A display panel defect detection method and system

Abstract

The application discloses a display panel defect detection method and system, comprising the following steps: obtaining original image data of a display panel; preprocessing the original image data to generate standardized image data; extracting multi-dimensional defect features based on the standardized image data; inputting the multi-dimensional defect features into a pre-trained defect classification model to output a defect type and a confidence; generating a defect distribution map based on the defect type and the confidence, and adjusting a detection threshold based on the defect distribution map; optimizing the defect classification model based on the adjusted detection threshold, and generating a dynamic quality control strategy. The application has the following advantages and effects: breaking through the limitation of a single data source by multi-spectral fusion and multi-dimensional features, covering surface and internal defects; adapting to changes in the production line environment by dynamic threshold and model optimization, and ensuring long-term detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of display panel inspection technology, and in particular to a method and system for detecting defects in display panels. Background Technology

[0002] As display panel manufacturing processes become increasingly complex and sophisticated, defect detection technology faces the need to upgrade from "macroscopic visible defect identification" to "precise classification of microscopic multimodal defects." Traditional detection methods primarily rely on visible light imaging technology, using single-dimensional image features such as edge contours and grayscale differences to determine defects. However, the types of defects in display panels are becoming increasingly diverse, including microcracks, bright spots, dark spots, color shifts, and thermal anomalies. Especially in high-end products such as new flexible screens and Mini-LicroLEDs, defect morphologies exhibit cross-modal correlations; for example, infrared thermal radiation anomalies may be accompanied by microtexture changes under visible light. Consequently, traditional methods suffer from a core problem: insufficient multimodal feature fusion and dynamic adaptability, specifically manifested as follows:

[0003] Existing technologies for multispectral data fusion typically employ fixed-weight fusion strategies, such as directly linearly superimposing visible and infrared images. This ignores the dynamic differences in signal-to-noise ratios (SNR) between different spectral bands and the correlation of physical characteristics in local areas. For instance, in areas with uneven illumination, the SNR of visible light images drops significantly, while infrared images may retain valuable information due to thermal radiation characteristics. However, fixed fusion weights can lead to noise amplification or loss of valuable features. Furthermore, defect classification models rely on static detection thresholds and cannot dynamically adjust classification boundaries based on changes in the production line environment, such as panel batch differences, equipment aging, and new types of defects. This results in fluctuating false positive and false negative rates throughout the production cycle, making it difficult to meet the quality control requirements of high-precision manufacturing. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for detecting defects in display panels, so as to solve the problems mentioned in the background art.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0006] This invention provides a method for detecting defects in a display panel, comprising the following steps:

[0007] S100. Acquire the raw image data of the display panel; wherein the raw image data is acquired by a multispectral imaging device, including visible light band sub-images and infrared band images;

[0008] S200. The original image data is preprocessed to generate standardized image data; wherein, the preprocessing includes noise suppression, illumination equalization and image registration, and the pixel resolution of the standardized image data is uniformly set to a preset threshold.

[0009] S300. Extract multi-dimensional defect features based on the standardized image data; wherein, the multi-dimensional defect features include texture distribution features, brightness gradient features, and chromaticity shift features;

[0010] S400. Input the multi-dimensional defect features into the pre-trained defect classification model and output the defect type and confidence level.

[0011] S500: Generate a defect distribution map based on the defect type and confidence level, and adjust the detection threshold based on the defect distribution map;

[0012] S600 optimizes the defect classification model based on the adjusted detection threshold and generates a dynamic quality control strategy.

[0013] By adopting the above technical solution, and integrating visible light and infrared band images, the sensitivity of different spectra to panel micro-defects is fully utilized, enhancing the ability to capture defect information. Standardized preprocessing ensures the consistency of image input, providing a unified data foundation for subsequent model learning and improving the model's generalization ability. Joint extraction of texture distribution, brightness gradient, and chromaticity shift features covers the differences in the physical characteristics of display panel defects. The detection threshold is dynamically adjusted based on historical data with the mean confidence level to avoid missed or false detections caused by fixed thresholds. At the same time, the incremental learning mechanism combined with model drift detection ensures that the classification model is continuously optimized as the production line environment changes. Finally, a defect distribution map is constructed and the detection threshold is optimized based on feedback, achieving a dynamic balance between detection accuracy and sensitivity. The overall solution breaks away from the traditional fixed-weight fusion strategy, is applicable to various complex backgrounds and diverse defect types, and has the advantages of multimodal feature fusion and dynamic adaptability.

[0014] A further provision is that S200 includes the following sub-steps:

[0015] The multispectral images in the original image data are subjected to noise suppression processing; wherein, the noise suppression processing includes:

[0016] Gaussian filtering is applied to the visible light band sub-image to remove noise, resulting in the first intermediate image;

[0017] Nonlocal mean denoising is applied to the infrared band sub-images to obtain the second intermediate image;

[0018] Image registration is performed between the first intermediate image and the second intermediate image, and spatial offset error between multispectral images is eliminated by a feature point matching algorithm; wherein, the feature point matching algorithm includes a scale-invariant feature transformation algorithm and a random sample consensus algorithm;

[0019] The registered first intermediate image and the second intermediate image are fused pixel-level to generate a fused image; wherein, the value of each pixel in the fused image is the weighted sum of the corresponding pixels in the two registered images, and the weights are dynamically adjusted according to the signal-to-noise ratio of the pixel region;

[0020] The fused image is subjected to illumination equalization processing, and the contrast-limited adaptive histogram equalization algorithm is used to eliminate uneven illumination distribution.

[0021] The image after illumination equalization is uniformly transformed to a preset threshold resolution to generate the standardized image data; wherein, the preset threshold is a fixed pixel density value set according to the display panel detection standard to ensure the size consistency of all input images.

[0022] By employing the above technical solutions, the visible light sub-image uses Gaussian filtering to smooth noise while preserving high-frequency texture details, while the infrared sub-image uses non-local mean denoising to effectively suppress thermal noise and preserve thermal radiation distribution characteristics. The combination of feature point matching based on scale-invariant feature transformation and random sampling consensus algorithm solves the image misalignment problem caused by parallax of multispectral cameras, ensuring pixel-level alignment between the visible light and infrared sub-images and avoiding fusion artifacts. Weights are dynamically allocated based on the signal-to-noise ratio difference between the visible light and infrared images within a local window, prioritizing the preservation of details in high-quality sub-images. This results in improved contrast in defective areas and reduced background interference in the fused image. The limited contrast adaptive histogram equalization algorithm limits the local contrast enhancement amplitude, avoiding edge artifacts caused by excessive enhancement, and is particularly suitable for uniform backgrounds on display panels, such as OLED solid color displays.

[0023] A further setting is that the process of dynamically adjusting the signal-to-noise ratio includes:

[0024] Within the fused image, the signal-to-noise ratio of the visible light band sub-image and the signal-to-noise ratio of the infrared band sub-image are calculated separately, using local windows of a preset size as units.

