Method for generating realistic features and a module generation system using the same

Abstract

The present invention provides a method for generating realistic features, which includes a step A, a step B, a step C and a step D. In the step A, a generative AI module generates a feature type set according to a plurality of images corresponding to samples to be identified. In the step A, a generative AI module generates a feature type set according to a plurality of images corresponding to samples to be identified. In step B, the generative AI module generates string information based on the user's response to the feature type set, and the generative AI module analyzes and segment the string information through a natural language database, and generates and sends an analysis result, and the generative AI module generates response result set for the feature type based on the analysis results. in the step C, the generative AI module is used to learn the feature type question response result set to generate a determining feature type set for determining whether the feature type is a realistic feature; in the In step D, the generative AI module is used to learn the set of determining feature types to generate a set of images with realistic features. The present invention provides a module generation system using the method for generating realistic features.

Description

[Technical Field]

[0001] The present invention relates to a method for generating realistic features and a module generation system using the method for generating realistic features, particularly to a system in the industrial field that can generate modules using the method for generating realistic features in real time according to the needs of users or customers. [Previous Technology]

[0002] In the prior art, image recognition in various industrial fields typically learns from images of defects, such as defective products, defective assembly processes, or continuous images with specific feature combinations, to effectively identify defective or specific products in defect detection. However, image recognition usually requires a large number of defect images to learn from, and in practice, it is often impossible to provide a sufficient number of defect images with the required feature combinations, resulting in poor defect detection performance. On the other hand, because the number of defects is inherently small, belonging to the statistical domain without a quantifiable population, it is not easy for users to directly provide diverse and large numbers of defect images. Furthermore, the defects to be identified often require significant human intervention, such as screening, judgment, and elimination steps, and the establishment of diverse decision-making processes to generate a robust judgment mechanism, which relies heavily on human experience to influence the quality of the generated images. However, there are often not enough human resources to perform the aforementioned screening, judgment, and elimination steps.

[0003] Furthermore, in various current industrial sectors, to analyze the integrated operation of industrial machines with different structures, the operation of vehicles and various voice and image recognition systems, or to analyze data captured by various sensing devices, it is often necessary to write customized modules for the data content of each industrial project. However, due to the significant differences and specializations in project situations across various industrial sectors, even for the same project's data content, as users or customers use it for longer periods, additional needs for modifying existing modules will continuously arise. In conventional technologies, addressing these modification needs often requires a significant amount of rework time, requiring system engineers, front-end engineers, database engineers, back-end engineers, and image recognition engineers to modify the entire physical or virtual module. This approach often consumes a large amount of time, manpower, and cost. Moreover, this need for modifications by various engineers recurs periodically, potentially leading to even greater time, manpower, and cost.

[0004] Furthermore, while conventional automated generation tools allow users to fine-tune the input data and computational architecture, their logical architecture and the required function sets are usually fixed. When the technical problem to be solved changes or is adjusted significantly, it is still inevitable that system engineers, front-end engineers, database engineers, back-end engineers, and image recognition engineers will need to make substantial modifications to the entire automated generation module.

[0005] Therefore, in order to overcome the aforementioned problems, the present invention was developed. [Summary of the Invention]

[0006] The main objective of this invention is to provide a method and system for generating image sets with realistic characteristics. This method can achieve good image detection results even with small sample data and allows users to adjust the data according to their needs. It generates the image judgment module, front-end module, and back-end module that the user actually needs in real time, thereby overcoming the technical problem that a large amount of time and cost is required to make major modifications to the entire module to meet subsequent changes in needs. Furthermore, by statistically transmitting the generation results through a generative AI module, the image judgment module, front-end module, and back-end module that the user actually needs can be generated in real time.

[0007] To achieve the aforementioned objective, the present invention provides a method for generating realistic features, comprising steps A, B, C, and D. In step A, a generative AI module generates a feature type set based on multiple images corresponding to the sample to be identified; in step B, the generative AI module analyzes and segments the string information generated by the user's response to the feature type set using a natural language database, generates and sends a response analysis result, and generates a feature type response result set based on the response analysis result; in step C, the generative AI module learns the feature type question response result set to generate a judgment feature type set for determining whether a feature type is realistic; in step D, the generative AI module learns the judgment feature type set to generate an image set with realistic features.

[0008] In practice, step A further includes: generating the feature type set by the generative AI module based on at least one of the feature types in the plurality of images corresponding to the sample to be identified; wherein the plurality of images corresponding to the sample to be identified includes a first image source consisting of at least one real image and a second image source from an image database.

