Product appearance defect importance degree evaluation method and system

Abstract

The invention provides a method and a system for evaluating the importance degree of appearance flaws of a product, and belongs to the fields of industrial quality inspection, data quality evaluation and the like. The method comprises the steps of obtaining a score marking result through double marking, generating a defect target and evaluating the importance degree, and predicting the importance degree of the product appearance defect through a trained importance degree regression model based on marking score data. According to the method, a relatively consistent continuous result can be given for the severity of the appearance defects of the product, and when subjective changes occur in the standard, the scheme can adapt to the subjective changes only by adjusting the threshold value, so that high workload and time cost caused by re-labeling data are avoided. The problems of subjectivity and reliability in product appearance defect quality evaluation are solved, and the method can be popularized and applied in the field of industrial quality inspection or visual inspection.

Description

Technical Field

[0001] This invention belongs to the fields of industrial quality inspection and data quality assessment, and specifically relates to a method and system for assessing the importance of product appearance defects. Background Technology

[0002] In the field of data quality assessment, traditional product appearance defect severity rating schemes typically use machine learning or deep learning methods to manually divide collected defect images into a finite number of severity level categories. This data is then compiled into a dataset and fed into a model for defect level classification learning. After the model has learned, it outputs a rating result, which includes the level category and a confidence score for that category. However, confidence scores alone cannot accurately express the severity of the defect. Furthermore, machine learning or deep learning methods require high accuracy in dataset labeling, and the data labeling standards themselves have a degree of subjectivity, which significantly reduces the reliability of defect severity detection. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a method and system for evaluating the importance of product appearance defects, which can solve the above-mentioned problems.

[0004] A method for assessing the importance of product appearance defects includes the following steps: obtaining scoring annotation results through double annotation, classifying and comparing product appearance defects to obtain objective scoring annotation results; generating defect targets by generating defect targets through data transformation and providing relative scores based on the transformation operation; and assessing importance by predicting the importance of product appearance defects based on the labeled scoring data and a trained importance regression model.

[0005] Furthermore, the grading system categorizes defect data into three levels of severity: minor, moderate, and severe. Minor defects are those that are not clearly visible in the image or are of good quality. Moderate defects are those that can be distinguished or whose category boundaries are somewhat confusing. Severe defects are those that are clearly identifiable and whose category is clearly defined.

[0006] Furthermore, the comparison labels are used to determine the ranking by comparing which of similar defects is more serious, and a severity score is calculated based on the ranking.

[0007] Furthermore, the formula for calculating the severity score of defects is: Defect Severity Score = Defect Severity Level Score + Comparison Score; where the severity level scores correspond as follows: Slight is 0 points, Average is 0.3 points, and Severe is 0.6 points; Comparison Score = Number of times the comparison result is more severe / Total number of comparisons × Total score of the interval; The total scores of the intervals for Slight, Average, and Severe are 0.3, 0.3, and 0.4, respectively.

[0008] Furthermore, the data transformations during defect target generation include scaling transformation and rotation transformation. Rotation transformation does not change the target score, and scaling transformation does not change the defect level, but the score within the level will be adjusted according to the scaling factor and the following formula.

[0009] Within-level score = min(scaling factor × current within-level score, maximum within-level score); where the within-level score is the defective target score minus the severity score.

[0010] Furthermore, the importance regression model includes a backbone network based on MobileViT, pooling layers, and fully connected layers, and the loss function used for model training is the Huber loss function.

[0011] This invention also provides a product appearance defect importance assessment system, comprising: a labeling module for performing dual labeling of product appearance defects (level labeling and comparison labeling) to obtain objective scoring labeling results; a defect target generation module, which uses a Stable Diffusion model to transform the labeled and scored defect data to generate defect targets and provides relative scores based on specific transformation operations; and a model training module, which uses an importance regression model and trains the importance regression model based on the labeled scoring data to predict the importance of product appearance defects.

[0012] Furthermore, the rating label in the labeling module divides the data into three categories according to severity level: minor, moderate, and severe. The comparison label in the labeling module determines the ranking based on which of the similar defects is more severe, and calculates the severity score based on the ranking.