[0025] Based on the signal-to-noise ratio difference between the visible light band sub-image and the infrared band sub-image, fusion weights are dynamically assigned; wherein:

[0026] If the signal-to-noise ratio of the visible light sub-image is higher than that of the infrared sub-image, then a first weight ratio is assigned to the visible light sub-image, and a complementary second weight ratio is assigned to the infrared sub-image.

[0027] If the signal-to-noise ratio of the visible light sub-image is lower than or equal to the signal-to-noise ratio of the infrared sub-image, then a balanced weight ratio is assigned to the visible light sub-image and the infrared sub-image.

[0028] Based on the assigned weight ratio between the visible light sub-image and the infrared sub-image, the corresponding pixel values ​​of the registered visible light band sub-image and the infrared band sub-image are weighted and fused to generate the pixel values ​​of the fused image; wherein, the allocation of the weight ratio between the visible light sub-image and the infrared sub-image is further corrected by combining the edge gradient magnitude of the image within the local window, specifically including:

[0029] If the edge gradient magnitude of the visible light band sub-image exceeds a preset threshold, the weight ratio of the visible light band sub-image is increased.

[0030] If the thermal radiation intensity of the infrared band sub-image exceeds a preset threshold, the weight ratio of the infrared band sub-image is increased.

[0031] By adopting the above technical solutions, the dynamic signal-to-noise ratio adjustment mechanism effectively improves the local contrast and feature fidelity of the fused image through a region-aware weighting strategy. In multispectral image fusion, different regions may exhibit different signal-to-noise ratios. This mechanism evaluates the signal-to-noise ratio of each region through a local window, dynamically optimizes the fusion weights, and avoids information loss problems that may be caused by fixed-ratio fusion. The fusion weight correction strategy is further refined by introducing edge gradient magnitude and thermal radiation intensity, thereby enhancing the processing of target detail boundaries and infrared active regions. This method is particularly suitable for detecting detail-sensitive defects such as microcracks, bright spots, and thermal anomalies, bringing significant performance improvements to multispectral fusion and maintaining stable detection performance in complex lighting or low-contrast scenarios.

[0032] A further setting is that S300 includes the following sub-steps:

[0033] The standardized image data is subjected to a three-level discrete wavelet transform to extract the energy values ​​of each level of high-frequency sub-bands, and the energy values ​​of each level are weighted and summed according to a preset ratio to obtain the texture distribution features.

[0034] The Sobel operator is used to calculate the horizontal and vertical gradient maps of the standardized image data. The square root of the sum of the squares of the corresponding pixel values ​​in the two gradient maps is taken to generate a comprehensive gradient magnitude map. The mean and variance of the comprehensive gradient magnitude map are used as the brightness gradient features.

[0035] The standardized image data is converted from the RGB color space to the Lab color space, the color difference value ΔE of each pixel is calculated, a color difference matrix is ​​generated, and the variance and kurtosis of the color difference matrix are used as chromaticity shift features.

[0036] By adopting the above technical solution and introducing three-level discrete wavelet transform, Sobel operator, and color difference analysis, this method achieves comprehensive feature extraction from texture, brightness, and chromaticity. Wavelet transform can extract high-frequency texture information at multiple scales, making it suitable for identifying fine-grained defects such as particles and scratches. The comprehensive gradient magnitude map calculated by the Sobel operator effectively describes the brightness change trend and can capture structural anomalies such as blurred edges and bright lines. Color difference ΔE and its statistics locate problems such as color spots and color casts from the perspective of color consistency. The three types of features together form a high-dimensional representation space, providing more comprehensive and discriminative input data for the defect classification model, effectively improving the model's recognition robustness and classification accuracy.

[0037] A further provision is that S400 includes the following sub-steps:

[0038] S410. Based on the texture distribution features, brightness gradient features, and chromaticity shift features, a multimodal interaction network is constructed; specifically including:

[0039] The correlation matrix between features is calculated using a cross-modal attention mechanism to generate interaction weights; the cross-modal attention mechanism includes:

[0040] Channel-by-channel correlation calculation is performed between texture features and brightness gradient features to generate channel attention maps;

[0041] Spatial domain correlation analysis is performed on the chromaticity shift features and fusion results to generate a spatial attention map;

[0042] The channel attention map and the spatial attention map are fused according to dynamic weights to generate a joint feature representation;

[0043] S420. The joint feature representation is input into a heterogeneous hybrid model for classification decision-making and mapped to a preset defect type label; the heterogeneous hybrid model includes a deep residual network branch, a graph neural network branch, and a collaborative decision-making module; wherein,

[0044] In the deep residual network branch, a ResNeXt module with adaptive dilated convolution is used to dynamically adjust the dilation rate according to the resolution of the input features, and to reduce computational redundancy and improve inference efficiency through cross-stage local connection structure.

[0045] In the graph neural network branch, a spatial topology graph of the defect region is constructed, nodes represent pixel features, and edge weights are calculated based on the gradient similarity of adjacent pixels. Then, a graph attention network is used to aggregate node information to enhance the context awareness of irregular defects.

[0046] In the collaborative decision-making module, the outputs of the deep residual network branch and the graph neural network branch are probabilistically fused to obtain the fusion probability;

[0047] S430: Generate post-calibration confidence and associate the output with the defect type label; specifically including:

[0048] The fusion probability is subjected to probability calibration processing, and the probability distribution is adjusted using a temperature scaling algorithm to obtain the calibrated probability value; wherein, the temperature parameter of the temperature scaling algorithm is optimized based on the training data distribution of the defect classification model;

[0049] According to the preset defect type label mapping table, the calibrated probability value is associated with the corresponding defect type label; wherein, the defect type label mapping table includes multiple defect type names and corresponding classification probability thresholds, and the classification probability threshold of each defect type is dynamically set based on the detection frequency of that type of defect in historical detection data;

[0050] If the calibrated probability value exceeds the classification probability threshold for the corresponding defect type, the defect type and the corresponding calibrated probability value are output as the confidence level; if the calibrated probability value does not exceed the classification probability threshold, the detection result is marked as an unidentified defect, and a manual review process is triggered.

[0051] The output defect types and their confidence levels are recorded in the historical detection database, and detection logs are generated based on timestamps and panel numbers.

[0052] By adopting the above technical solutions, a multimodal interactive network is constructed and a deep residual network and graph neural network are heterogeneously fused, significantly enhancing the model's understanding and classification capabilities for complex defects. The cross-modal attention mechanism strengthens the synergistic effect between texture, brightness, and chromaticity features, enabling the model to focus on key feature regions during the learning process and improving the semantic consistency of feature representation. The graph neural network module achieves sensitive identification of irregular and diffuse defects through spatial topology mapping and graph attention aggregation. The collaborative decision-making module integrates the outputs of the two branches, comprehensively considering local details and global structural information, further improving classification accuracy. The temperature scaling calibration algorithm optimizes the interpretability and credibility of the output confidence, making the model more controllable and practically applicable in engineering.