[0009] In practice, step B further includes: using the generative AI module to perform content parsing based on the response analysis results, and learning the corresponding feature types from the feature type database for the feature types, thereby generating a set of response results for the feature type.

[0010] In practice, step C further includes: learning the set of response results for the feature type of question using the generative AI module to generate a set of judgment features, and generating the set of judgment feature types using the generative AI module based on the set of judgment features.

[0011] In practice, step D further includes: learning at least one of the combinations of multiple judgment feature types in the feature type set using the generative AI module, and performing appearance change image processing on the multiple images corresponding to the sample to be identified, either those without flaws or those with partial flaws, to generate the image set with realistic features.

[0012] In practice, step D further includes: using the generative AI module to learn at least one of the combinations of multiple judgment feature types in the judgment feature type set, and defecting the flawless or partially flawed images corresponding to the samples to be identified, thereby generating the image set with realistic features.

[0013] In practice, step B further includes: using the generative AI module to analyze and segment the string information generated by the user's response to the feature type set, and using the generative AI module to generate and send the response analysis result through the natural language database, and the generative AI module assigns values ​​to the user's response to the combination of each feature type in the feature type combination set based on the response analysis result, thereby generating a feature type response result set.

[0014] The present invention further provides a module generation system using a method for generating realistic features, comprising: an input module, a generative AI module, an image judgment module, a function group generation unit, a background module generation unit, and a front-end module generation unit. The input module receives the user's response to the image corresponding to the analysis result desired by the desired module, converts the response into a string for tag classification, and generates string information. The generative AI module is connected to the input module and generates a feature type set based on multiple samples to be identified, wherein the multiple samples to be identified are images corresponding to the analysis result desired by the desired module. The generative AI module analyzes and segments the response using a natural language database based on the string information generated by the user's response to the feature type set, generates and sends the response analysis result, and generates a feature type response result set based on the response analysis result. The image judgment module is connected to the input module and executes the method described in the foregoing embodiments based on the multiple samples to be identified. The generative AI module learns from the set of realistic features from the multiple samples to be identified, which is required to generate the analysis result corresponding to the desired module. The function group generation unit is connected to the image judgment module and generates a first function group based on the optimal step of the generative AI module for the analysis result to be generated by the desired module. It also generates a second function group based on the optimal step of the generative AI module for displaying the part of the desired module that the user needs to view on the display interface. The background module generation unit is connected to the function group generation unit and generates the background module through the first function group and the image judgment module. The foreground module generation unit is connected to the function group generation unit and generates the foreground module through the second function group and the image judgment module.

[0015] In practice, the display interface is connected to the function group generating unit, and the display interface displays multiple samples to be identified that have been confirmed by the user.

[0016] In implementation, the aforementioned system further includes a database generation unit, which generates the image database, the natural language database, or the feature type database based on the image corresponding to the analysis result to be generated by the module to be generated.

[0017] In implementation, the background module generating unit is further connected to the image database, the natural language database and the feature type database and performs the following steps, including: filtering the original data from the image database, the natural language database and the feature type database to generate filtered data; performing calculations on the filtered data to generate calculated data; and performing statistical analysis on the calculated data to generate the analysis result.

[0018] In implementation, the aforementioned system further includes a setting unit, which is connected to the background module generation unit and the front-end module generation unit, and is used to set the data type of the required input data in the background module generation unit and the front-end module generation unit, the background logic architecture corresponding to the background module generation unit, or the front-end logic architecture corresponding to the front-end module generation unit according to the multiple steps executed corresponding to the module to be generated; wherein the front-end logic architecture is set according to the content viewed by the user on the display interface and the content required to be input by the input module, and the second function group is set accordingly; wherein the background logic architecture is set according to the user's analysis needs, and the first function group is set accordingly.

[0019] In implementation, the generative AI module is connected to the setting unit, which provides the setting unit to set the back-end logic architecture corresponding to the back-end module generation unit or the front-end logic architecture corresponding to the front-end module generation unit according to the multiple steps performed corresponding to the module to be generated.

[0020] In implementation, the generative AI module is connected to the backend module generation unit to provide the first function set of the backend module generation unit, and the generative AI module is connected to the frontend module generation unit to provide the second function set of the frontend module generation unit.

[0021] To further understand the present invention, the following preferred embodiments are described in detail with reference to drawings and reference numbers, illustrating the specific composition of the present invention and the effects it achieves.

