A nonlinear intelligent evaluation method for screen appearance defects

Abstract

This invention discloses a nonlinear intelligent evaluation method for screen appearance defects, comprising the following steps: imaging the screen using an industrial camera, detecting and extracting defect areas, and calculating the multidimensional feature vector of each defect; manually evaluating a series of typical defect samples, associating each defect sample and its feature vector with the evaluation results to construct a standard database; training a nonlinear evaluation model based on the collected evaluation database; evaluating the model performance through cross-validation, calculating accuracy, precision, recall, and F1 score; for newly detected defect samples, extracting feature vectors, inputting the prediction results of the trained evaluation model, and outputting the confidence level of the defect evaluation to assist in manual review.

Description

Technical Field

[0001] This invention relates to the field of display device quality inspection, and specifically to a nonlinear intelligent evaluation method for screen appearance defects. Background Technology

[0002] With the rapid development of display device technology, the appearance quality of screens is receiving increasing attention from consumers and manufacturers. Screen appearance inspection is a key step in the quality control process of display devices, mainly targeting and evaluating defects such as bright spots, dark spots, foreign objects, and scratches that may appear on the screen surface.

[0003] Currently, the industry mainly uses two methods for detecting screen appearance defects: manual visual inspection and automated optical inspection (AOI). Manual visual inspection relies on the experience and judgment of quality inspectors, which has problems such as strong subjectivity, low efficiency, and poor consistency. While the AOI method can improve inspection efficiency and consistency, most existing AOI systems use a judgment method based on linear thresholds. That is, a fixed threshold is set for each defect feature (such as contrast, area, length, aspect ratio, etc.). When a feature exceeds the corresponding threshold, the defect is judged as unqualified.

[0004] However, this linear threshold judgment method differs significantly from how the human eye perceives and evaluates defects. The human eye's evaluation of defects is a non-linear process that integrates multiple features, with different features interacting and requiring trade-offs. For example, for stain-like defects on a screen, when the grayscale value is large (the defect is brighter), the human eye pays more attention to the grayscale value and is relatively more tolerant of the area; while when the grayscale value is small, the human eye's tolerance for the defect area decreases.

[0005] Furthermore, existing linear threshold judgment methods are difficult to adapt to the differentiated needs of defect evaluation standards for different product models and application scenarios, resulting in test results that do not match actual quality requirements, leading to missed or over-detection, which in turn affects product yield and user experience.

[0006] Therefore, there is an urgent need to develop a nonlinear screen appearance defect evaluation method that can simulate the characteristics of human eye perception, comprehensively consider multiple defect features, and flexibly adapt to different product requirements. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a nonlinear screen appearance defect evaluation method to achieve the following objectives: establishing a comprehensive defect evaluation model that conforms to the characteristics of human eye perception, breaking the limitation of independent judgment between different defect features; utilizing the experience and knowledge of professional engineers and quality inspectors to construct an intelligent evaluation system; providing a flexible intervention mechanism so that the model can adapt to the specific constraints of different product models and application scenarios; improving the accuracy of defect detection, reducing missed and over-detection, and optimizing production efficiency and product quality.

[0008] This invention proposes a nonlinear intelligent evaluation method for screen appearance defects, comprising the following steps:

[0009] Step S1: Image the screen using an industrial camera, detect and extract defect areas, and calculate the multidimensional feature vector for each defect;

[0010] Step S2: Manually evaluate a series of typical defect samples, associate each defect sample and its feature vector with the evaluation results, and construct a standard database;

[0011] Step S3: Train a nonlinear evaluation model based on the collected evaluation database;

[0012] Step S4: Evaluate model performance through cross-validation, and calculate accuracy, precision, recall, and F1 score;

[0013] Step S5: For newly detected defect samples, extract feature vectors, input the prediction results of the trained evaluation model, and output the confidence level of the defect evaluation to assist in manual review.

[0014] As a further aspect of the present invention, in step S1: the multidimensional features of the defect include contrast, area, length, aspect ratio, shape complexity, and location, and the feature vector of each defect is represented as:

[0015] X=[x1,x2,...,x n ], where x n This represents the value of the nth feature.

