings, cargo, and vehicles) that cause occlusions to the 3D objects.

[0028] The environmental parameter generation unit 401 generates a parameter group by fixing the values ​​of some parameters while varying the values ​​of other parameters. The fixed parameters are selected, for example, to suit the usage conditions (such as the installation environment and season) of the camera that captures the image to be subjected to AI processing in real space. At this time, the usage conditions of the camera may be received from the user 103, and the fixed parameters may be automatically selected according to the content of the usage conditions, or the fixed parameters themselves may be received from the user 103. There may be one or more fixed parameters. Furthermore, multiple parameter groups may be generated while changing the fixed parameters. Furthermore, multiple parameters selected randomly may be set without fixed parameters, in which case one parameter can be regarded as one parameter group.

[0029] The pseudo image generation unit 402 arranges a camera, 3D objects, light sources, background, occlusion, and the like in a virtual 3D space based on the parameter group generated by the environmental parameter generation unit 401, and generates a group of pseudo images by 3D rendering from the camera in the virtual 3D space. The 3D rendering method may be ray tracing, radiosity, scanline, Z-buffer method, toon rendering, or any other 3D rendering method. The 3D rendering method may be changed each time depending on the parameters used.

[0030] The annotation data generation unit 403 generates a group of annotation data suitable for the processing task of the AI ​​processing unit 202 for the group of pseudo images generated by the pseudo image generation unit 402, based on the group of parameters generated by the environmental parameter generation unit 401. As an example, if the processing task of the AI ​​processing unit 202 is object recognition, each piece of annotation data is composed of the class name of an object (e.g., person, dog, cat, car, train, building, park, sea, etc.) appearing in each pseudo image. As another example, if the processing task of the AI ​​processing unit 202 is object detection, each piece of annotation data is composed of the class name of the object to be detected appearing in each pseudo image and bounding box information for that object.

[0031] The pseudo training data evaluation unit 404 performs processing as shown in the flowchart in FIG. 5 . That is, the pseudo training data evaluation unit 404 first performs AI processing similar to that performed by the AI ​​processing unit 202 using the existing AI model 211 on the group of pseudo images generated by the pseudo image generation unit 402 (step S101). Next, the pseudo training data evaluation unit 404 compares the results of the AI ​​processing with the group of annotation data generated by the annotation data generation unit 403 to calculate an evaluation value (step S102). As described above, the evaluation value is an index such as the AI ​​accuracy rate, precision, recall, F-measure, IoU, MOTA, or SSIM, and is expressed, for example, as a value ranging from 0 to 1.

[0032] Next, the pseudo training data evaluation unit 404 compares the calculated evaluation value with a preset threshold T2 (step S103). If the evaluation value is below the threshold T2, this indicates that the existing AI model 211 has low accuracy in AI processing of the target pseudo image group, and the target pseudo image group and annotation data group are effective for relearning the AI ​​model. Therefore, if the evaluation value is below the threshold T2, the pseudo training data evaluation unit 404 combines the target pseudo image group and annotation data group to generate pseudo training data (step S104). Note that the parameters fixed by the environmental parameter generation unit 401 when the pseudo training data was generated are considered to be effective for relearning the AI ​​model. Therefore, this parameter information may be presented to the user 103 to provide material for the user 103 to accumulate knowledge.

[0033] On the other hand, if the evaluation value is equal to or greater than the threshold T2, this indicates that the accuracy of AI processing for the target pseudo image group in the existing AI model 211 is sufficiently high, and it can be said that the target pseudo image group and annotation data group are not effective for relearning the AI ​​model. Therefore, if the evaluation value is equal to or greater than the threshold T2, the pseudo training data evaluation unit 404 discards the target pseudo image group and annotation data group (step S105), and repeats the process of generating another parameter group using the environmental parameter generation unit 401 until an evaluation value below the threshold T2 is obtained.