Implementation Method

[0022] Referring to Figure 1, the present invention provides a module generation system using a method for generating an image set with realistic characteristics, comprising: an input module 1, an image judgment module 2, a function group generation unit 3, a background module generation unit 4, a front-end module generation unit 5, a setting unit 6, and a generative AI module 7. The input module 1 is used to receive the user's response to the image corresponding to the analysis result to be generated by the module to be generated, and converts the response into a string for tag classification processing and generates string information. The generative AI module 7 is connected to the input module 1 and generates a feature type set based on multiple samples to be identified, wherein the multiple samples to be identified are images corresponding to the analysis results that the module wants to generate. The generative AI module 7 generates and sends response analysis results based on the string information generated by the user's response to the feature type set, and analyzes and segments the response through the natural language database 8. The generative AI module 7 generates a feature type response result set based on the response analysis results. The image judgment module 2 is connected to the input module 1 and performs the method as described in the foregoing and subsequent embodiments to generate a set of realistic features from the multiple samples to be identified, which is required for the analysis results that the module wants to generate. The generative AI module 7 learns the set of images with realistic features. The function group generation unit 3 is connected to the image judgment module 2 and generates a first function group based on the optimal step corresponding to the analysis result to be generated by the generative AI module 7 for the module to be generated. It also generates a second function group based on the optimal step required by the generative AI module to display the portion of the module to be generated that the user needs to view on the display interface. The background module generation unit 4 is connected to the function group generation unit 3 and generates a background module via the first function group and the image judgment module 2. The foreground module generation unit 5 is connected to the function group generation unit and generates a foreground module via the second function group and the image judgment module 2.

[0023] Referring to Figure 2A, the present invention provides a method for generating an image set with realistic features, comprising the following steps A to D: Step A: A generative AI module generates a feature type set based on multiple images corresponding to samples to be identified; Step B: The generative AI module analyzes and segments the string information generated by the user's response to the feature type set using a natural language database, generates and sends a response analysis result, and generates a feature type response result set based on the response analysis result; Step C: The generative AI module learns the feature type question response result set to generate a judgment feature type set for judging whether the feature type is realistic; and Step D: The generative AI module learns the judgment feature type set to generate an image set with realistic features.

[0024] In another embodiment, the aforementioned system further includes a display interface (not shown), which is connected to the function group generation unit 3, and the display interface displays a plurality of samples to be identified by the user.

[0025] In another embodiment, the aforementioned system further includes a database generation unit (not shown), which is connected to the generative AI module 7 and generates the natural language database 8, the image database 9, or the feature type database 10 based on the image corresponding to the analysis result to be generated by the module to be generated. That is, the generative AI module 7 generates string information based on the user's response to the structure of the generated database, and analyzes and segments the string information through the natural language database to generate and send the response analysis result. The generative AI module 7 generates a set of response results for the generated database based on the response analysis result. In the aforementioned embodiment, the input module 1 is further configured to receive the user's response to the content of the generated databases, convert the response into a string for tag classification processing, and generate the aforementioned string information.

[0026] In another embodiment, the background module generation unit 4 is further connected to the image database 9, the natural language database 8, and the feature type database 10 and performs the following steps, including: filtering the original data from the image database 9, the natural language database 8, and the feature type database 10 to generate filtered data; performing calculations on the filtered data to generate calculated data; and performing statistical analysis on the calculated data to generate the analysis result. For example, referring to the embodiments of Figures 3A to 7A of the present invention, in order to filter the time information contained in the original image to generate filtered data, the filtering condition is the data from 4 pm to 6 pm on July 10, 2024; then, the average value of each of the multiple data in the aforementioned steps is taken; then, the calculated data is statistically analyzed, for example, by ANOVA-test, to generate an analysis result on whether there is a significant difference in the operation time of the aforementioned steps.

[0027] In another embodiment, the aforementioned system further includes a setting unit 6, which is connected to the background module generation unit 4 and the front-end module generation unit 5. The setting unit 6 is used to set the data type of the required input data in the background module generation unit 4 and the front-end module generation unit 5, the background logic architecture corresponding to the background module generation unit, or the front-end logic architecture corresponding to the front-end module generation unit, according to the multiple steps executed corresponding to the module to be generated. The data type includes: string type, integer type, floating-point type, character type, Boolean type, etc., depending on the data analysis requirements, and will not be elaborated here. The front-end logic architecture is set according to the content viewed by the user on the display interface and the content required for input by the input module, based on the background logic architecture; wherein the background logic architecture is set according to the user's analysis needs, based on the first function group. In another embodiment, the background logic architecture corresponding to the background module generation unit 4 is to perform calculations on the aforementioned function group and output the required logic architecture according to the user's analysis needs; the setting unit 6 sets the front-end logic architecture according to the user's needs and the content that the user needs to view on the display interface corresponding to the background logic architecture, based on the front-end logic architecture corresponding to the front-end module generation unit 5.