[0013] Furthermore, data transformation includes scaling and rotation operations. Rotation does not change the target score, and scaling does not change the defect level, but the score within the level will be adjusted according to the scaling factor.

[0014] Furthermore, the importance regression model includes the backbone network, pooling layer, and fully connected layer based on MobileViT.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: the solution of this application can provide a more consistent and continuous result on the severity of product appearance defects. In the standard alignment process, especially for defect standard alignment in complex scenarios, it can improve the alignment efficiency by more than 50%. At the same time, when the standard changes subjectively, the solution only needs to adjust the threshold to adapt, avoiding the high workload and time cost brought about by re-annotating data. It can be promoted and applied in the fields of industrial quality inspection or visual inspection. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the overall scheme for scoring the importance of this invention; Figure 2 This is a schematic diagram of the importance regression model. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] A method for assessing the importance of product appearance defects, see [link to relevant documentation]. Figure 1 The overall flowchart for the importance rating scheme includes the following steps.

[0019] The dual-labeling method obtains the rating labeling results by classifying and comparing product appearance defects, thus obtaining objective rating labeling results.

[0020] Flaw target generation involves generating flaw targets through data transformation and providing a relative score based on the transformation operation.

[0021] Importance assessment, based on labeled scoring data, uses a trained importance regression model to predict the importance of product appearance defects.

[0022] Detailed Explanation: To address the subjectivity and reliability issues in product appearance defect quality assessment, we first provide relatively objective scoring results through grade labeling and comparison labeling. Then, we generate defect targets through various data transformations and provide relative scores based on the specific transformation operations. Finally, based on the scoring data, we use an importance regression model to assess the importance of product appearance defects. By using target importance scoring in actual industrial production, we can both improve the reliability of target detection and guide customers to determine target data boundaries through quantitative scoring.

[0023] The manual scoring of product appearance defects is mainly divided into two modules: one is the grade labeling; the other is the comparison labeling.

[0024] Among them, the grade labeling divides the defect data into three categories according to the severity level: minor, moderate and severe: (1) minor, which is not obvious in imaging or good products; (2) moderate, which can be distinguished by careful identification or the category boundary definition is somewhat confusing; (3) severe, which is obviously identifiable and has a clear category.

[0025] The comparative annotation method uses a comparison of similar defects to determine the ranking, and then calculates a severity score based on the ranking.

[0026] When selecting comparison samples, the flawed samples with the most similar current scores are screened and compared. The calculation method for flawed sample scores, i.e. the formula for calculating the severity score of flaws, is as follows.

[0027] The severity score is calculated as follows: Severity Level Score + Comparison Score. The severity levels correspond as follows: Minor is 0 points, Average is 0.3 points, and Severe is 0.6 points.

[0028] The comparison score is calculated as follows.

[0029] The comparison score = number of times the comparison result was more severe / total number of comparisons × total score of the interval; where the total scores of the intervals for mild, moderate and severe are 0.3, 0.3 and 0.4 respectively.

[0030] The defect target generation process employs a Stable Diffusion-based model to migrate, or transform, the labeled defect targets across different categories. During migration, scaling and rotation transformations are applied to the defects. Rotation transformations do not change the target score, and scaling transformations do not change the defect level; however, the scores within each level are adjusted according to the scaling factor, as detailed below.

[0031] Within-level score = min(scaling factor × current within-level score, maximum within-level score); where the within-level score is the defective target score minus the severity score.

[0032] The structure of the importance regression model proposed in this invention is as follows: Figure 2 As shown, the importance regression model includes a backbone network based on MobileViT, a pooling layer, and a fully connected layer. The loss function adopted is the Huber loss function, which is as follows.

[0033] .

[0034] δ is the decision parameter. The prediction bias is used. Different loss calculation methods are employed based on the relationship between the prediction bias and δ. When the prediction bias is less than δ, Huber Loss uses the squared error; while when the prediction bias is greater than δ, linear error is used.

