Image-analysis-based system and method for fault detection of LCD system

By acquiring images with a high-resolution camera and performing homomorphic filtering, CLAHE contrast enhancement, and multi-scale feature fusion, the accuracy and robustness issues of Mura defect detection in LCD manufacturing were resolved, achieving more efficient Mura defect identification.

WO2026232725A1PCT designated stage Publication Date: 2026-11-12SHANGHAI ZEMSO ELECTRONICS TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/120261
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-09
Filing Date
2025-09-10
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively detect Mura defects in LCD manufacturing, especially low-contrast gradient Mura defects, resulting in high rates of missed detections and false positives, as well as poor robustness of the detection system.

Method used

An image analysis-based approach is employed, using high-resolution cameras to acquire images, performing homomorphic filtering and CLAHE contrast enhancement, multi-scale feature enhancement and fusion, combined with spatial-semantic dual-dimensional feature enhancement, to conduct statistical analysis of the Mura region, obtain statistical feature encoding vectors for the Mura region, and finally identify Mura defects.

Benefits of technology

It improves the accuracy of Mura defect detection, reduces misjudgments and missed detections caused by interference factors such as light sources and noise, and ensures the stability of the detection system under different production conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025120261_12112026_PF_FP_ABST
    Figure CN2025120261_12112026_PF_FP_ABST
Patent Text Reader

Abstract

The present application relates to the field of intelligent detection. Provided are an image-analysis-based system and method for fault detection of an LCD system. The method comprises: first, acquiring an LCD panel image captured by a high-resolution industrial camera; then, performing homomorphic filtering and CLAHE contrast enhancement processing on the image, so as to obtain an enhanced LCD panel image; then, performing multi-scale feature-enhanced fusion on the enhanced image, so as to obtain a multi-scale fused feature map of the LCD panel image; subsequently, performing statistical analysis on the feature map on the basis of segmented Mura regions, so as to obtain a statistical-feature encoding vector for the Mura regions; and finally, on the basis of the encoding vector, obtaining a Mura defect recognition result. In this way, features of Mura defects can be captured more accurately, thereby reducing misjudgment and missed detection caused by interference factors such as a light source and noise, and thus improving the accuracy of Mura defect detection; moreover, a detection system can adapt to different production conditions, thereby ensuring the stability of a detection result.
Need to check novelty before this filing date? Find Prior Art

Description

Image Analysis-Based Fault Detection System and Method for LCD Systems Technical Field

[0001] This application relates to the field of intelligent detection, and more specifically, to an LCD system fault detection system and method based on image analysis. Background Technology

[0002] With the widespread application of liquid crystal display (LCD) technology in consumer electronics, industrial control, and other fields, quality inspection during the manufacturing process has become a crucial link in ensuring product performance. Among the many types of LCD defects, Mura defects have become a difficult problem to detect in the industry due to their unique visual appearance. These defects manifest as cloud-like patterns with uneven brightness in the display area, and their causes involve multiple manufacturing processes, such as abnormalities in the backlight assembly and disordered arrangement of liquid crystal molecules.

[0003] Traditional manual visual inspection methods rely on engineers' subjective judgment and experience, resulting in inherent drawbacks such as low inspection efficiency and inconsistent standards. This is especially true when dealing with low-contrast, gradient murex defects, where the human eye's perception threshold for brightness differences is easily affected by external factors such as ambient lighting and viewing angle, leading to high rates of missed and false detections. Although automated inspection technologies based on machine vision have developed in recent years, they still face multiple technical bottlenecks in practical applications: during image acquisition, abnormal brightness distribution due to insufficient light source uniformity can easily be confused with real defects; camera noise and surface reflection interference significantly affect the image signal-to-noise ratio. In the image processing stage, traditional algorithms lack sensitivity to the gradient features of murex defects, threshold segmentation struggles to accurately capture low-brightness-difference defect areas, and simple edge detection operators cannot effectively distinguish between artifacts caused by uneven backlighting and real defects. More importantly, existing methods lack the ability to fuse and analyze multi-scale defect features and cannot establish an objective quantitative evaluation system consistent with subjective perception, making it difficult to guarantee the robustness of the inspection system under different production batches and environmental parameters.

[0004] Therefore, an optimized image analysis-based fault detection scheme for LCD systems is desired. Summary of the Invention

[0005] This application addresses the shortcomings of existing technologies by providing an LCD system fault detection system and method based on image analysis.