[0028] In one embodiment, the generative AI module 7 of the present invention generates the aforementioned front-end and back-end logical architecture, code, images, natural language, and feature types. The present invention extensively utilizes large-scale language models such as the GPT model. The GPT model performs adversarial and reinforcement training on the pre-learned function sets of the back-end module, the logical architecture of this back-end module (e.g., interface connections based on API communication protocols), images with flawed features, images with realistic features, and statements used by users in the question-and-answer process. Therefore, for subsequent embodiments, the corresponding... The first set of functions is generated by the optimal step corresponding to the analysis result to be generated by the module. The second set of functions is generated by the generative AI module based on the optimal step required to display the portion of the module to be generated that the user needs to view on the display interface during the multiple data processing steps. The realistic image generation module generation unit (not shown) generates a set of realistic features from the multiple samples to be identified, corresponding to the analysis result to be generated by the module, based on the generative AI module, to generate an image with realistic image features. Therefore, the various modules generated by the method and system of the present invention can generate realistic defect images for a specific project according to the user's needs, thereby generating a module with good image detection performance even with small sample data by producing a sufficient amount of images with realistic image features.

[0029] In another embodiment, the generative AI module 7 is connected to the setting unit 6 so that the setting unit 6 can set the background logic architecture corresponding to the background module generation unit 4 or the front-end logic architecture corresponding to the front-end module generation unit 5 according to the multiple steps performed corresponding to the module to be generated.

[0030] In another embodiment, the generative AI module 7 is connected to the background module generation unit 4 to provide the first function set of the background module generation unit 4, and the generative AI module 7 is connected to the front-end module generation unit 5 to provide the second function set of the front-end module generation unit 5.

[0031] In another embodiment, the background module generation unit 4 is connected to the generative AI module 7. The generative AI module 7 generates and sends response analysis results based on the string information generated by the user's response to the generated background module, and analyzes and segments the string information through a natural language database. The generative AI module 7 then generates a set of response results for the generated background module based on the response analysis results. In this embodiment, the system of the present invention can make the generative AI module 7 act as an AI backend engineer through the display interface. The user can adjust the code of the generated module and the corresponding backend logic architecture by interacting with the AI ​​backend engineer through question and answer. Similarly, in another embodiment, the front-end module generation unit 5 is connected to the generative AI module 7. The generative AI module 7 generates and sends response analysis results based on the string information generated by the user's response to the generated front-end module, and analyzes and segments the string information using a natural language database. The generative AI module 7 then generates a set of response results for the generated back-end module based on these response analysis results. In this embodiment, the system of the present invention can make the generative AI module 7 act as an AI front-end engineer through the display interface. The user interacts with the AI ​​front-end engineer through question-and-answer sessions, allowing the generative AI module 7 to adjust the generated code and the corresponding front-end logic architecture. In the aforementioned embodiment, the input module 1 is further configured to receive the user's response to the generated back-end module and the generated front-end module, convert the response into a string for tag classification processing, and generate the aforementioned string information.

[0032] The following details the method and system of the present invention. Please refer to the embodiments shown in Figures 3A to 7A of the present invention. These embodiments correspond to the analysis results to be generated for the module to be produced, that is, analyzing the performance of this project and generating virtual defect images for each step of this project. The corresponding images are multiple screenshots of the automated assembly process of the circuit board. First, the correct sequence and defect-free images are described below. In Figure 3A, the automated mechanical device (not shown) first tightens the screw (object) in the dashed box of step 1 of this process; then, in Figure 4A, the automated mechanical device installs the ribbon cable (object) of step 2 of this process in the position of the dashed box; then, in Figure 5A, the automated mechanical device installs another set of ribbon cables (objects) of step 3 of this process in the position of the dashed box; then, in Figure 6A, the automated mechanical device installs another set of ribbon cables of step 4 of this process in the position of the dashed box; finally, in Figure 7A, the automated mechanical device first tightens the other screw in the dashed box of step 5 of this process.