[0016] Specifically, the system includes: Contrast: quantifying the grayscale difference between the defect area and the background to intuitively reflect the visual salience of the defect; Area: based on the total number of pixels in the defect area, converted into physical units (mm²), accurately representing the actual size of the defect; Length: extracting the maximum axial length of the defect area to measure its spatial extent; Aspect Ratio: clearly depicting the morphological characteristics of the defect (e.g., linear, blocky, circular differences) through the ratio of the defect's major axis to its minor axis; Complexity: assessing the complexity of the defect's morphology based on the irregularity of its contour (e.g., jaggedness, smoothness); and Position: labeling the distribution of the defect on the screen with normalized relative position coordinates, facilitating a comprehensive evaluation by combining the sensitivity characteristics of different screen areas (e.g., defects are more easily detected in the center area than at the edges). This provides structured input data for subsequent machine learning models. Perceptual digitization: transforming subjective human judgment into calculable objective features, eliminating the randomness of manual evaluation.

[0017] It can achieve: Model empowerment: providing high-quality input for nonlinear machine learning, driving an accuracy improvement of 15%-20%; Scenario adaptation: through physical calibration and normalization design, the system can be quickly deployed to different products, equipment and inspection scenarios; Technical scalability: supporting smooth upgrades from single defects to multiple types of defects, and from manual features to deep learning features.

[0018] As a further aspect of the present invention, in step S2, the standard database constructed is:

[0019] D={(x1,y1),(x2,y2),...,(x m ,y m )};where x m Let y be the feature vector of the m-th defect sample. m This is the evaluation result of the experts on this sample.

[0020] As a further aspect of the present invention, in step S3, the method for training the nonlinear evaluation model includes: using a logistic regression model to predict the probability of a defect passing detection; using an SVM to separate Pass and Fail samples by finding the optimal separating hyperplane; using a random forest or gradient boosting tree to improve classification performance by combining the results of multiple decision trees; and using a multilayer perceptron, where the input layer receives the defect feature vector, and the hidden layer performs a nonlinear transformation to output the defect evaluation result. This improves the consistency and traceability of the evaluation criteria.

[0021] This can eliminate discrepancies in human judgment: In traditional manual quality inspection, different engineers may give different conclusions about the same defect due to subjective preferences (e.g., A thinks "a scratch length of 1.2mm is acceptable," while B thinks "must be <1mm"). Through a majority voting mechanism of an expert team (e.g., the majority vote of 5 engineers in Example 1), the database can accumulate unified evaluation standards, improving batch-to-batch consistency by 25%.

[0022] Standard traceability: The database records the original features and annotation results of each sample, supporting later traceability analysis (such as "why sample X was judged as Fail"), which facilitates the optimization of annotation rules or model parameters.

[0023] Other specific methods for training nonlinear evaluation models include: designing the nonlinear evaluation model and training it based on a collected expert evaluation database. This invention primarily employs the following machine learning models:

[0024] Logistic Regression:

[0025] Logistic regression models are used to predict the probability of defects passing detection.

[0026] Where β0, β1, ..., β n The model parameters are obtained using the maximum likelihood estimation method:

[0027] ;

[0028] Support Vector Machine (SVM):

[0029] SVM separates Pass and Fail samples by finding the optimal separating hyperplane:

[0030] Where K(X,X) i ) is the kernel function, and polynomial kernels, Gaussian kernels, etc. can be used to achieve nonlinear classification.

[0031] Decision tree ensemble method:

[0032] Ensemble methods such as random forests or gradient boosting trees improve classification performance by combining the results of multiple decision trees.

[0033] ;where h t (X) is a single decision tree model, W t These are the weighting coefficients.

[0034] Neural network model:

[0035] A multilayer perceptron receives defect feature vectors at the input layer, and outputs defect evaluation results through nonlinear transformations in the hidden layers.

[0036] z (l+1) =W （l） a (l) +b (l) ;

[0037] a (l+1) =g(z (l+1) ); where g(·) is the activation function, such as ReLU, Sigmoid, etc.