[0006] According to one aspect of this application, an image analysis-based fault detection method for an LCD system is provided, comprising:

[0007] Acquire images of LCD panels captured by a high-resolution industrial camera;

[0008] The LCD panel image is subjected to homomorphic filtering and CLAHE contrast enhancement to obtain an enhanced LCD panel image;

[0009] The enhanced LCD panel image is subjected to multi-scale feature enhancement and fusion to obtain a multi-scale fused feature map of the LCD panel image, including: performing wavelet decomposition and feature extraction on the enhanced LCD panel image to obtain a low-frequency approximate sub-band feature map, a horizontal detail sub-band feature map, a vertical detail sub-band feature map, and a diagonal detail sub-band feature map; and performing spatial-semantic dual-dimensional feature enhancement and fusion processing on the low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map to obtain the multi-scale fused feature map of the LCD panel image.

[0010] Statistical analysis based on Mura segmentation regions is performed on the multi-scale fusion feature map of the LCD panel image to obtain the statistical feature encoding vector of the Mura region;

[0011] Based on the statistical feature encoding vector of the Mura region, the Mura defect identification result is obtained.

[0012] According to another aspect of this application, an image analysis-based LCD system fault detection system is provided, comprising:

[0013] Image acquisition module, used to acquire LCD panel images captured by a high-resolution industrial camera;

[0014] An image preprocessing module is used to perform homomorphic filtering and CLAHE contrast enhancement on the LCD panel image to obtain an enhanced LCD panel image.

[0015] An image encoding module is used to perform multi-scale feature enhancement and fusion on the enhanced LCD panel image to obtain a multi-scale fused feature map of the LCD panel image. The image encoding module is used to: perform wavelet decomposition and feature extraction on the enhanced LCD panel image to obtain a low-frequency approximate sub-band feature map, a horizontal detail sub-band feature map, a vertical detail sub-band feature map, and a diagonal detail sub-band feature map; and perform spatial-semantic dual-dimensional feature enhancement and fusion processing on the low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map to obtain the multi-scale fused feature map of the LCD panel image.

[0016] The Mura region analysis module is used to perform statistical analysis on the multi-scale fusion feature map of the LCD panel image based on Mura segmentation regions to obtain the Mura region statistical feature encoding vector.

[0017] The defect identification module is used to obtain the Mura defect identification result based on the statistical feature encoding vector of the Mura region.

[0018] This application, by adopting the above technical solution, has significant technical effects:

[0019] The image analysis-based LCD system fault detection system and method provided in this application first acquires LCD panel images captured by a high-resolution industrial camera. Then, homomorphic filtering and CLAHE contrast enhancement processing are applied to the image to obtain an enhanced LCD panel image. Next, multi-scale feature enhancement fusion is performed on the enhanced image to obtain a multi-scale fused feature map of the LCD panel image. Subsequently, statistical analysis is performed on this feature map based on Mura segmentation regions to obtain statistical feature encoding vectors for the Mura regions. Finally, based on these encoding vectors, the Mura defect identification result is obtained. This approach can more accurately capture the features of Mura defects, reduce misjudgments and missed detections caused by interference factors such as light sources and noise, and improve the detection accuracy of Mura defects. Simultaneously, the detection system can adapt to different production conditions, ensuring the stability of the detection results. Attached Figure Description

[0020] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0021] Figure 1 is a flowchart of an image analysis-based LCD system fault detection method according to an embodiment of this application.

[0022] Figure 2 is a flowchart of step S3 in the LCD system fault detection method based on image analysis according to an embodiment of this application.

[0023] Figure 3 is a flowchart of step S31 in the LCD system fault detection method based on image analysis according to an embodiment of this application.

[0024] Figure 4 is a flowchart of step S32 in the LCD system fault detection method based on image analysis according to an embodiment of this application.

[0025] Figure 5 is a flowchart of step S321 in the LCD system fault detection method based on image analysis according to an embodiment of this application.

[0026] Figure 6 is a flowchart of steps S321-3 in the LCD system fault detection method based on image analysis according to an embodiment of this application.

[0027] Figure 7 is a flowchart of step S4 in the LCD system fault detection method based on image analysis according to an embodiment of this application.

[0028] Figure 8 is a system block diagram of an LCD system fault detection system based on image analysis according to an embodiment of this application. Detailed Implementation

[0029] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0030] With the widespread application of liquid crystal display (LCD) technology in many fields, quality inspection during its manufacturing process is crucial. Mura defects, as one of the many types of defects in LCDs, have become a difficult problem to detect in the industry due to their unique visual appearance. Their causes involve multiple manufacturing stages, such as abnormalities in the backlight assembly and disordered arrangement of liquid crystal molecules.