[0033] Please refer to the embodiments shown in Figures 8A to 12A of the present invention. This embodiment is the correct process for an operator to assemble a circuit board. First, in Figure 8A, the operator tightens the screws in the dashed box; then, in Figure 9A, the operator installs the ribbon cable in the position of the dashed box; then, in Figure 10A, the operator installs another set of ribbon cables in the position of the dashed box; then, in Figure 11A, the operator installs another set of ribbon cables in the position of the dashed box; finally, in Figure 12A, the operator tightens the other screw in the dashed box. Figures 8A to 12A only show the operator's hands, but the method and system of the present invention are not limited to recognizing and using only the image of the operator's hands, but can also recognize and use the operator's overall actions.

[0034] The method and system of the present invention will be described in detail below. In step A, the generative AI module generates a set of feature types based on multiple images corresponding to the sample to be identified, i.e., for example, screenshots of the automated assembly process of the aforementioned circuit board (Figures 3A to 12A) and screenshots of the operator assembling the circuit board (Figures 3B to 12B). For example, the location of screw installation in each step, the location of cable installation in each step, whether the operator's hand is near the object being installed in each step, the state of screw tightening / untightening, the state of screw installation being correct / failed, the state of cable installation being correct / failed, etc.

[0035] Furthermore, in step B, referring to Figure 2B, the generative AI module 7 generates string information based on the user's response to the feature type set via the input module 1. For example, if the feature types generated by the generative AI module 7 are not accurate enough, the positions of the summarized feature types need to be adjusted, or even if the feature types are incorrect, the user can respond via the input module 1. For example, the user can use the input module 1 to tell the generative AI module 7 how the positions should be adjusted to be more accurate, what additional features must be included, or which of the aforementioned images corresponds to what feature type, etc. Subsequently, the generative AI module 7 analyzes and segments the string information using the natural language database 8, generates and sends response analysis results, and produces a feature type response result set based on these results. In other words, the generative AI module 7, through the user's response, ensures that the exported feature type response result set accurately reflects various actual features, such as the correct position range of screw installation in each step, the correct position range of ribbon cable installation in each step, the correct position range of the operator's hand (whether it is near the object being installed in each step and will not be judged as a defect / incorrect image), the correct position range of screw tightening / loosening, the correct position range of screw installation correct / failed, and the correct position range of ribbon cable installation correct / failed, etc. The input module 1 is further configured to receive the user's response to the content of the generated image, convert the response into a string for tag classification processing, and generate the aforementioned string information. Therefore, in this embodiment, the system of the present invention can enable the generative AI module 7 to act as an image engineer through the display interface. The user interacts with the AI ​​image engineer through question-and-answer dialogue, allowing the generative AI module 7 and the image judgment module 2 to work together to adjust the features contained in the generated image. Furthermore, in step C, the generative AI module 7 learns the set of response results for the feature type questions, generating a set of judgment feature types for determining whether a feature type is realistic. For example, multiple features for determining the location of screw installation in each step, multiple features for determining the location of cable installation in each step, multiple features for determining whether the operator's hand is near the object being installed in each step, multiple features for determining whether the screw is tightened / untightened, multiple features for determining whether the screw is installed correctly / failed, and multiple features for determining whether the cable is installed correctly / failed, etc. The response result generated from the screenshots of the automated assembly process of the aforementioned circuit board (Figures 3A to 12A) and the screenshots of the operator assembling the circuit board (Figures 3B to 12B) yields a set of judgment feature types for determining whether a feature type is realistic. Each element in this set of judgment feature types is used to determine whether the aforementioned images are realistic.In other words, the method and system of the present invention can utilize small sample data, such as less than 500, 400, 300, 200, 100, or 50 samples, to generate multiple images that conform to realism, thereby enabling generative AI modules to learn and generate images with realistic features and diverse flaw features.

[0036] Furthermore, in step D, the generative AI module 7 learns each of the judgment feature type set, and generates an image set with realistic features through the image database 10 and the feature type database 10. As mentioned above, since image recognition usually requires a large number of defective images for the generative AI module 7 to learn from, in practice, in many industrial technical fields, it is often impossible to provide a sufficient number of defective images with the required feature combinations, resulting in poor defect detection performance. Therefore, the image set with realistic features generated in step D of the present invention is a diverse, large number of defective images that conform to the real situation.