[0053] A further setting is that S500 includes the following sub-steps:

[0054] S510. Generate a defect distribution map based on the defect type and confidence level; wherein, the defect distribution map includes multiple defect type labels and their corresponding average confidence levels; specifically:

[0055] Based on the defect type and confidence level output in S400, the frequency of occurrence of each defect type in the same batch of inspections is statistically analyzed, and its average confidence level is calculated; the average confidence level is the arithmetic mean of the confidence levels of all inspection results under the same defect type.

[0056] S520. Adjust the detection threshold based on the defect distribution map; specifically:

[0057] Obtain the historical average confidence score of the same type of defect in the historical detection data. The historical average confidence score is the average confidence score of the defect type in the previous M detections.

[0058] Determine whether the average confidence score of the same defect type in N consecutive detections exceeds a preset percentage of the corresponding historical average confidence score; where N is a preset positive integer and N≤M, and the preset percentage is a threshold adjustment ratio set according to the detection accuracy requirements;

[0059] If the average confidence level of the same defect type in N consecutive detections exceeds the preset percentage of the historical average confidence level, then the detection threshold for that defect type is lowered; otherwise, the detection threshold for that defect type is raised.

[0060] S530. The adjusted detection threshold is associated with the defect distribution map and stored for easy retrieval.

[0061] By adopting the above technical solutions, a closed loop is established between the model detection results and the actual distribution feedback through the construction of the defect distribution map and the mechanism for adjusting the detection threshold based on it. By calculating the frequency and confidence average of each type of defect in different detection batches, a statistical trend of defect detection is formed, providing theoretical support for dynamically adjusting the classification threshold. The dynamic adjustment mechanism of the detection threshold enables the system to adapt to the actual deviations of different production batches, improve the sensitivity of defect identification, and reduce the false judgment rate. It achieves high stability output in multi-batch data environments, improves the consistency of defect detection and the overall quality control level, and is particularly suitable for automated production line scenarios with frequent fluctuations in production conditions.

[0062] A further provision is that S600 includes the following sub-steps:

[0063] S610. The adjusted detection threshold is associated with and stored in the detection log in the historical detection database to construct a detection parameter evolution map; wherein, the detection parameter evolution map includes the detection threshold change trajectory, the confidence level mean fluctuation curve, the defect type distribution heat map, and the quality control strategy version history on the time series.

[0064] S620. Based on the evolution map of the detection parameters, perform model drift detection and calculate the performance degradation index of the defect classification model;

[0065] S630, Trigger the incremental learning mechanism to update the defect classification model:

[0066] S640. Generate dynamic quality control strategies based on the updated defect classification model:

[0067] S650: Verify the effectiveness of dynamic quality control strategies and manage version rollbacks.

[0068] By adopting the above technical solutions, the parameter evolution map and dynamic quality control strategy form a complete closed loop from data analysis to control decision-making. It not only records the evolution trajectory of threshold, confidence level, and defect type parameters, but also links them to model update strategies, establishing a full lifecycle quality tracking system. Model drift detection measures short-term and long-term shifts using the Euclidean distance of confidence level and the JS divergence of defect distribution, enabling quantitative monitoring of model performance. It can provide early warnings of model failure risks, avoiding quality control lags. In practical applications, it effectively solves the performance degradation problem caused by model aging over time and changes in data distribution, ensuring the continuous accuracy and stability of production testing.

[0069] A further setting is that S620 specifically includes:

[0070] The Euclidean distance between the mean confidence score of the most recent K tests and the mean confidence score of the historical tests is selected as the short-term drift.

[0071] Calculate the Jensen-Shannon divergence of various defects in the defect type distribution heatmap as the long-term distribution offset;

[0072] If the short-term drift exceeds a preset first threshold or the long-term distribution offset exceeds a preset second threshold, the defect classification model is determined to have experienced performance degradation.

[0073] Specifically, S630 includes:

[0074] When performance degradation of the defect classification model is detected, new defect sample data is extracted from the historical detection database, and the following processing is performed on the new defect sample data:

[0075] An adaptive oversampling algorithm is used to balance the number of samples for different defect types; wherein, the oversampling ratio in the adaptive oversampling algorithm is dynamically adjusted according to the detection frequency of each type of defect, and the oversampling ratio of low-frequency defects is inversely proportional to the detection frequency;

[0076] The balanced sample data is input into the defect classification model, the underlying parameters of the deep residual network branch are frozen, and only the graph attention layer parameters of the graph neural network branch and the fusion weight parameters of the collaborative decision module are updated.

[0077] Incremental training is performed using a momentum optimizer, with the learning rate decaying exponentially with each training epoch. The loss function is cross-entropy with class weights.

[0078] Specifically, S640 includes:

[0079] The rate of change in detection sensitivity for each type of defect in the updated defect classification model is statistically analyzed. If the rate of increase in sensitivity exceeds a preset threshold, the sampling frequency of the inspection station on the production line is adjusted. Specifically:

[0080] For the defect types with the highest sensitivity improvement rate, reduce the sampling frequency;

[0081] For defect types with the lowest sensitivity improvement rate, increase the sampling frequency;

[0082] The sampling frequency adjustment parameters, detection thresholds, and model version information are packaged into a dynamic quality control strategy and stored in the quality control strategy version history of the detection parameter evolution map. The current dynamic quality control strategy version is marked as V. n , where n is the incrementing sequence number;

[0083] The S650 specifically includes:

[0084] Within T hours after the current dynamic quality control strategy is implemented, collect data on the actual production line's defect missed detection rate and false detection rate.

[0085] If the decrease in the false negative rate exceeds the preset third threshold and the increase in the false positive rate is less than the preset fourth threshold, then the current dynamic quality control strategy will be marked as effective, and the version of the current dynamic quality control strategy will be changed to V. n Set as the default dynamic quality control strategy version;

[0086] Otherwise, perform the following operations:

[0087] Retrieve the last effective dynamic quality control strategy version V from the quality control strategy version history of the detection parameter evolution map. n-1 ;

[0088] The last effective dynamic quality control strategy version V n-1 The strategy parameters are reloaded into the defect classification model, the strategy rollback is completed, and a report is generated.

[0089] By adopting the above technical solutions, the incremental learning mechanism enables rapid model adaptation and evolution for newly emerging or difficult-to-identify defect types through dynamic sample balancing and local optimization of graph neural networks. The design that the oversampling ratio is inversely proportional to the detection frequency ensures the learning weight of small sample defects, mitigating training bias caused by sample imbalance. The dynamic control strategy adjusts the sampling frequency according to the increase in sensitivity, achieving optimal resource allocation and effectively improving detection efficiency and accuracy. The strategy verification and rollback mechanism ensures the controllability and security of model updates. Once an ineffective strategy is found, it can be quickly rolled back to the previous effective version, preventing bad strategies from continuously affecting production quality and enhancing the robustness and adaptability of the system.