[0031] Traditional manual visual inspection relies on the subjective judgment of engineers, resulting in low efficiency and inconsistent standards. When dealing with low-contrast gradient murex defects, it is susceptible to interference from factors such as ambient lighting and viewing angle, leading to high rates of missed detections and false positives. While automated inspection technologies based on machine vision have made progress, they face multiple bottlenecks in practical applications: during image acquisition, insufficient light source uniformity, camera noise, and surface reflection interference affect detection; during image processing, traditional algorithms have low sensitivity to murex gradient features, making it difficult to accurately capture low-brightness-difference defect areas and effectively distinguish between artifacts and real defects. Furthermore, existing methods lack multi-scale feature fusion analysis capabilities and objective quantitative evaluation systems, making it difficult to guarantee the robustness of the detection system.

[0032] Based on this, the technical concept of this application is to construct a full-process detection system for Mura defects in LCD panels through a multi-level image enhancement and multi-scale feature fusion framework. Specifically, after acquiring the original image using a high-resolution camera, the illumination component and reflection component are separated by homomorphic filtering, and the local contrast is enhanced by the CLAHE algorithm, effectively overcoming uneven background illumination and surface reflection interference. Subsequently, the image is decomposed into different frequency bands and directional features to extract defect features with a larger receptive field, enhancing the sensitivity to low-contrast gradient defects. Then, through multi-scale feature fusion and semantic segmentation, the boundary distribution of different morphological Mura defects is accurately located. Finally, statistical features such as texture and brightness distribution of the defect area are extracted to establish a defect quantification model, achieving an objective evaluation consistent with human subjective perception. This scheme suppresses background noise interference through frequency-spatial domain joint enhancement technology, improves the detection capability of small defects through multi-resolution analysis, and solves the problem of traditional algorithms lacking quantitative standards for Mura defects through statistical feature modeling, significantly improving the system's adaptability to complex industrial environments while ensuring detection sensitivity.

[0033] Figure 1 is a flowchart of an image analysis-based LCD system fault detection method according to an embodiment of this application. As shown in Figure 1, the image analysis-based LCD system fault detection method according to an embodiment of this application includes: S1, acquiring an LCD panel image captured by a high-resolution industrial camera; S2, performing homomorphic filtering and CLAHE contrast enhancement on the LCD panel image to obtain an enhanced LCD panel image; S3, performing multi-scale feature enhancement fusion on the enhanced LCD panel image to obtain a multi-scale fused feature map of the LCD panel image; S4, performing statistical analysis based on Mura segmentation regions on the multi-scale fused feature map of the LCD panel image to obtain a statistical feature encoding vector of the Mura region; S5, obtaining a Mura defect identification result based on the statistical feature encoding vector of the Mura region.

[0034] In step S1, an image of the LCD panel is acquired using a high-resolution industrial camera. It should be understood that the information in the LCD panel image is essentially a digital mapping of the panel's optical performance. Spatially, the grayscale value of each pixel accurately reflects the brightness intensity of the corresponding area. This brightness distribution characteristic is directly related to the uniformity of the liquid crystal molecule arrangement and the optical diffusion efficiency of the backlight module. For example, when the grayscale value of a local area fluctuates abnormally, it may indicate uneven liquid crystal layer thickness or abnormal driving voltage leading to changes in transmittance, which is a typical manifestation of Mura defects. At the texture feature level, high-resolution images can clearly present the periodic arrangement structure of the TFT array. When this regularity is disrupted, it may mean pixel misalignment caused by packaging process defects or external force damage. Such texture anomalies are often accompanied by specific forms of Mura defects. In summary, by performing in-depth analysis of the information in the acquired LCD panel image, Mura defects can be effectively identified.