[0038] The nonlinear evaluation model is the "intelligent decision-making core" of this invention, which is reflected in: cognitive simulation: reproducing the nonlinear comprehensive judgment of humans on multi-dimensional defects through machine learning algorithms, solving the perceptual bias of traditional linear methods; scene adaptation: the multi-model system supports full scene coverage from simple rules to complex images, taking into account both accuracy and efficiency; rule fusion: in collaboration with the virtual point intervention mechanism, it achieves the dual constraint of "expert experience + hard rules", ensuring that the evaluation results meet both industry standards and product customization needs; technological evolution: the model architecture supports the upgrade from traditional machine learning to deep learning (such as introducing CNN to extract image features), laying the foundation for the intelligent upgrade of future defect detection.

[0039] As a further aspect of the present invention, step S3 also includes guiding model behavior by adding virtual sample points to the training data. Specifically, this involves determining absolute constraints based on product specifications and quality standards, generating virtual sample points that meet the constraints in the feature space, and then mixing the original expert evaluation data with the virtual sample points to form an enhanced dataset. .

[0040] The virtual point intervention mechanism is the core innovation of this invention in achieving "flexible rule embedding". Its value is reflected in: rule digitization: transforming product specifications into computable virtual samples, realizing the algorithm-level implementation of quality standards; scenario agility: quickly adapting to different products without relying on new data, significantly reducing cross-scenario deployment costs; decision controllability: preventing the model from deviating from business rules through mandatory constraints, ensuring that the detection results meet industry standards and customer requirements.

[0041] Technological foresight: This lays the foundation for future "rule transfer learning of unlabeled data," such as automatically generating virtual points through natural language processing parsing specifications. This mechanism, together with nonlinear models and expert databases, forms a synergistic closed loop, jointly constructing a new generation of intelligent defect evaluation system based on "data-driven + rule-guided" principles.

[0042] As a further aspect of the present invention, step S5 specifically involves: for newly detected defective samples, extracting feature vector X.new Input the prediction results from the trained model: Output the confidence level of the defect evaluation to assist in manual review: .

[0043] In this invention, evaluating model performance through cross-validation and calculating accuracy, precision, recall, and F1 score specifically includes: evaluating model performance through cross-validation and calculating accuracy, precision, recall, and F1 score.

[0044] ;

[0045] Model performance can be improved by optimizing model hyperparameters through methods such as grid search.

[0046] The nonlinear screen appearance defect evaluation method of the present invention has the following technical effects:

[0047] (1) Improve the accuracy of defect evaluation: Compared with the traditional linear threshold judgment method, the present invention simulates the human eye perception characteristics for comprehensive evaluation, which can improve the accuracy by 15%-20% and is more in line with the actual human eye perception of defects.

[0048] (2) Reduce the rate of missed detection and over-detection: By capturing the complex relationship between features through a nonlinear model, the rate of missed detection is reduced by more than 10% and the rate of over-detection is reduced by more than 12%, effectively balancing product quality and production efficiency.

[0049] (3) Strong adaptability: Through the virtual point intervention mechanism, the model can be flexibly adjusted according to the requirements of different product models and application scenarios, making it more adaptable.

[0050] (4) Improve consistency: Eliminate the differences in subjective judgment among different quality inspectors, achieve the unification of defect evaluation standards, and improve batch consistency by 25%.

[0051] (5) Significant economic benefits: By reducing manual inspection steps and improving inspection accuracy, quality inspection costs can be reduced by about 30%, while product yield can be increased by about 5%, resulting in significant economic benefits.

[0052] To more clearly illustrate the structural features and effects of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0053] Figure 1 This is a system architecture diagram of the nonlinear screen appearance defect evaluation method of the present invention;

[0054] Figure 2 This is a flowchart of the defect feature extraction process of the present invention;

[0055] Figure 3This is a flowchart illustrating the database construction process based on expert experience in this invention.

[0056] Figure 4 This is a flowchart of the training process for the nonlinear evaluation model of this invention;

[0057] Figure 5 This is a schematic diagram of the virtual point intervention mechanism of the present invention;

[0058] Figure 6 , Figure 7 Example diagrams of nonlinear decision boundaries for different combinations of defect features, where Figure 6 Example plot of nonlinear decision boundary for stain defects, y-axis range is [0-255], x-axis range is [0-0.1]; Figure 7 The figure shows an example of the nonlinear decision boundary for white spot defects. The y-axis range is [0-160] and the x-axis range is [0-0.05].