[0090] The present invention also provides a display panel defect detection system, including a detection system comprising:

[0091] The data acquisition module is configured to acquire raw image data of the display panel through a multispectral imaging device, wherein the raw image data includes visible light band sub-images and infrared band images;

[0092] The preprocessing module is configured to perform noise suppression, image registration, pixel fusion, and illumination equalization on the original image data to generate standardized image data; the preprocessing module includes:

[0093] The visible light noise suppression unit uses a Gaussian filter to process the visible light band sub-image.

[0094] The infrared noise suppression unit uses a nonlocal mean denoising algorithm to process infrared band sub-images.

[0095] The registration and fusion unit achieves multispectral image registration based on scale-invariant feature transformation and random sample consensus algorithm, and performs pixel fusion according to the dynamic weight of signal-to-noise ratio;

[0096] The feature extraction module is configured to perform three-level discrete wavelet transform, Sobel gradient calculation and Lab color difference analysis on standardized image data to extract texture distribution features, brightness gradient features and color shift features.

[0097] The defect classification module includes a multimodal interactive network and a heterogeneous hybrid model. The multimodal interactive network fuses multi-dimensional defect features through a cross-modal attention mechanism. The heterogeneous hybrid model includes a deep residual network branch and a graph neural network branch, and outputs the defect type and calibrated confidence based on the collaborative decision module.

[0098] The dynamic optimization module is configured to adjust the detection threshold based on the defect distribution map, update the defect classification model parameters through an incremental learning mechanism, and generate a dynamic quality control strategy.

[0099] The storage module stores the evolution graph of detection parameters, historical detection database, and quality control strategy version records to support model drift detection and strategy rollback.

[0100] In summary, the present invention has the following beneficial effects:

[0101] By overcoming the limitations of single data sources through multispectral fusion and multidimensional features, this invention covers both surface and internal defects; through dynamic thresholds and model optimization, it adapts to changes in the production line environment and ensures long-term detection accuracy; through an automated closed-loop process, it automates the entire process from data acquisition to strategy generation, significantly improving detection efficiency and process optimization speed; this invention can be widely applied to production line inspection of display panels such as OLED, LCD, and Micro-LED, and has significant industrial application value. Attached Figure Description

[0102] Figure 1 This is a flowchart illustrating an embodiment;

[0103] Figure 2 This is a system structure block diagram for an embodiment. Detailed Implementation

[0104] The present invention will be further described in detail below with reference to the accompanying drawings.

[0105] As attached Figure 1 As shown;

[0106] This embodiment discloses a method for detecting defects in a display panel, including the following steps:

[0107] S100: Acquire raw image data of the display panel; wherein, the raw image data is acquired by a multispectral imaging device, including visible light band sub-images and infrared band images;

[0108] S200. Preprocess the original image data to generate standardized image data; wherein, the preprocessing includes noise suppression, illumination equalization and image registration, and the pixel resolution of the standardized image data is uniformly set to a preset threshold.

[0109] S300. Extract multi-dimensional defect features based on standardized image data; among which, multi-dimensional defect features include texture distribution features, brightness gradient features, and chromaticity shift features;

[0110] S400: Input multi-dimensional defect features into a pre-trained defect classification model and output the defect type and confidence level;

[0111] S500: Generate a defect distribution map based on the defect type and confidence level, and adjust the detection threshold based on the defect distribution map;

[0112] S600 optimizes the defect classification model based on the adjusted detection threshold and generates a dynamic quality control strategy.

[0113] Specifically, S200 includes the following sub-steps:

[0114] Noise suppression processing is performed on the multispectral images in the original image data; the noise suppression processing includes:

[0115] Gaussian filtering is applied to the visible light band sub-image to denoise, resulting in the first intermediate image. The kernel function of the Gaussian filter is a two-dimensional Gaussian distribution with a size of 5×5 pixels and a standard deviation σ=1.5. High-frequency noise is suppressed through convolution operation.

[0116] Non-local mean denoising is applied to the infrared band sub-image to obtain the second intermediate image; wherein, the similar block size of the non-local mean denoising is 3×3 pixels, the search window is 11×11 pixels, the attenuation parameter h=10, and noise suppression is achieved by weighted average pixel value;

[0117] Image registration is performed between the first and second intermediate images, and spatial offset errors between the multispectral images are eliminated using a feature point matching algorithm. This feature point matching algorithm includes a scale-invariant feature transformation algorithm and a random sample consensus algorithm. Specifically:

[0118] The scale-invariant feature transform algorithm is used to extract key points and descriptors from two images;

[0119] The affine transformation matrix is ​​calculated by eliminating mismatched pairs using a random sampling consensus algorithm.

[0120] The coordinate system of the infrared band sub-image is mapped to that of the visible light band sub-image, and the spatial offset error after registration is less than 0.5 pixels;

[0121] The registered first intermediate image and the second intermediate image are fused pixel by pixel to generate a fused image; wherein, the value of each pixel in the fused image is the weighted sum of the corresponding pixels in the two registered images, and the weights are dynamically adjusted according to the signal-to-noise ratio of the pixel region;

[0122] The fused image is subjected to illumination equalization processing, and a contrast-limited adaptive histogram equalization algorithm is used to eliminate uneven illumination distribution; wherein the parameters of the contrast-limited adaptive histogram include:

[0123] The image block size is 32×32 pixels;

[0124] The contrast limit threshold is set to 2.0 to suppress excessive enhancement;

[0125] Eliminate inter-block boundary effects through bilinear interpolation;

[0126] The image after illumination equalization is uniformly transformed to a preset threshold resolution to generate standardized image data. The preset threshold is a fixed pixel density value set according to the display panel detection standard to ensure the size consistency of all input images.

[0127] Specifically, the process of dynamically adjusting the signal-to-noise ratio includes:

[0128] Within the fused image, the signal-to-noise ratio of the visible light band sub-image and the signal-to-noise ratio of the infrared band sub-image are calculated separately, using a local window of a preset size of 5×5 pixels as the unit.

[0129] The fusion weights are dynamically assigned based on the signal-to-noise ratio difference between the visible light band sub-image and the infrared band sub-image; where:

[0130] If the signal-to-noise ratio of the visible light band sub-image is higher than that of the infrared band sub-image, then the visible light band sub-image is assigned a first weight ratio, and the infrared band sub-image is assigned a complementary second weight ratio.

[0131] If the signal-to-noise ratio of the visible light band sub-image is lower than or equal to that of the infrared band sub-image, then a balanced weight ratio is assigned to the visible light band sub-image and the infrared band sub-image.