[0035] In step S2, homomorphic filtering and CLAHE contrast enhancement are applied to the LCD panel image to obtain an enhanced LCD panel image. Correspondingly, considering that uneven illumination and surface reflection interference during image acquisition significantly affect defect identification, the distribution of light sources in industrial inspection environments is difficult to achieve absolute uniformity. Images captured by cameras often contain low-frequency brightness fluctuations caused by differences in backlight module luminance or external environmental reflections. This interference manifests as large-area gradual brightness shifts in the spatial domain, and its frequency characteristics overlap with the low-frequency gradual characteristics of genuine Mura defects. Furthermore, changes in the relative position of the high-reflectivity material of the LCD surface and the inspection equipment easily generate specular reflections, leading to abnormal saturation of local pixel values ​​or the formation of highlight artifacts. These interference factors can mask the subtle contrast changes of genuine defects, making it difficult for traditional threshold segmentation algorithms to effectively distinguish defect areas from optical artifacts. Therefore, in order to enhance the local contrast of the image while eliminating the adverse effects of uneven illumination, noise, and reflection, this application applies homomorphic filtering and CLAHE contrast enhancement to the LCD panel image to obtain an enhanced LCD panel image. In other words, homomorphic filtering decomposes the image into illumination and reflection components through frequency domain filtering operations, dynamically compressing the low-frequency global brightness distribution while enhancing high-frequency detail features, effectively suppressing low-frequency brightness drift caused by uneven backlighting. The CLAHE algorithm, on the other hand, performs contrast-limited histogram equalization in local sub-regions, overcoming the limitations of traditional histogram equalization in adjusting global contrast. It enhances the local contrast of Mura defects while avoiding information distortion caused by over-enhancement of reflective areas. The combination of these two algorithms forms a dual enhancement mechanism of "global brightness correction - local contrast optimization": homomorphic filtering eliminates large-area background interference, and CLAHE enhances the gradient changes at defect edges, jointly improving the signal-to-noise ratio between low-contrast defects and the background.

[0036] In step S3, multi-scale feature enhancement fusion is performed on the enhanced LCD panel image to obtain a multi-scale fused feature map of the LCD panel image. Specifically, Figure 2 is a flowchart of step S3 in the LCD system fault detection method based on image analysis according to an embodiment of this application. As shown in Figure 2, step S3 includes: S31, performing wavelet decomposition and feature extraction on the enhanced LCD panel image to obtain a low-frequency approximate sub-band feature map, a horizontal detail sub-band feature map, a vertical detail sub-band feature map, and a diagonal detail sub-band feature map; S32, performing spatial-semantic dual-dimensional feature enhancement fusion processing on the low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map to obtain the multi-scale fused feature map of the LCD panel image.

[0037] In step S31, wavelet decomposition and feature extraction are performed on the enhanced LCD panel image to obtain a low-frequency approximate sub-band feature map, a horizontal detail sub-band feature map, a vertical detail sub-band feature map, and a diagonal detail sub-band feature map. Specifically, Figure 3 is a flowchart of step S31 in the LCD system fault detection method based on image analysis according to an embodiment of this application. As shown in Figure 3, step S31 includes: S311, performing wavelet decomposition on the enhanced LCD panel image to obtain a low-frequency approximate sub-band matrix, a horizontal detail sub-band matrix, a vertical detail sub-band matrix, and a diagonal detail sub-band matrix; S312, performing feature extraction based on dilated convolutional coding on the low-frequency approximate sub-band matrix, the horizontal detail sub-band matrix, the vertical detail sub-band matrix, and the diagonal detail sub-band matrix to obtain the low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map.

[0038] In step S311, wavelet decomposition is performed on the enhanced LCD panel image to obtain a low-frequency approximate sub-band matrix, a horizontal detail sub-band matrix, a vertical detail sub-band matrix, and a diagonal detail sub-band matrix. Correspondingly, considering that the morphological features of defects often exhibit multi-scale distribution characteristics, traditional image processing methods, when directly processing full-frequency domain information, can lead to feature confusion due to the coupling of high-frequency noise and low-frequency background interference. This is especially true for gradient-type Mura defects with contrast below the human eye's recognition threshold, where subtle brightness gradient changes are easily submerged in the background texture. More importantly, the diffuse edges of Mura defects exhibit different responses in different spatial directions. For example, horizontally extending band-like defects may exhibit a stronger feature response in the vertical detail sub-band, and traditional single-scale feature extraction methods cannot effectively separate this directionally sensitive feature. Therefore, in order to analyze and process detail information in different directions more meticulously, this application performs wavelet decomposition on the enhanced LCD panel image to obtain a low-frequency approximate sub-band matrix, a horizontal detail sub-band matrix, a vertical detail sub-band matrix, and a diagonal detail sub-band matrix. In this way, the image is decomposed into a low-frequency approximation subband and three high-frequency detail subbands (horizontal, vertical, and diagonal) using orthogonal wavelet basis functions. The low-frequency approximation subband focuses on the overall brightness distribution of the image, effectively characterizing the low-frequency brightness gradient features of Mura defects; the horizontal detail subband enhances the abrupt response of vertical edges in the image, making it suitable for detecting linear defects with horizontal extension characteristics; the vertical detail subband enhances horizontal edge features, exhibiting higher sensitivity to vertically distributed Mura regions; and the diagonal detail subband simultaneously captures diagonal texture changes, used to identify irregularly shaped cloud-like defects. This multi-directional, multi-scale decomposition method overcomes the limitation of traditional spatial domain processing on the single feature dimension, providing a physically meaningful feature representation foundation for subsequent feature encoding.