[0059] Figure 8 , Figure 9 This is a comparison chart showing the effectiveness of the present invention in stain defect evaluation, wherein... Figure 8 Using a simple decision-making logic to evaluate defects, the false negative rate was 35%, and the false positive rate was 74%. Figure 9 To evaluate the defect logic using nonlinear decision-making, the rate of missed detection is 5%, and the rate of over-detection is 20%. Detailed Implementation

[0060] The present invention will now be further described in conjunction with the accompanying drawings and relevant knowledge, and will be described clearly and completely. Obviously, the described applications are only some embodiments of the present invention, and not all embodiments.

[0061] Example 1: Refer to Figures 1-9 As shown, this invention provides a nonlinear intelligent evaluation method for screen appearance defects, comprising the following steps:

[0062] Step S1: Imaging the screen using an industrial camera, detecting and extracting defect areas, and calculating the multidimensional feature vector for each defect; specifically, this involves defect data acquisition and feature extraction. The system uses an industrial camera to image the screen, detects and extracts defect areas, and calculates the multidimensional feature vector for each defect. Defect features mainly include, but are not limited to: Contrast: the grayscale difference between the defect area and the background, reflecting the salience of the defect; Area: the total number of pixels in the defect area, convertible to mm²; Length: the maximum axial length of the defect area; Aspect Ratio: the ratio of the major axis to the minor axis of the defect area; and Complexity: the complexity of the defect contour.

[0063] Position: The relative coordinates of the defect's position on the screen;

[0064] Each defect can be represented as a feature vector:

[0065] X=[x1,x2,...,x n ], where x n This represents the value of the nth feature.

[0066] Step S2: Manually evaluate a series of typical defect samples, and associate each defect sample and its feature vector with the evaluation results to construct a standard database; specifically, expert database construction: invite professional engineers and quality inspection experts to manually evaluate a series of typical defect samples, and associate each defect sample and its feature vector with the expert evaluation results (Pass / Fail) to construct a standard database.

[0067] D={(x1,y1),(x2,y2),...,(x m ,y m )};where x m Let y be the feature vector of the m-th defect sample. m This represents the expert's evaluation of the sample (Pass=0, Fail=1).

[0068] Step S3: Train a nonlinear evaluation model based on the collected evaluation database; specifically, design the nonlinear evaluation model: train a nonlinear evaluation model based on the collected expert evaluation database. This invention mainly employs the following machine learning models:

[0069] Logistic Regression is a model used to predict the probability of a defect passing detection.

[0070] Where β0, β1, ..., β n The model parameters are obtained using the maximum likelihood estimation method: ;

[0071] Support Vector Machine (SVM): SVM separates Pass and Fail samples by finding the optimal separating hyperplane. Where K(X,X) i ) is the kernel function, and polynomial kernels, Gaussian kernels, etc. can be used to achieve nonlinear classification.

[0072] Decision tree ensemble methods: Ensemble methods such as random forests or gradient boosting trees improve classification performance by combining the results of multiple decision trees. ;where h t (X) is a single decision tree model, W t These are the weighting coefficients.

[0073] Neural network model: Multilayer perceptron, the input layer receives the defect feature vector, and through the nonlinear transformation of the hidden layer, the output is the defect evaluation result.

[0074] z (l+1) =W （l） a (l) +b (l) ;a (l+1) =g(z (l+1) ); where g(·) is the activation function, such as ReLU, Sigmoid, etc.

[0075] Step S4: Evaluate model performance through cross-validation, calculating accuracy, precision, recall, and F1 score; specifically, a virtual point intervention mechanism: To meet the specific product's defect restrictions, this invention proposes a virtual point intervention mechanism, which guides model behavior by adding virtual sample points to the training data.

[0076] Limitation condition identification: Based on product specifications and quality standards, determine absolute limiting conditions, such as "defects with an area exceeding 0.2 mm² must be judged as Fail".

[0077] Virtual sample generation involves generating virtual sample points in the feature space that meet certain constraints. For example, for area constraints, a series of virtual samples with an area greater than 0.2 mm² and other features varying within a reasonable range are generated and marked as Fail.

[0078] The process of generating virtual sample points can be represented as: Among them, threshold i Let g(X) be the threshold value for the i-th feature, and g(X) be the composite constraint condition.