[0132] Based on the assigned weight ratio of the visible light sub-image and the infrared sub-image, the corresponding pixel values ​​of the registered visible light band sub-image and the infrared band sub-image are weighted and fused to generate the pixel values ​​of the fused image. The weight ratio of the visible light sub-image and the infrared sub-image is further corrected by incorporating the edge gradient magnitude of the image within a local window, specifically including:

[0133] If the edge gradient magnitude of the visible light band sub-image exceeds a preset threshold, the weight ratio of the visible light band sub-image will be increased.

[0134] If the thermal radiation intensity of the infrared sub-image exceeds a preset threshold, the weight ratio of the infrared sub-image will be increased.

[0135] Example 1

[0136] A multispectral imaging device was used to image the OLED display panel, simultaneously acquiring images in the visible light band and the infrared band, both with a resolution of 2048×1536 pixels.

[0137] A 5×5 Gaussian filter (σ=1.5) was applied to the visible photon image to eliminate high-frequency noise. Experiments show that this parameter reduces the noise standard deviation from 12.3 to 3.8 while preserving edge details.

[0138] Non-local mean denoising (3×3 similar blocks, 11×11 search window, h=10) was applied to the infrared sub-images, improving the thermal radiation noise signal-to-noise ratio from 18.6dB to 29.4dB.

[0139] Applying CLAHE processing (contrast limit threshold of 2.0) to the fused image blocks reduced the illumination uniformity index from 45.7 to 18.2.

[0140] Unify the resolution to a standard value (e.g., 2560×1920 pixels) through affine transformation.

[0141] The signal-to-noise ratio (SNR) of the visible light band sub-image and the signal-to-noise ratio (SNR) of the infrared band sub-image are calculated using a 5×5 local window.

[0142] If SNR_Vis > SNR_IR + 2dB, then the visible light weight α = 0.7 and the infrared weight β = 0.3.

[0143] If the edge gradient magnitude of the visible light band sub-image is greater than 20 or the thermal radiation intensity of the infrared band sub-image is greater than 150, the corresponding weight is increased by 0.1, with an upper limit of 0.8.

[0144] The verification results of Example 1 are as follows:

[0145] In testing 100 samples containing microcracks (width <5μm), the defect detection rate of traditional fixed-weight fusion was 78%, while the detection rate of this method was increased to 93% after dynamic fusion, and the false detection rate was reduced from 12% to 6%.

[0146] Specifically, S300 includes the following sub-steps:

[0147] A three-level discrete wavelet transform is performed on the standardized image data to extract the energy values ​​of each level of high-frequency sub-bands. The energy values ​​of each level are then weighted and summed according to a preset ratio to obtain the texture distribution features.

[0148] The Sobel operator is used to calculate the horizontal and vertical gradient maps of the standardized image data. The square root of the sum of the squares of the corresponding pixel values ​​in the two gradient maps is taken to generate a comprehensive gradient magnitude map. The mean and variance of the comprehensive gradient magnitude map are used as the brightness gradient features.

[0149] The standardized image data is converted from the RGB color space to the Lab color space. The color difference value ΔE of each pixel is calculated to generate a color difference matrix. The variance and kurtosis of the color difference matrix are used as color shift features. The color difference value ΔE is the color difference between the current pixel and the mean value of the Lab color space of the defect-free area of ​​the display panel.

[0150] Example 2

[0151] Feature analysis was performed on 300 samples (including three types of defects: bright spots, dark spots, and color deviation). The classification accuracy of multi-dimensional feature combinations reached 89.2%, while the accuracy of single features (such as texture only) was only 67.5%.

[0152] Specifically, S400 includes the following sub-steps:

[0153] S410. Based on texture distribution features, brightness gradient features, and chromaticity shift features, a multimodal interaction network is constructed; specifically including:

[0154] The correlation matrix between features is calculated using a cross-modal attention mechanism to generate interaction weights; the cross-modal attention mechanism includes:

[0155] Channel-by-channel correlation calculation is performed between texture features and brightness gradient features to generate channel attention maps;

[0156] Spatial domain correlation analysis is performed on the chromaticity shift features and fusion results to generate a spatial attention map;

[0157] The channel attention map and the spatial attention map are fused with dynamic weights to generate a joint feature representation;

[0158] S420. The joint feature representation is input into the heterogeneous hybrid model for classification decision-making and mapped to a preset defect type label; the heterogeneous hybrid model includes a deep residual network branch, a graph neural network branch, and a collaborative decision-making module; wherein,

[0159] In the deep residual network branch, a ResNeXt module with adaptive dilated convolution is used to dynamically adjust the dilation rate according to the resolution of the input features, and to reduce computational redundancy and improve inference efficiency through cross-stage local connection structure.

[0160] In the graph neural network branch, a spatial topology graph of the defect region is constructed, where nodes represent pixel features and edge weights are calculated based on the gradient similarity of adjacent pixels. Then, a graph attention network is used to aggregate node information to enhance the contextual awareness of irregular defects.

[0161] In the collaborative decision-making module, the outputs of the deep residual network branch and the graph neural network branch are probabilistically fused to obtain the fusion probability.

[0162] The fusion probability formula is:

[0163] P final =σ(W c ·[P ResNeXt ,P GAT ]+b c )

[0164] Among them, W c and b c Let P be a learnable parameter, σ be the Sigmoid activation function, and P be the P-value. final Let P be the fusion probability. ResNeXt P represents the classification probability value output by the branch of the deep residual network. GAT The classification probability value output by the branch of the graph neural network;

[0165] S430: Generate post-calibration confidence and associate the output with the defect type label; specifically including:

[0166] The fusion probability is subjected to probability calibration, and the probability distribution is adjusted using a temperature scaling algorithm to obtain the calibrated probability value; wherein, the temperature parameter of the temperature scaling algorithm is optimized based on the training data distribution of the defect classification model.

[0167] The specific implementation of the temperature scaling algorithm is as follows:

[0168] The fusion probability P final The input to the calibration function is in the following form:

[0169] P calibrated =Softmax(z / T)

[0170] Where z represents the logits before fusion; T is a temperature parameter that controls the "smoothing" of the confidence level. The temperature parameter T is obtained by minimizing the negative log-likelihood loss on the validation set to ensure that the calibrated probability best represents the true label distribution; in this method, the validation set is derived from the historical training data of the defect classification model; P calibrated This is the calibrated probability value.

[0171] According to the preset defect type label mapping table, the calibrated probability value is associated with the corresponding defect type label; wherein, the defect type label mapping table includes multiple defect type names and corresponding classification probability thresholds, and the classification probability threshold of each defect type is dynamically set based on the detection frequency of that type of defect in historical detection data;

[0172] If the calibrated probability value exceeds the classification probability threshold for the corresponding defect type, the defect type and the corresponding calibrated probability value are output as the confidence level; if the calibrated probability value does not exceed the classification probability threshold, the detection result is marked as an unidentified defect, and the manual review process is triggered.

[0173] The output defect types and their confidence levels are recorded in the historical detection database, and detection logs are generated based on timestamps and panel numbers.