[0039] In step S312, feature extraction based on dilated convolutional coding is performed on the low-frequency approximation subband matrix, the horizontal detail subband matrix, the vertical detail subband matrix, and the diagonal detail subband matrix to obtain the low-frequency approximation subband feature map, the horizontal detail subband feature map, the vertical detail subband feature map, and the diagonal detail subband feature map. Correspondingly, considering that the low-frequency approximation subband matrix, the horizontal detail subband matrix, the vertical detail subband matrix, and the diagonal detail subband matrix respectively contain information of different scales and directions of the image, the fixed receptive field of a conventional convolutional kernel may be insufficient to capture the Mura region formed by large-scale brightness gradients when processing low-frequency approximation subbands due to insufficient coverage; while in high-frequency detail subbands, an excessively small convolutional window is prone to over-focusing on local noise and ignoring the continuous features of defect edges. Furthermore, the morphological differences of the Mura defect in the horizontal, vertical, and diagonal directions require the feature extractor to have orientation awareness. However, traditional convolutional operations lack specificity in their response to directional features, making it difficult to effectively extract key information from subbands in different directions. Therefore, in the technical solution of this application, feature extraction based on dilated convolutional coding is performed on the low-frequency approximate subband matrix, the horizontal detail subband matrix, the vertical detail subband matrix, and the diagonal detail subband matrix to obtain the low-frequency approximate subband feature map, the horizontal detail subband feature map, the vertical detail subband feature map, and the diagonal detail subband feature map. Specifically, for the low-frequency approximation subband, a dilated convolutional layer with a larger dilation rate is used to expand the receptive field without increasing the number of parameters, enabling it to cover a wider range of brightness gradient regions and effectively capture the overall distribution pattern of low-contrast mura. In the three high-frequency detail subbands (horizontal, vertical, and diagonal), directional dilated convolutional kernel groups are used. By adjusting the dilation direction of the convolutional kernels to align with the spatial characteristics of each subband, for example, matching the horizontal detail subband with a horizontally extending dilated convolutional pattern, the response sensitivity to edge features in specific directions is enhanced. This encoding method breaks through the isotropic limitation of traditional convolutional operations, enabling the feature extraction process to adapt to the spatial-frequency characteristics of different frequency bands.

[0040] In step S32, the low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map are subjected to spatial-semantic dual-dimensional feature enhancement fusion processing to obtain the multi-scale fusion feature map of the LCD panel image. Specifically, Figure 4 is a flowchart of step S32 in the image analysis-based LCD system fault detection method according to an embodiment of this application. As shown in Figure 4, step S32 includes: S321, performing spatial-semantic dual-dimensional feature enhancement on the low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map to obtain enhanced low-frequency approximate sub-band feature map, enhanced horizontal detail sub-band feature map, enhanced vertical detail sub-band feature map, and enhanced diagonal detail sub-band feature map; S322, fusing the enhanced low-frequency approximate sub-band feature map, the enhanced horizontal detail sub-band feature map, the enhanced vertical detail sub-band feature map, and the enhanced diagonal detail sub-band feature map to obtain the multi-scale fusion feature map of the LCD panel image.