[0034] As an example, a processing flow for generating pseudo-learning data effective for relearning an AI model used for person detection will be described. The environmental parameter generation unit 401 fixes relevant parameters so that a camera in a virtual 3D space is placed directly above a person object, and generates a parameter group in which parameters such as the gender, clothing, background, and distance between the person object and the camera are varied. The pseudo image generation unit 402 places a camera, a person object, a background, etc. in the virtual 3D space based on the parameter group generated by the environmental parameter generation unit 401, and generates a pseudo image group by performing 3D rendering from the camera in the virtual 3D space. The annotation data generation unit 403 generates an annotation data group having correct bounding box information for the person object in the pseudo image group based on the parameter group generated by the environmental parameter generation unit 401. The pseudo learning data evaluation unit 404 compares the bounding box information resulting from AI processing of the pseudo image group generated by the pseudo image generation unit 402 using the existing AI model 211 with the annotation data group having correct bounding box information generated by the annotation data generation unit 403 to calculate IoU (Intersection over Union) as an evaluation value, and if the average IoU value is below a threshold T2 (e.g., 0.3), combines this pseudo image group with the annotation data group to generate pseudo learning data. As a result, if it is found that the performance of the AI ​​model for person detection for images of people taken from directly above is poor, by presenting this information to the user 103, the user 103 can gain knowledge that such parameters are effective for relearning.

[0035] The retraining unit 303 retrains the AI ​​model using the pseudo training data generated by the pseudo training data generation unit 302. That is, the retraining unit 303 retrains the AI ​​model using a combination of a group of pseudo images and a group of annotation data determined to be effective for retraining. The retraining of the AI ​​model may be performed by retraining from scratch using random initial parameters, or by fine-tuning using parameters of the existing AI model 211. At this time, some or all of the training data used when generating the existing AI model 211 may be additionally used. Furthermore, the new retrained AI model may have the same structure as the existing AI model 211 or may have a different structure, but the AI ​​processing function (i.e., the format of the AI ​​processing result output from the AI ​​processing unit 202) is the same.

[0036] The update unit 304 updates the AI ​​model 211 used in the AI ​​processing in the AI ​​processing unit 202 by replacing it with the AI ​​model re-learned by the relearning unit 303. At this time, the update unit 304 stores and manages the existing AI model 211 in the AI ​​model database 212 so that it can be reused later. As shown in FIG. 6 , the AI ​​model database 212 is composed of items such as an item number identifying each AI model, the generation date and time and storage date and time of each AI model, and information on parameters fixed by the environmental parameter generation unit 401, but other items may be added. When updating the AI ​​model 211 by operating only the update unit 304 in response to an instruction from the user 103, the AI ​​model 211 used in the AI ​​processing in the AI ​​processing unit 202 is replaced with an AI model selected by the user 103 from among the many AI models stored in the AI ​​model database 212. By appropriately updating the AI ​​model 211 in accordance with changes in the domain (season, weather, time of day, camera angle of view, camera installation location, etc.), it is possible to expect to maintain a high-performance image processing system.

[0037] As described above, in the image processing system of this example, the environmental parameter generation unit 401 generates a group of parameters defining a virtual 3D space, the pseudo image generation unit 402 and the annotation data generation unit 403 generate a group of pseudo images and a group of annotation data based on the group of parameters, the pseudo learning data evaluation unit 404 performs AI processing on the group of pseudo images using the AI ​​model 211, and compares the results of the AI ​​processing with the group of annotation data to determine whether the group of pseudo images and the group of annotation data are effective for relearning the AI ​​model 211. The re-learning unit 303 then re-learns the AI ​​model 211 using the group of pseudo images and the group of annotation data determined to be effective for relearning the AI ​​model 211. This configuration allows for efficient collection of learning data effective for relearning the AI ​​model 211, enabling relearning of the AI ​​model 211. Furthermore, the work required by the user 103 during relearning can be minimized, reducing the burden on the user 103.