[0174] Specifically, S500 includes the following sub-steps:

[0175] S510. Generate a defect distribution map based on the defect type and confidence level; wherein, the defect distribution map includes multiple defect type labels and their corresponding average confidence levels; specifically:

[0176] Based on the defect types and confidence levels output in S400, the frequency of occurrence of each defect type in the same batch of inspections is statistically analyzed, and its mean confidence level is calculated; the mean confidence level is the arithmetic mean of the confidence levels of all inspection results under the same defect type.

[0177] S520. Adjust the detection threshold based on the defect distribution map; specifically:

[0178] Obtain the historical average confidence score of the same type of defect in the historical detection data. The historical average confidence score is the average confidence score of the defect type in the previous M detections.

[0179] Determine whether the average confidence score of the same defect type in N consecutive detections exceeds a preset percentage of the corresponding historical average confidence score; where N is a preset positive integer and N≤M, and the preset percentage is a threshold adjustment ratio set according to the detection accuracy requirements;

[0180] If the average confidence level of the same defect type in N consecutive tests exceeds the preset percentage of the historical average confidence level, then the detection threshold for that defect type is lowered; specifically: original detection threshold × (1 - adjustment coefficient); otherwise, the detection threshold for that defect type is raised; specifically: original detection threshold × (1 + adjustment coefficient); the adjustment coefficient ranges from 0.05 to 0.15.

[0181] S530. The adjusted detection threshold is associated with and stored in the defect distribution map for easy retrieval.

[0182] Specifically, S600 includes the following sub-steps:

[0183] S610. The adjusted detection threshold is associated with and stored in the detection log in the historical detection database to construct a detection parameter evolution map; wherein, the detection parameter evolution map includes the detection threshold change trajectory in the time series, the confidence level mean fluctuation curve, the defect type distribution heat map, and the quality control strategy version history.

[0184] S620. Based on the evolution spectrum of detection parameters, perform model drift detection and calculate the performance degradation index of the defect classification model;

[0185] S630, Trigger the incremental learning mechanism to update the defect classification model:

[0186] S640. Generate dynamic quality control strategies based on the updated defect classification model:

[0187] S650: Verify the effectiveness of dynamic quality control strategies and manage version rollbacks.

[0188] Specifically, the S620 includes:

[0189] The Euclidean distance between the mean confidence score of the most recent K tests and the mean confidence score of the historical tests is selected as the short-term drift.

[0190] Calculate the Jensen-Shannon divergence of various defects in the defect type distribution heatmap as the long-term distribution offset;

[0191] If the short-term drift exceeds the preset first threshold or the long-term distribution offset exceeds the preset second threshold, the defect classification model is deemed to have experienced performance degradation.

[0192] The S630 specifically includes:

[0193] When performance degradation of the defect classification model is detected, new defect sample data is extracted from the historical detection database, and the following processing is performed on the new defect sample data:

[0194] An adaptive oversampling algorithm is used to balance the number of samples for different defect types. The oversampling ratio in the adaptive oversampling algorithm is dynamically adjusted according to the detection frequency of each type of defect, and the oversampling ratio of low-frequency defects is inversely proportional to the detection frequency.

[0195] The balanced sample data is input into the defect classification model, the underlying parameters of the deep residual network branch are frozen, and only the graph attention layer parameters of the graph neural network branch and the fusion weight parameters of the collaborative decision module are updated.

[0196] Incremental training is performed using a momentum optimizer, with the learning rate decaying exponentially with each training epoch. The loss function is cross-entropy with class weights.

[0197] The S640 specifically includes:

[0198] The rate of change in detection sensitivity for each type of defect in the updated defect classification model is statistically analyzed. If the rate of increase in sensitivity exceeds a preset threshold, the sampling frequency of the inspection station on the production line is adjusted. Specifically:

[0199] For the defect type with the highest sensitivity improvement rate, reduce the sampling frequency to 50% of the original frequency.

[0200] For the defect type with the lowest sensitivity improvement rate, increase the sampling frequency to 150% of the original frequency.

[0201] The sampling frequency adjustment parameters, detection thresholds, and model version information are packaged into a dynamic quality control strategy and stored in the quality control strategy version history of the detection parameter evolution map. The current dynamic quality control strategy version is marked as V. n , where n is the incrementing sequence number;

[0202] Example 3

[0203] Sensitivity change rate statistics:

[0204] The sensitivity of "bright spot" detection increased from 85% to 93% (an improvement of 9.4%), and the sampling frequency decreased from 100% to 50%.

[0205] The sensitivity for "color deviation" increased from 78% to 81% (an improvement of 3.8%), and the sampling frequency increased from 100% to 150%.

[0206] Dynamic quality control strategy verification: After 24 hours of execution, the false negative rate decreased from 5.2% to 3.1%, and the false positive rate increased from 4.8% to 5.0% (< threshold 5.5%), and the strategy was marked as effective.

[0207] The S650 specifically includes:

[0208] Within T hours after the current dynamic quality control strategy is implemented, collect data on the actual production line's defect missed detection rate and false detection rate.

[0209] If the decrease in the false negative rate exceeds the preset third threshold and the increase in the false positive rate is less than the preset fourth threshold, then the current dynamic quality control strategy will be marked as effective, and the version of the current dynamic quality control strategy will be changed to V. n Set as the default dynamic quality control strategy version;

[0210] Otherwise, perform the following operations:

[0211] Retrieve the last effective dynamic quality control strategy version V from the quality control strategy version history of the detection parameter evolution map. n-1 ;

[0212] The last effective dynamic quality control strategy version V n-1 The strategy parameters are reloaded into the defect classification model, the strategy rollback is completed, and a report is generated.

[0213] Example 4

[0214] Table 1 shows the comparison data of production line testing after applying this method with the traditional method in Example 4;

[0215] Table 1. Comparison Data of Production Line Testing

[0216] index Traditional methods Example 4 Average detection rate 82.3％ 95.7％ False positive rate 8.9％ 4.1％

[0217] As attached Figure 2 As shown;

[0218] This embodiment also discloses a display panel defect detection system, including a detection system comprising:

[0219] The data acquisition module is configured to acquire raw image data of the display panel through a multispectral imaging device. The raw image data includes visible light band sub-images and infrared band images.

[0220] The preprocessing module is configured to perform noise suppression, image registration, pixel fusion, and illumination equalization on the original image data to generate standardized image data; the preprocessing module includes:

[0221] The visible light noise suppression unit uses a Gaussian filter to process the visible light band sub-image.

[0222] The infrared noise suppression unit uses a nonlocal mean denoising algorithm to process infrared band sub-images.

[0223] The registration and fusion unit achieves multispectral image registration based on scale-invariant feature transformation and random sample consensus algorithm, and performs pixel fusion according to the dynamic weight of signal-to-noise ratio;

[0224] The feature extraction module is configured to perform three-level discrete wavelet transform, Sobel gradient calculation and Lab color difference analysis on standardized image data to extract texture distribution features, brightness gradient features and color shift features.