[0035] The present invention also provides a product appearance defect importance assessment system, the system comprising the following modules.

[0036] The annotation module is used for dual annotation of product appearance defects, including severity rating and comparison annotation, to obtain objective scoring results. The severity rating in the annotation module categorizes data into three levels: minor, moderate, and severe. The comparison annotation determines the ranking based on the severity of similar defects and calculates the severity score based on the ranking.

[0037] The defect target generation module uses the Stable Diffusion model to transform the labeled and scored defect data to generate defect targets, and provides a relative score based on the specific transformation operation. Data transformation includes scaling and rotation operations. Rotation does not change the target score, and scaling does not change the defect level, but the score within each level will be adjusted according to the scaling factor.

[0038] The model training module employs an importance regression model, trained based on labeled scoring data, to predict the importance of product appearance defects. The importance regression model comprises a MobileViT-based backbone network, pooling layers, and fully connected layers.

[0039] The above methods and systems enable a rapid assessment of the importance of product appearance defects, avoiding the high workload and time costs associated with re-labeling data, and solving the issues of subjectivity and reliability in the quality assessment of product appearance defects.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for assessing the importance of product appearance defects, characterized in that, Includes the following steps: The dual-labeling method obtains the rating labeling results by classifying and comparing product appearance defects, thus obtaining objective rating labeling results. Defect target generation involves generating defect targets through data transformation and providing a relative score based on the transformation operation. Importance assessment, based on labeled scoring data, uses a trained importance regression model to predict the importance of product appearance defects.

2. The method according to claim 1, characterized in that: The grading system categorizes defect data into three levels of severity: minor, moderate, and severe. Minor defects are those that are not clearly visible in the image or are good products. Moderate defects are those that can be distinguished or whose category boundaries are somewhat confusing. Severe defects are those that are clearly identifiable and whose category is clearly defined.

3. The method according to claim 1, characterized in that: The comparison labeling method determines the ranking by comparing similar defects to see which is more serious, and calculates the severity score based on the ranking.

4. The method according to claim 3, characterized in that, The formula for calculating the severity score of defects is: Defect Severity Score = Defect Severity Level Score + Comparison Score; where the severity level scores correspond as follows: Slight is 0 points, Average is 0.3 points, and Severe is 0.6 points; Comparison Score = Number of times the comparison result is more severe / Total number of comparisons × Total score of the interval; The total scores of the intervals for Slight, Average, and Severe are 0.3, 0.3, and 0.4, respectively.

5. The method according to claim 1, characterized in that: The data transformations during defect target generation include scaling and rotation transformations. Rotation transformations do not change the target score, while scaling transformations do not change the defect level. However, the scores within each level are adjusted according to the scaling factor, based on the following formula: Within-level score = min(scaling factor × current within-level score, maximum within-level score); where the within-level score is the defective target score minus the severity score.

6. The method according to claim 1, characterized in that, The importance regression model consists of a backbone network based on MobileViT, pooling layers, and fully connected layers. The loss function used for model training is the Huber loss function.

7. A system for assessing the importance of product appearance defects, characterized in that, The system includes: The annotation module is used to perform dual annotation of product appearance defects, including grade annotation and comparison annotation, to obtain objective scoring annotation results; The defect target generation module uses the Stable Diffusion model to transform the labeled and scored defect data to generate defect targets, and provides a relative score based on the specific transformation operation. The model training module uses an importance regression model, which is trained based on labeled scoring data to predict the importance of product appearance defects.

8. The system according to claim 7, characterized in that, The rating labeling in the labeling module divides the data into three categories: minor, moderate, and severe according to the severity level. The comparison labeling in the labeling module determines the ranking based on which of the similar defects is more severe, and calculates the severity score based on the ranking.

9. The system according to claim 7, characterized in that, Data transformation includes scaling and rotation operations. Rotation does not change the target score, while scaling does not change the defect level, but the score within the level will be adjusted according to the scaling factor.

10. The system according to claim 7, characterized in that: The importance regression model includes the backbone network, pooling layer, and fully connected layer based on MobileViT.

Images