[0037] In another embodiment, step A further includes: generating the feature type set by the generative AI module 7 based on at least one of the feature types in the plurality of images corresponding to the sample to be identified; wherein the plurality of images corresponding to the sample to be identified includes a first image source consisting of at least one real image and a second image source from the image database 9, meaning that the image source of the present invention may be not only real images, but also virtual images learned by the generative AI module 7. Furthermore, the method of the present invention can generate the feature type set by taking at least one of the feature types in the images corresponding to the sample to be identified. That is, the present invention can adjust the number of at least one of the following according to actual needs, such as the training needs of the generative AI module: the images of the screw installation location, the images of the cable installation location, the images of the operator's hand position, the images of screw tightening / untightening features, the images of screw installation correct / failed installation features, and the images of cable installation correct / failed installation features, thereby adjusting the training direction of the generative AI module 7 and maximizing the system performance of the present invention.

[0038] In another embodiment, please also refer to FIG2B. Step B further includes: using the generative AI module 7 to perform content parsing based on the response analysis results, and learning the corresponding feature types from the feature type database 10 for the feature types, thereby generating a set of response results for the feature type. The generative AI module 7 compares the resulting set of feature type response results generated by the user's response with the feature type database 10, which contains feature types established before or before this step. Specifically, it compares the following adjustments in each step of the response analysis results: fine-tuning of the screw installation location, fine-tuning of the cable installation location, fine-tuning of whether the operator's hand is near the object being installed, fine-tuning of the screw tightening / loosening feature, fine-tuning of the screw installation correct / failed feature, and fine-tuning of the cable installation correct / failed feature, with the feature type database 10, which contains feature types established before or before this step, to generate a set of feature type response results that are clearly defined based on the user's feedback.

[0039] In another embodiment, step C further includes: learning the set of response results for the feature type of question using the generative AI module 7 to generate a set of judgment features, and generating features that can truly determine whether the aforementioned situation is a real defect based on the set of judgment features, so as to avoid generating defect images that do not conform to the real situation.

[0040] In another embodiment, step D further includes: learning at least one of the combinations of multiple judgment feature types in the feature type set using the generative AI module 7, and performing appearance change image processing on the multiple images corresponding to the samples to be identified, either flawless or partially flawed, to generate the image set with realistic features. The image processing includes, for example, at least one of random gamma, scaling, fisheye, blur, flip, normalization, transpose, crop, optical distortion, hue saturation value, mesh distortion, elastic transformation, random mesh shuffling, filling, translation, random brightness, random contrast, motion blur, contrast-limited adaptive histogram, equalized center blur, Gaussian blur, affine transformation, inversion, and Gaussian noise. This means that, in order to optimize and diversify the realistic images that the methods and systems of the present invention can generate, the images that are flawless (i.e., the images of the correct installation action in the aforementioned embodiments) or have partial defects in the multiple images corresponding to the sample to be identified are subjected to appearance modification image processing. For example, the images of Figures 3A to 12A in the aforementioned automated assembly process of the circuit board are subjected to appearance modification image processing; or the screenshots of the assembly process of the circuit board with the operator's hand image in Figures 3B to 12B (with defects) are subjected to appearance modification processing, thereby generating various appearance-modified images with defects that conform to the real situation.

[0041] Furthermore, in another embodiment, step D further includes: using the generative AI module 7 to learn at least one of the combinations of multiple judgment feature types in the judgment feature type set, and defecting the flawless or partially flawed images corresponding to the sample to be identified, thereby generating the image set with realistic features. The method and system of the present invention can, according to actual needs, use the generative AI module 7 to learn at least one of the combinations of multiple judgment feature types in the judgment feature type set, and imperfect the flawless (images in Figures 3A to 12A) or partially flawed (Figures 3B to 12B) images corresponding to the sample to be identified. Imperfectization means that the flawless or partially flawed images corresponding to the sample to be identified are combined according to the needs, for example, the flawed part (dashed line) in the image of 7B and the flawed part (dashed line) in the image of 8B, to generate the image set with realistic features.

[0042] Furthermore, in another embodiment, step B further includes: using the generative AI module 7 to analyze and segment the string information generated by the user's response to the feature type set, and using the generative AI module to analyze and segment the string information through the natural language database, generating and sending the response analysis result, and the generative AI module assigning values ​​to the user's response to the combination of each feature type in the feature type combination set according to the response analysis result. For example, in this embodiment, the surface information from the aforementioned captured image is assigned to the coordinate system established by the production line in this embodiment; or, in another embodiment, different priority weights are given to the aforementioned various feature types. For example, the features used to determine whether the operator's hand position is near the object being installed in each step are given the highest priority (highest value), the features used to determine whether the screw is installed correctly / failed are given the second highest priority (second highest value), the features used to determine whether the cable is installed correctly / failed are given the third highest priority (third highest value), etc., to generate the feature type response result set.