[0225] The defect classification module includes a multimodal interactive network and a heterogeneous hybrid model. The multimodal interactive network integrates multi-dimensional defect features through a cross-modal attention mechanism, while the heterogeneous hybrid model includes a deep residual network branch and a graph neural network branch. Based on the collaborative decision-making module, it outputs the defect type and calibrated confidence.

[0226] The dynamic optimization module is configured to adjust the detection threshold based on the defect distribution map, update the defect classification model parameters through an incremental learning mechanism, and generate a dynamic quality control strategy.

[0227] The storage module stores the evolution graph of detection parameters, historical detection database, and quality control strategy version records to support model drift detection and strategy rollback.

[0228] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A method for detecting defects in a display panel, characterized in that, It includes the following steps: S100. Acquire the raw image data of the display panel; wherein the raw image data is acquired by a multispectral imaging device, including visible light band sub-images and infrared band images; S200. The original image data is preprocessed to generate standardized image data; wherein, the preprocessing includes noise suppression, illumination equalization and image registration, and the pixel resolution of the standardized image data is uniformly set to a preset threshold; the preset threshold is a fixed pixel density value set according to the display panel detection standard to ensure the size consistency of all input images. S300. Extract multi-dimensional defect features based on the standardized image data; wherein, the multi-dimensional defect features include texture distribution features, brightness gradient features, and chromaticity shift features; S400. Input the multi-dimensional defect features into the pre-trained defect classification model and output the defect type and confidence level. S500. Generate a defect distribution map based on the defect type and confidence level, and adjust the detection threshold based on the defect distribution map; specifically including the following sub-steps: S510. Generate a defect distribution map based on the defect type and confidence level; wherein, the defect distribution map includes multiple defect type labels and their corresponding average confidence levels; specifically: Based on the defect type and confidence level output in S400, the frequency of occurrence of each defect type in the same batch of inspections is statistically analyzed, and its average confidence level is calculated; the average confidence level is the arithmetic mean of the confidence levels of all inspection results under the same defect type. S520. Adjust the detection threshold based on the defect distribution map; specifically: Obtain the historical average confidence score of the same type of defect in the historical detection data. The historical average confidence score is the average confidence score of the defect type in the previous M detections. Determine whether the average confidence score of the same defect type in N consecutive detections exceeds a preset percentage of the corresponding historical average confidence score; where N is a preset positive integer and N≤M, and the preset percentage is a threshold adjustment ratio set according to the detection accuracy requirements; If the average confidence level of the same defect type in N consecutive detections exceeds the preset percentage of the historical average confidence level, then the detection threshold for that defect type is lowered; otherwise, the detection threshold for that defect type is raised. S530. The adjusted detection threshold is associated with the defect distribution map and stored for easy retrieval; S600 optimizes the defect classification model based on the adjusted detection threshold and generates a dynamic quality control strategy.

2. The method for detecting defects in a display panel according to claim 1, characterized in that: S200 includes the following sub-steps: The multispectral images in the original image data are subjected to noise suppression processing; wherein, the noise suppression processing includes: Gaussian filtering is applied to the visible light band sub-image to remove noise, resulting in the first intermediate image; Nonlocal mean denoising is applied to the infrared band sub-images to obtain the second intermediate image; Image registration is performed between the first intermediate image and the second intermediate image, and spatial offset error between multispectral images is eliminated by a feature point matching algorithm; wherein, the feature point matching algorithm includes a scale-invariant feature transformation algorithm and a random sample consensus algorithm; The registered first intermediate image and the second intermediate image are fused pixel-level to generate a fused image; wherein, the value of each pixel in the fused image is the weighted sum of the corresponding pixels in the two registered images, and the weights are dynamically adjusted according to the signal-to-noise ratio of the pixel region; The fused image is subjected to illumination equalization processing, and the contrast-limited adaptive histogram equalization algorithm is used to eliminate uneven illumination distribution. The image after illumination equalization is uniformly transformed to a preset threshold resolution to generate the standardized image data.

3. The method for detecting defects in a display panel according to claim 2, characterized in that: The process of dynamically adjusting the signal-to-noise ratio includes: Within the fused image, the signal-to-noise ratio of the visible light band sub-image and the signal-to-noise ratio of the infrared band sub-image are calculated separately, using local windows of a preset size as units. Based on the signal-to-noise ratio difference between the visible light band sub-image and the infrared band sub-image, fusion weights are dynamically assigned; wherein: If the signal-to-noise ratio of the visible light sub-image is higher than that of the infrared sub-image, then a first weight ratio is assigned to the visible light sub-image, and a complementary second weight ratio is assigned to the infrared sub-image. If the signal-to-noise ratio of the visible light sub-image is lower than or equal to the signal-to-noise ratio of the infrared sub-image, then a balanced weight ratio is assigned to the visible light sub-image and the infrared sub-image. Based on the assigned weight ratio between the visible light sub-image and the infrared sub-image, the corresponding pixel values ​​of the registered visible light sub-image and the infrared sub-image are weighted and fused to generate the pixel values ​​of the fused image; wherein, the allocation of the weight ratio between the visible light sub-image and the infrared sub-image is further corrected by combining the edge gradient magnitude of the image within the local window, specifically including: If the edge gradient magnitude of the visible light band sub-image exceeds a preset threshold, the weight ratio of the visible light band sub-image is increased. If the thermal radiation intensity of the infrared band sub-image exceeds a preset threshold, the weight ratio of the infrared band sub-image is increased.

4. The method for detecting defects in a display panel according to claim 1, characterized in that: S300 includes the following sub-steps: The standardized image data is subjected to a three-level discrete wavelet transform to extract the energy values ​​of each level of high-frequency sub-bands, and the energy values ​​of each level are weighted and summed according to a preset ratio to obtain the texture distribution features. The Sobel operator is used to calculate the horizontal and vertical gradient maps of the standardized image data. The square root of the sum of the squares of the corresponding pixel values ​​in the two gradient maps is taken to generate a comprehensive gradient magnitude map. The mean and variance of the comprehensive gradient magnitude map are used as the brightness gradient features. The standardized image data is converted from the RGB color space to the Lab color space, the color difference value ΔE of each pixel is calculated, a color difference matrix is ​​generated, and the variance and kurtosis of the color difference matrix are used as chromaticity shift features.