[0079] Hybrid data training combines the original expert evaluation data with virtual sample points to create an augmented dataset. The model is trained using augmented datasets to ensure that it can learn both the non-linear evaluation criteria of experts and strictly adhere to the absolute constraints of the product.

[0080] And model evaluation and optimization: Evaluate model performance through cross-validation, and calculate accuracy, precision, recall, and F1 score.

[0081] ;

[0082] Model performance can be improved by optimizing model hyperparameters through methods such as grid search.

[0083] Step S5: For newly detected defect samples, extract feature vectors, input the prediction results of the trained evaluation model, and output the confidence score of the defect evaluation to assist in manual review. Specifically, defect evaluation decision: For newly detected defect samples, extract feature vector X. new Input the prediction results from the trained model: Simultaneously, it can output the confidence level of the defect evaluation to assist in manual review. .

[0084] The nonlinear modeling method of this invention breaks through the linear threshold limitation, simulating the human eye's comprehensive judgment of multiple features, improving accuracy by 15%-20%. Virtual point intervention: It can quickly adapt to different product specifications without new data, shortening the deployment cycle from weeks to hours. Digitalization of expert experience: Through majority voting annotation and database accumulation, it eliminates differences in human judgment, improving batch consistency by 25%. Through a dual mechanism of "data-driven + rule-guided," a high-precision, highly adaptable, and traceable intelligent quality inspection system is constructed, providing core technical support for the automation upgrade of the screen manufacturing industry.

[0085] This invention utilizes a nonlinear machine learning model to comprehensively evaluate multidimensional defect features, simulating human visual perception for defect assessment. The expert evaluation database construction step involves collecting evaluation results from professional quality inspection engineers to build a model training dataset. A virtual point intervention mechanism guides model behavior by adding virtual sample points that meet specific constraints during model training. The nonlinear machine learning model includes at least one of logistic regression, support vector machine, random forest, or neural network. Virtual point generation considers the specific constraints of different defect types, achieving differentiated evaluation standards.

[0086] Specifically, the accuracy of defect evaluation is significantly improved, breaking through the limitations of traditional linear threshold methods that rely on "independent judgment of a single feature." By fusing multi-dimensional features (such as contrast × area × location) through a non-linear model, it simulates the human eye's comprehensive perception logic of defects based on "salience + size + shape + location." Data support: Compared to linear methods, accuracy is improved by 15%-20%. For example, in stain defect evaluation, the logistic regression model, by introducing feature interaction terms, increases accuracy from 72% to 80%, effectively identifying scenarios where linear methods easily misjudge, such as "high contrast, small area" or "low contrast, large area."

[0087] Application value: It reduces misjudgments caused by simplified judgment logic, making the detection results closer to the user's actual perception, and is especially suitable for products with high requirements for visual experience (such as high-end displays).

[0088] The model achieves a dual reduction in both false negative and false positive rates, balancing quality and efficiency. By learning complex relationships between features (such as the "length × directionality" of scratches and the "area × roundness" of particles), the nonlinear model avoids the "one-size-fits-all" flaw of linear thresholds. For example, the false negative rate is reduced by more than 10% by capturing "multi-feature collaborative defects" (such as shallow spots with low contrast but complex shapes) that traditional methods may overlook.

[0089] The pass rate is reduced by more than 12%: This reduces the misjudgment of good products caused by a single feature triggering the threshold (such as small-area high-contrast defects in the edge region being judged as Pass due to low position weight).

[0090] Production impact: The production line can reduce rework costs caused by missed inspections, while reducing the loss of good products caused by over-inspection. The measured overall efficiency of the production line is improved by about 15%.

[0091] The virtual point mechanism drives flexible adaptation across multiple scenarios. Adaptability: By generating virtual samples that conform to product specifications (e.g., "area > 0.2mm² must be judged as Fail" and "center area defect threshold reduced by 30%), the model can quickly adapt to different product models, quality standards, or application scenarios (e.g., differences in lighting between indoor and outdoor displays) without re-labeling data. Deployment efficiency: When a new model is launched, only the virtual point generation rules need to be adjusted (e.g., modifying area thresholds and position weights), and the system can complete adaptation within hours, shortening the cycle by more than 90% compared to traditional methods (which require re-collecting and labeling data).