[0038] Furthermore, in the image processing system of this example, the environmental parameter generation unit 401 repeatedly generates parameter groups while changing the value of at least one parameter in the parameter groups until it finds a pseudo image group and an annotation data group that are determined to be effective for relearning the AI ​​model 211. This makes it possible to find a parameter group that is effective for relearning the AI ​​model 211 and generate a pseudo image group and an annotation data group by repeatedly generating pseudo image groups and annotation data groups under various parameter conditions.

[0039] Furthermore, in the image processing system of this example, the environmental parameter generation unit 401 fixes the value of at least one parameter in the parameter group, but the fixed parameter may be set to suit the usage conditions of the video acquisition device 101 (such as an industrial camera or a surveillance camera) that acquires the image to be processed by AI. This makes it possible to re-train an AI model adapted to individual usage conditions, such as domain changes or performance degradation of the video acquisition device 101.

[0040] In the above description, the re-learning unit 303 re-learns the AI ​​model 211 using a group of pseudo images and a group of annotation data that have been determined to be effective for re-learning the AI ​​model 211, but it may also use a portion of a group of pseudo images and a group of annotation data that have not been determined to be effective for re-learning the AI ​​model 211 for re-learning the AI ​​model 211. This makes it possible to prevent the AI ​​model 211 from being over-learned.

[0041] Although the embodiments of the present invention have been described above, these embodiments are merely illustrative and do not limit the technical scope of the present invention. The present invention can take on various other embodiments, and various modifications such as omissions and substitutions can be made without departing from the spirit of the present invention. These embodiments and modifications thereof are included in the scope and spirit of the invention described in this specification, etc., and are included in the invention described in the claims and their equivalents.

[0042] Furthermore, the present invention can be provided not only as devices such as those described above or as systems composed of these devices, but also as methods executed by these devices, programs for realizing the functions of these devices using a processor, and storage media for storing such programs in a computer-readable manner.

[0043] The present invention can be used in an AI model re-learning system that re-learns AI models used in image processing, and can be applied to a wide range of fields such as public infrastructure.

[0044] 101: Video acquisition device, 102: Image processing device, 103: User, 104: Display output device, 201: Still image acquisition unit, 202: AI processing unit, 203: AI update unit, 204: Information transmission unit, 211: AI model, 212: AI model database, 301: AI update determination unit, 302: Pseudo learning data generation unit, 303: Re-learning unit, 304: Update unit, 401: Environmental parameter generation unit, 402: Pseudo image generation unit, 403: Annotation data generation unit, 404: Pseudo learning data evaluation unit

Claims

1. In an AI model retraining system that retrains an AI model used for image processing, a data generation unit that generates a group of images and a group of annotation data that are candidates for training data used for retraining the AI model based on a group of parameters that define a virtual 3D space, an image processing is performed on the group of images by the AI model, and by comparing the result of the image processing with the group of annotation data, a data evaluation unit that determines whether the group of images and the group of annotation data are effective for retraining the AI model, and a retraining unit that retrains the AI model using the group of images and the group of annotation data determined to be effective for retraining the AI model. An AI model retraining system characterized by comprising:

2. In the AI model retraining system according to claim 1, as the parameters, one or more parameters related to a camera that photographs the virtual 3D space, one or more parameters related to 3D objects arranged in the virtual 3D space, one or more parameters related to a light source existing in the virtual 3D space, one or more parameters related to the background of the virtual 3D space, and one or more parameters related to occlusion that generates occlusion for the target. An AI model retraining system further comprising a parameter generation unit that generates a plurality of parameters, wherein the parameter generation unit repeats the generation of the parameter group while changing the value of at least one parameter in the parameter group until the group of images and the group of annotation data determined to be effective for retraining the AI model are found.

3. In the AI model retraining system according to claim 2, the parameter generation unit fixes the value of at least one parameter in the parameter group so as to match the usage situation of a camera that photographs an image that is the object of image processing by the AI model, and repeats the generation of the parameter group. An AI model retraining system characterized by this.

4. In the AI model retraining system according to claim 1, the retraining unit uses a part of the group of images and the group of annotation data that are not determined to be effective for retraining the AI model for retraining the AI model. An AI model retraining system characterized by this.

Images