5. The method for detecting defects in a display panel according to claim 1, characterized in that: S400 includes the following sub-steps: S410. Based on the texture distribution features, brightness gradient features, and chromaticity shift features, construct a multimodal interaction network; Specifically, it includes: The correlation matrix between features is calculated using a cross-modal attention mechanism to generate interaction weights; the cross-modal attention mechanism includes: Channel-by-channel correlation calculation is performed between texture features and brightness gradient features to generate channel attention maps; Spatial domain correlation analysis is performed on the chromaticity shift features and fusion results to generate a spatial attention map; The channel attention map and the spatial attention map are fused according to dynamic weights to generate a joint feature representation; S420. The joint feature representation is input into a heterogeneous hybrid model for classification decision-making and mapped to a preset defect type label; the heterogeneous hybrid model includes a deep residual network branch, a graph neural network branch, and a collaborative decision-making module; wherein, In the deep residual network branch, a ResNeXt module with adaptive dilated convolution is used to dynamically adjust the dilation rate according to the resolution of the input features, and to reduce computational redundancy and improve inference efficiency through cross-stage local connection structure. In the graph neural network branch, a spatial topology graph of the defect region is constructed, nodes represent pixel features, and edge weights are calculated based on the gradient similarity of adjacent pixels. Then, a graph attention network is used to aggregate node information to enhance the context awareness of irregular defects. In the collaborative decision-making module, the outputs of the deep residual network branch and the graph neural network branch are probabilistically fused to obtain the fusion probability; S430: Generate post-calibration confidence and associate the output with the defect type label; specifically including: The fusion probability is subjected to probability calibration processing, and the probability distribution is adjusted using a temperature scaling algorithm to obtain the calibrated probability value; wherein, the temperature parameter of the temperature scaling algorithm is optimized based on the training data distribution of the defect classification model; According to the preset defect type label mapping table, the calibrated probability value is associated with the corresponding defect type label; wherein, the defect type label mapping table includes multiple defect type names and corresponding classification probability thresholds, and the classification probability threshold of each defect type is dynamically set based on the detection frequency of that defect type in historical detection data; If the calibrated probability value exceeds the classification probability threshold for the corresponding defect type, the defect type and the corresponding calibrated probability value are output as the confidence level; if the calibrated probability value does not exceed the classification probability threshold, the detection result is marked as an unidentified defect, and a manual review process is triggered. The output defect types and their confidence levels are recorded in the historical detection database, and detection logs are generated based on timestamps and panel numbers.

6. The method for detecting defects in a display panel according to claim 1, characterized in that: S600 includes the following sub-steps: S610. The adjusted detection threshold is associated with and stored in the detection log in the historical detection database to construct a detection parameter evolution map; wherein, the detection parameter evolution map includes the detection threshold change trajectory, the confidence level mean fluctuation curve, the defect type distribution heat map, and the quality control strategy version history on the time series. S620. Based on the evolution map of the detection parameters, perform model drift detection and calculate the performance degradation index of the defect classification model; S630, Trigger the incremental learning mechanism to update the defect classification model: S640. Generate dynamic quality control strategies based on the updated defect classification model: S650: Verify the effectiveness of dynamic quality control strategies and manage version rollbacks.

7. The method for detecting defects in a display panel according to claim 6, characterized in that: Specifically, S620 includes: The Euclidean distance between the mean confidence score of the most recent K tests and the mean confidence score of the historical tests is selected as the short-term drift. Calculate the Jensen-Shannon divergence of various defects in the defect type distribution heatmap as the long-term distribution offset; If the short-term drift exceeds a preset first threshold or the long-term distribution offset exceeds a preset second threshold, the defect classification model is determined to have experienced performance degradation. Specifically, S630 includes: When performance degradation of the defect classification model is detected, new defect sample data is extracted from the historical detection database, and the following processing is performed on the new defect sample data: An adaptive oversampling algorithm is used to balance the number of samples for different defect types; wherein, the oversampling ratio in the adaptive oversampling algorithm is dynamically adjusted according to the detection frequency of each type of defect, and the oversampling ratio of low-frequency defects is inversely proportional to the detection frequency; The balanced sample data is input into the defect classification model, the underlying parameters of the deep residual network branch are frozen, and only the graph attention layer parameters of the graph neural network branch and the fusion weight parameters of the collaborative decision module are updated. Incremental training is performed using a momentum optimizer, with the learning rate decaying exponentially with each training epoch. The loss function is cross-entropy with class weights. Specifically, S640 includes: The rate of change in detection sensitivity for each type of defect in the updated defect classification model is statistically analyzed. If the rate of increase in sensitivity exceeds a preset threshold, the sampling frequency of the inspection station on the production line is adjusted. Specifically: For the defect types with the highest sensitivity improvement rate, reduce the sampling frequency; For defect types with the lowest sensitivity improvement rate, increase the sampling frequency; The sampling frequency adjustment parameters, detection thresholds, and model version information are packaged into a dynamic quality control strategy and stored in the quality control strategy version history of the detection parameter evolution map. The current dynamic quality control strategy version is marked as V. n , where n is the incrementing sequence number; The S650 specifically includes: Within T hours after the current dynamic quality control strategy is implemented, collect data on the actual production line's defect missed detection rate and false detection rate. If the decrease in the false negative rate exceeds the preset third threshold and the increase in the false positive rate is less than the preset fourth threshold, then the current dynamic quality control strategy will be marked as effective, and the version of the current dynamic quality control strategy will be changed to V. n Set as the default dynamic quality control strategy version; Otherwise, perform the following operations: Retrieve the last effective dynamic quality control strategy version V from the quality control strategy version history of the detection parameter evolution map. n-1 ; The last effective dynamic quality control strategy version V n-1 The strategy parameters are reloaded into the defect classification model, the strategy rollback is completed, and a report is generated.

8. A display panel defect detection system, applied to the display panel defect detection method according to any one of claims 1-7, characterized in that, The system includes a detection system, which comprises: The data acquisition module is configured to acquire raw image data of the display panel through a multispectral imaging device, wherein the raw image data includes visible light band sub-images and infrared band images; The preprocessing module is configured to perform noise suppression, image registration, pixel fusion, and illumination equalization on the original image data to generate standardized image data; the preprocessing module includes: The visible light noise suppression unit uses a Gaussian filter to process the visible light band sub-image. The infrared noise suppression unit uses a nonlocal mean denoising algorithm to process infrared band sub-images. The registration and fusion unit achieves multispectral image registration based on scale-invariant feature transformation and random sample consensus algorithm, and performs pixel fusion according to the dynamic weight of signal-to-noise ratio; The feature extraction module is configured to perform three-level discrete wavelet transform, Sobel gradient calculation and Lab color difference analysis on standardized image data to extract texture distribution features, brightness gradient features and color shift features. The defect classification module includes a multimodal interactive network and a heterogeneous hybrid model. The multimodal interactive network fuses multi-dimensional defect features through a cross-modal attention mechanism. The heterogeneous hybrid model includes a deep residual network branch and a graph neural network branch, and outputs the defect type and calibrated confidence based on the collaborative decision module. The dynamic optimization module is configured to adjust the detection threshold based on the defect distribution map, update the defect classification model parameters through an incremental learning mechanism, and generate a dynamic quality control strategy. The storage module stores the evolution graph of detection parameters, historical detection database, and quality control strategy version records to support model drift detection and strategy rollback.

Images