[0043] In another embodiment of the present invention, the present invention further provides a method for generating images with realistic defects, including steps A' to E'. Step A' includes: acquiring a plurality of real images with defects corresponding to samples to be identified; Step B' further includes: generating a defect type question set using the generative AI module based on the real images with defects corresponding to samples to be identified; wherein Step C' further includes: generating a defect type question response result set based on responses from a plurality of users to the defect type question set; Step D' includes: generating a defect feature judgment set using the generative AI module 7 based on the defect type question answer result set; Step E' includes: generating a virtual defect image set using the generative AI module 7 based on the defect feature set; Step F' includes: generating a set of images with realistic defects using the generative AI module based on the virtual defect feature image set through learning. In another embodiment, the system of the present invention further includes a realistic image generation module generating unit, which is connected to the image judgment module 2, and the realistic image generation module generating unit performs the aforementioned steps A' to E'.

[0044] Therefore, the present invention has the following advantages: 1. The method and system of the present invention for generating image sets with realistic features can achieve good image detection results even with small sample data, and allows users to adjust according to their needs, instantly generating the image judgment module, front-end module, and back-end module actually needed by the user. 2. The present invention improves the model learning effect of various image recognition by increasing the number of training images and the diversity of image content. 3. The present invention can generate diverse images with realistic defects through generative AI modules and image databases, feature type databases, and natural language databases connected to them, after training such as adversarial training. 4. The method and system of the present invention use the generated images with realistic defects to further check the automated processes of related industries, and the modules generated by the system of the present invention can significantly improve the error rate of existing automated mechanical assembly equipment in related industries. 5. The method and system of the present invention only need to use images captured by general photography devices such as smartphones to instantly train the captured images, and automatically construct the module by providing the required questions and answers through the dialog box of the AI ​​module, saving a lot of time spent on familiar modules, which is extremely convenient.

[0045] The above description is a specific embodiment of the present invention and the technical means used. Many changes and modifications can be derived from the disclosure or teachings herein, which can still be regarded as equivalent changes to the concept of the present invention. The effects produced do not exceed the essential spirit covered by the specification and drawings, and should all be regarded as within the technical scope of the present invention. This is hereby stated.

[0046] In summary, based on the content disclosed above, the present invention can indeed achieve the intended purpose of the invention, providing a method for generating realistic features and a module generation system using the method for generating realistic features, which has great industrial applicability. Therefore, an invention patent application is filed in accordance with the law. [Simplified Explanation of the Diagram]

[0047] Figure 1 is a block diagram of an embodiment of the module generation system of the present invention using a method for generating an image set with realistic features.

[0048] Figure 2A is a flowchart of an embodiment of the method for generating an image set with realistic features according to the present invention.

[0049] Figure 2B is a schematic diagram of another embodiment of the method for generating an image set with realistic features according to the present invention.

[0050] Figures 3A to 12A are schematic diagrams of the normal process of an embodiment of the method for generating an image set with realistic features according to the present invention.

[0051] Figures 3B to 12B are schematic diagrams of flawed features of another embodiment of the method for generating an image set with realistic features according to the present invention.

Claims

1. A method for generating realistic features, comprising: Step A: A generative AI module generates a feature type set based on multiple images corresponding to the sample to be identified. Step B: The generative AI module analyzes and segments the string information generated by a user's response to the feature type set using a natural language database, generating and sending a response analysis result. Based on this response analysis result, the generative AI module generates a feature type response result set. Step C: The generative AI module learns from the feature type response result set to generate a set of judgment feature types for determining whether a feature type is realistic. Step D: The generative AI module learns at least one combination of multiple judgment feature types from the feature type set and performs appearance change image processing on the flawless or partially flawed images corresponding to the sample to be identified, generating an image set with realistic features.

2. A method for generating realistic features, comprising: Step A: A generative AI module generates a feature type set based on multiple images corresponding to the sample to be identified. Step B: The generative AI module analyzes and segments the string information generated by a user's response to the feature type set using a natural language database, generating and sending a response analysis result. Based on this response analysis result, the generative AI module generates a feature type response result set. Step C: The generative AI module learns from the feature type response result set to generate a judgment feature type set for determining whether a feature type is realistic. Step D: The generative AI module learns at least one combination of multiple judgment feature types from the judgment feature type set, and imperfects or partially imperfect images corresponding to the sample to be identified, generating an image set with realistic features.