[0092] Scenario Case: Differentiated virtual point rules were designed for multiple types of defects (stain / scratch / particle / mura), and the overall accuracy of the model improved from 75.6% to 93.2%, demonstrating its ability to generalize across defect types.

[0093] Standardized evaluation criteria and improved batch consistency highlight the value of standardization: By establishing a unified evaluation logic through an expert database (e.g., a majority vote of 5 engineers to annotate samples), the problem of inconsistent judgments for the same defect caused by individual differences in experience during manual quality inspection is eliminated. Actual test data shows that the consistency of evaluation results between batches increased from 65% to 90% (a 25% improvement).

[0094] Quality traceability: Model decisions can be traced back to feature vectors and expert-annotated data, which makes it easier to locate the root cause of quality problems (such as "a sudden increase in the missed detection rate of a certain batch due to the error in the extraction of scratch length features") and supports continuous process optimization.

[0095] Significant economic benefits, achieving a double harvest of cost reduction and efficiency improvement, with a 30% reduction in quality inspection costs: Reduced manual re-inspection steps: Model confidence assists in screening samples that require manual verification (e.g., confidence level <70%), resulting in a reduction of approximately 40% in actual manual intervention; Automation replaces some manual labor: A single production line can save 2-3 quality inspectors, reducing annual labor costs by over one million yuan (taking an average of 100,000 products inspected per day as an example).

[0096] A 5% increase in product yield: This reduces the number of defective products entering the market due to missed inspections, while also preventing the rejection of good products due to over-inspection. In a real-world test, the yield on one production line increased from 92% to 97%, resulting in an annual increase in production value exceeding ten million yuan. Long-term value: Through data accumulation and model iteration, system performance is continuously optimized, and the improvement in quality inspection costs and yield further amplifies over time.

[0097] The closed-loop collaboration of perception, algorithm, and rules: human eye perceives digitalization (multi-dimensional features) → nonlinear algorithm simulates decision-making (model training) → virtual point injection of rule constraints, forming a complete technical link of "bionic perception - intelligent decision-making - rule controllability", which is difficult to be replaced by a single technology (such as relying solely on linear algorithms or manual quality inspection).

[0098] Deeply adapted design for industrial scenarios: Feature engineering considers physical unit calibration (mm² / mm) and location normalization, and the model supports deployment across devices / products; the virtual point mechanism is compatible with absolute rules and flexible experience, adapting to the manufacturing industry's need for "standardization and customization to coexist".

[0099] The two-way value-added of data and algorithms: the richer the expert database, the more accurate the model; the more perfect the virtual point mechanism, the more comprehensive the rule coverage, forming a positive cycle of "data accumulation - algorithm optimization - scenario expansion" and building a technological moat.

[0100] Example 1: Stain Defect Evaluation Based on Logistic Regression

[0101] System setup, such as Figure 1 As shown, a screen appearance defect detection system was built, including a high-resolution industrial camera, a uniform light source, an automatic positioning platform, and an image processing workstation. The industrial camera has a resolution of 4096×3072 pixels, and its field of view covers the entire display screen area, ensuring that the actual size of each pixel is less than 0.01mm.

[0102] Defect detection and feature extraction, such as Figure 2As shown, the acquired screen images are first preprocessed, including image smoothing and background equalization. Then, an adaptive threshold segmentation method is used to detect potential defect regions, extracting the following features for each defect: Contrast: the average grayscale difference between the defect region and the background; Area: the total number of pixels in the defect region, converted to mm²; Length: the major axis of the circumscribed ellipse of the defect region, in mm; Aspect Ratio: the ratio of the major axis to the minor axis of the circumscribed ellipse; Position: the relative coordinates (x, y) of the defect center point on the screen.

[0103] Expert database construction, such as Figure 3 As shown, 100 sample units were randomly selected from the production line, containing various degrees of stain defects. Five professional quality control engineers were invited to evaluate each defect (Pass / Fail), recording the feature vector of each defect and the final evaluation result (determined by majority vote). An expert evaluation database containing 300 samples was constructed, including 180 Pass samples and 120 Fail samples.

[0104] Training of nonlinear evaluation models, such as Figure 4 As shown, a logistic regression model is used for training. Considering the potential interactions between features, a second-order interaction term is added as a feature in addition to the original features:

[0105] ;

[0106] For stain defects, the focus is on the interaction between contrast and area, introducing a cross term x. contrast× x area .

[0107] Virtual point intervention, such as Figure 5 As shown, based on customer needs and product specifications, the following hard constraints are set:

[0108] When the stain area exceeds 0.035 mm² and the contrast ratio is less than 70, it must be judged as Fail; when the stain area is less than 0.005 mm², it must be judged as Pass.

[0109] Generate 100 virtual points in the feature space that satisfy the above conditions:

[0110] 1. The area is uniformly distributed between 0.035-0.04 mm², and the contrast is uniformly and randomly distributed between grayscale values ​​of 0-70. These virtual points are marked as Fail and added to the training dataset.

[0111] 2. The area is uniformly distributed between 0.03 and 0.035 mm², and the contrast is uniformly and randomly distributed between grayscale values ​​of 0 and 70. These virtual points are labeled as "Pass" and added to the training dataset.

[0112] 3. The area is uniformly distributed between 0 and 0.005 mm², and the contrast is uniformly and randomly distributed between 0 and 255 gray values. These virtual points are labeled as Pass and added to the training dataset.

[0113] Model evaluation and optimization: Five-fold cross-validation was used to evaluate model performance, with the following results: accuracy: 80%; precision: 90.7%; recall: 67%. Regularization parameters and learning rate were optimized through grid search to finally determine the optimal model parameters.

[0114] In practical applications, the trained model is deployed to the production line to evaluate newly detected stain defects in real time. The system provides the evaluation result (Pass / Fail) and confidence level. When the confidence level is below 70%, the sample is marked as requiring manual review.

[0115] In this embodiment, a high-resolution industrial camera is used in the system setup phase to ensure high-precision image acquisition, providing a reliable data source for subsequent accurate defect detection. Introducing second-order interaction terms, especially contrast and area interaction terms, into the logistic regression model training better captures the interactions between features, improving the model's fitting ability and accuracy. The virtual point intervention mechanism, by setting hard constraints and generating virtual points to be added to the training dataset, ensures the model strictly adheres to product specifications, effectively reducing the false positive rate. Five-fold cross-validation is used to evaluate model performance, and grid search is used to optimize model parameters, ensuring good performance in practical applications. In practical applications, the system outputs evaluation results and confidence levels. Samples with low confidence levels are marked as requiring manual review, achieving human-machine collaboration and further improving the reliability of the detection results.

[0116] Reference Figures 1-9 As shown in Example 2, the system performs a comprehensive evaluation of multiple types of defects based on random forest. In this example, the system simultaneously processes multiple types of screen appearance defects, including four types of defects: stain, scratch, particle, and mura (uniform brightness).

[0117] Multi-type defect feature extraction: For different types of defects, in addition to common features, their unique features are extracted: Scratch-specific features: linearity, directionality; Particle-specific features: roundness, density; Mura-specific features: gradation degree, texture complexity.

[0118] The hierarchical model structure first classifies defects using a convolutional neural network, and then evaluates each type of defect using a separately trained random forest model.

[0119] The virtual dot strategy designs differentiated virtual dot generation strategies for different types of defects. For example, for scratch-type defects, a length exceeding 1.5mm must be judged as a Fail; for particle-type defects, a location in the center of the screen and an area exceeding 0.1mm² must be judged as a Fail.

[0120] In practice, compared with the traditional linear threshold method, the overall evaluation accuracy of this embodiment increased from 75.6% to 93.2%, the false negative rate decreased from 12.3% to 2.8%, and the false positive rate decreased from 15.7% to 3.5%.

[0121] This embodiment extracts unique features for different types of defects, enabling a more detailed description of defect characteristics and providing rich information for accurate classification and evaluation. The powerful classification capabilities of CNNs are combined with the excellent evaluation performance of random forests, fully leveraging the advantages of both models and improving the overall performance of the evaluation system. Differentiated virtual point generation strategies are designed for different types of defects, allowing the model to better adhere to the specific constraints of the product, enhancing the model's accuracy and reliability. Significant improvements have been achieved in overall evaluation accuracy, false negative rate, and false positive rate, demonstrating the effectiveness and superiority of this method in practical applications.