3. The method for generating realistic features as described in claim 1 or 2, wherein step A further comprises: The generative AI module generates the feature type set based on at least one of the feature types in the plurality of images corresponding to the sample to be identified; wherein the plurality of images corresponding to the sample to be identified includes a first image source consisting of at least one real image and a second image source from an image database.

4. The method for generating realistic features as described in claim 1 or 2, wherein step B further includes: The generative AI module performs content parsing based on the response analysis results and learns the corresponding feature types from a feature type database to generate a set of response results for that feature type.

5. The method for generating realistic features as described in claim 1 or 2, wherein step C further comprises: The generative AI module learns the set of response results for the feature type of question to generate a set of judgment features, and the generative AI module generates the set of judgment feature types based on the set of judgment features.

6. The method for generating realistic features as described in claim 1 or 2, wherein step B further comprises: The generative AI module generates string information based on a user's response to a set of feature types. It then analyzes and segments the string using a natural language database to generate and send the response analysis results. Based on these results, the generative AI module assigns values ​​to the user's responses to combinations of feature types within the feature type combination set, thus generating a set of feature type response results.

7. A module generation system using a method for generating realistic features, comprising: An input module receives a user's response to an image corresponding to the analysis result to be generated by the desired module, converts the response into a string for tag classification, and generates a string of information. A generative AI module is connected to the input module and generates a feature type set based on multiple samples to be identified, wherein the multiple samples to be identified are images corresponding to the analysis result to be generated by the desired module. The generative AI module analyzes and segments the response using a natural language database based on the string of information generated by the user's response to the feature type set, generates and sends a response analysis result, and generates a feature type response result set based on the response analysis result. An image judgment module, connected to the input module, performs the method described in any one of claims 1 to 6 to generate a set of realistic features from the plurality of samples to be identified, which is required for the analysis result to be generated by the module to be generated, based on the plurality of samples to be identified, and learns the set of images with realistic features by the generative AI module; a function group generation unit, connected to the image judgment module, generates a first function group based on the optimal step of the generative AI module for the analysis result to be generated by the module to be generated, and generates a second function group based on the optimal step of the generative AI module for displaying the part of the module to be generated that the user needs to view in a display interface in the plurality of data processing steps; A background module generation unit is connected to the function group generation unit and generates a background module through the first function group and the image judgment module; and a foreground module generation unit is connected to the function group generation unit and generates a foreground module through the second function group and the image judgment module.

8. A module generation system using a method for generating realistic features as described in claim 7, wherein the display interface is connected to the function group generation unit, and the display interface displays a plurality of samples to be identified as confirmed by the user.

9. The module generation system using the method for generating realistic features as described in claim 7, further comprising a database generation unit that generates an image database, the natural language database, or a feature type database based on the image corresponding to the analysis result to be generated for the module to be generated.

10. A module generation system using the method for generating realistic features as described in claim 9, wherein the background module generation unit is further connected to the image database, the natural language database, and a feature type database and performs the following steps, including: The raw data from the image database, the natural language database, and the feature type database are filtered to generate filtered data; the filtered data is processed to generate processed data; and the processed data is statistically analyzed to generate the analysis results.

11. The module generation system using the method for generating realistic features as described in claim 7, further comprising a setting unit connected to the background module generation unit and the foreground module generation unit, for setting the data type of the required input data in the background module generation unit and the foreground module generation unit, a background logic architecture corresponding to the background module generation unit, or a foreground logic architecture corresponding to the foreground module generation unit, according to the multiple steps performed corresponding to the module to be generated; wherein the foreground logic architecture is set according to the content viewed by the user on the display interface and the content required to be input by the input module, wherein the background logic architecture is set according to the user's analysis needs, wherein each of the functions in the first function group is set.

12. A module generation system using a method for generating realistic features as described in claim 11, wherein the generative AI module is connected to the setting unit to provide the setting unit with setting the back-end logic architecture corresponding to the back-end module generation unit or the front-end logic architecture corresponding to the front-end module generation unit according to multiple steps performed corresponding to the module to be generated.

13. A module generation system using a method for generating realistic features as described in claim 7, wherein the generative AI module is connected to the backend module generation unit to provide the first set of functions of the backend module generation unit, and the generative AI module is connected to the frontend module generation unit to provide the second set of functions of the frontend module generation unit.

Images