[0122] The technical principles of the present invention have been described above with reference to specific embodiments, which are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments; all technical solutions falling within the scope of the present invention's concept are within its protection scope. Those skilled in the art can conceive of other specific embodiments of the present invention without creative effort, and these embodiments will all fall within the protection scope of the present invention.

Claims

1. A nonlinear intelligent evaluation method for screen appearance defects, characterized in that, Includes the following steps: Step S1: Image the screen using an industrial camera, detect and extract defect areas, and calculate the multidimensional feature vector for each defect; Step S2: Manually evaluate a series of typical defect samples, associate each defect sample and its feature vector with the evaluation results, and construct a standard database; Step S3: Train a nonlinear evaluation model based on the collected standard database; Step S4: Evaluate model performance through cross-validation, and calculate accuracy, precision, recall, and F1 score; Step S5: For newly detected defect samples, extract feature vectors, input them into the trained nonlinear evaluation model, and output the confidence score of the defect evaluation to assist in manual review. Step S3 also includes guiding the model by adding virtual sample points to the training data. Specifically, based on product specifications and quality standards, determine absolute constraints and generate virtual sample points that meet the constraints in the feature space. Then, mix the standard database with the virtual sample points to form an enhanced dataset. Where D is the standard database; the process of generating the virtual sample points is expressed as follows: Among them, threshold i Let g(X) be the threshold value for the i-th feature, and g(X) be the composite constraint condition.

2. The nonlinear intelligent evaluation method for screen appearance defects as described in claim 1, characterized in that, In step S1: the multidimensional features of the defect include contrast, area, length, aspect ratio, shape complexity, and location. The feature vector of each defect is represented as: X=[x1,x2,...,x n ], where x n This represents the value of the nth feature.

3. The nonlinear intelligent evaluation method for screen appearance defects as described in claim 2, characterized in that, In step S2, the standard database constructed is as follows: D={(x1,y1),(x2,y2),...,(x m ,y m )};where x m Let y be the feature vector of the m-th defect sample. m This is the evaluation result of the experts on this sample.

4. The nonlinear intelligent evaluation method for screen appearance defects as described in claim 3, characterized in that, In step S3, the method for training the nonlinear evaluation model includes: using a logistic regression model to predict the probability of a defect passing detection; using an SVM to separate Pass and Fail samples by finding the optimal separating hyperplane; using a random forest or gradient boosting tree to improve classification performance by combining the results of multiple decision trees; and using a multilayer perceptron, where the input layer receives the defect feature vector and the nonlinear transformation of the hidden layer outputs the defect evaluation result.

5. The nonlinear intelligent evaluation method for screen appearance defects as described in claim 4, characterized in that, Training the nonlinear evaluation model specifically includes: predicting the probability of a defect passing detection using a logistic regression model. Where β0, β1, ..., β n The model parameters are obtained using the maximum likelihood estimation method: ; Support Vector Machine (SVM): SVM separates Pass and Fail samples by finding the optimal separating hyperplane. Where K(X,X) i ) is the kernel function; Decision tree ensemble methods: Random forest or gradient boosting tree ensemble methods improve classification performance by combining the results of multiple decision trees. ;where h t (X) is a single decision tree model, W t These are the weighting coefficients; Neural network model: Multilayer perceptron, the input layer receives the defect feature vector, and through the nonlinear transformation of the hidden layer, the output defect evaluation result is: z (l+1) =W （l） a (l) +b (l) ;a (l+1) =g(z (l+1) ); where g(·) is the activation function.

6. The nonlinear intelligent evaluation method for screen appearance defects as described in claim 5, characterized in that, Step S4 specifically involves: evaluating model performance through cross-validation, and calculating accuracy, precision, recall, and F1 score. ; 。 7. The nonlinear intelligent evaluation method for screen appearance defects as described in claim 6, characterized in that, Step S5 specifically involves: for newly detected defect samples, extracting the feature vector X. new Input the prediction results from the trained model: Output the confidence level of the defect evaluation to assist in manual review: .

Images