[0041] In step S321, the low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map are enhanced based on a spatial-semantic dual-dimensional feature enhancement to obtain enhanced low-frequency approximate sub-band feature maps, enhanced horizontal detail sub-band feature maps, enhanced vertical detail sub-band feature maps, and enhanced diagonal detail sub-band feature maps. Specifically, Figure 5 is a flowchart of step S321 in the image analysis-based LCD system fault detection method according to an embodiment of this application. As shown in Figure 5, step S321 includes: S321-1, extracting the channel feature vector at pixel position (i,j) from the low-frequency approximate sub-band feature map as the low-frequency approximate sub-band feature channel feature vector to be enhanced; S321-2, performing n random scans on the low-frequency approximate sub-band feature map to obtain n low-frequency approximate sub-band feature channel feature vectors as a sparse set of low-frequency approximate sub-band feature reference vectors; S321-3, performing implicit compensation enhancement based on a mask control component on the sparse set of low-frequency approximate sub-band feature reference vectors and the low-frequency approximate sub-band feature channel feature vector to be enhanced to obtain the low-frequency approximate sub-band feature enhancement component implicit coding vector; S321-4, fusing the low-frequency approximate sub-band feature enhancement component implicit coding vector and the low-frequency approximate sub-band feature channel feature vector to be enhanced to obtain the enhanced low-frequency approximate sub-band feature channel feature vector, wherein the enhanced low-frequency approximate sub-band feature channel feature vector is the channel feature vector at pixel position (i,j) of the enhanced low-frequency approximate sub-band feature map.

[0042] It is understandable that although wavelet decomposition achieves multi-scale representation of defect features in sub-band feature maps of different frequency bands and directions, each sub-band still suffers from inherent interference and limitations in feature representation. In the low-frequency approximation sub-band, the global brightness gradient caused by uneven backlighting overlaps with the low-frequency characteristics of the real defect, making it difficult for traditional feature enhancement methods to suppress background interference while preserving the overall defect morphology. In the high-frequency detail sub-band, high-frequency artifacts formed by camera noise and surface reflection are easily confused with defect edge features. Furthermore, the differentiated responses of different directional sub-bands to defect morphology (e.g., horizontal sub-bands are more sensitive to vertically extending defects) require the enhancement process to consider both spatial distribution and semantic association. More importantly, the local gradient characteristics of Mura defects may only manifest as weak signals in a single sub-band, requiring cross-regional contextual information mining to form an effective representation. However, conventional enhancement methods lack the ability to explicitly model implicit semantic associations in the feature space, resulting in the enhancement process being unable to adapt to the physical characteristics of different sub-bands. To this end, this application enhances the low-frequency approximate subband feature map, the horizontal detail subband feature map, the vertical detail subband feature map, and the diagonal detail subband feature map by performing spatial-semantic dual-dimensional feature enhancement on the low-frequency approximate subband feature map, the horizontal detail subband feature map, the vertical detail subband feature map, and the diagonal detail subband feature map to obtain enhanced low-frequency approximate subband feature maps, enhanced horizontal detail subband feature maps, enhanced vertical detail subband feature maps, and enhanced diagonal detail subband feature maps.

[0043] Specifically, the processing of the low-frequency approximate sub-band feature map is used as an example for detailed explanation. First, the low-frequency approximate sub-band feature map is randomly scanned to obtain a set of reference vectors covering different brightness distribution regions. The Poincaré distance is used to measure the hierarchical association between the feature vector to be enhanced and each reference vector in hyperbolic space, constructing a spatial modulation parameter matrix to quantify the spatial correlation strength. Simultaneously, the implicit semantic association between the feature to be enhanced and the reference vectors is calculated, mining their abstract similarity in brightness gradient patterns to generate a semantic association coding matrix. Subsequently, the semantic association coding matrix is ​​used as a basic masking control component to semantically filter the reference vectors, and the spatial modulation matrix is ​​used as a secondary masking control component to implement spatial correlation modulation, forming a two-layer mask control information compensation mechanism. This process dynamically injects reference vector information that conforms to the defect brightness distribution pattern into the feature to be enhanced. For example, when a region has a slight brightness change suspected of being a defect, reference regions with similar gradient patterns are selected through semantic association, and their contribution weight is strengthened based on spatial correlation. Finally, the enhanced low-frequency approximate sub-band feature map is generated through feature fusion.

[0044] Specifically, in this embodiment of the application, step 321-1 includes: extracting the channel feature vector at the (i,j)th pixel position from the low-frequency approximate sub-band feature map as the low-frequency approximate sub-band feature channel feature vector to be enhanced, which can be expressed by the following formula:

[0045]

[0046]

[0047] in, It is a low-frequency approximate sub-band feature map. For the set of real numbers, and They are The height and width of each feature matrix along the channel dimension. yes The number of channels, yes The channel feature vector at the (i,j)th pixel position. It is the eigenvector of the low-frequency approximate sub-band feature channel to be enhanced. It is a sparse set of low-frequency approximate subband feature reference vectors.

[0048] It is understandable that extracting the channel feature vector at a specific position (i,j) from the low-frequency approximate sub-band feature map as the target for enhancement aims to focus on the brightness distribution characteristics of that region. Specifically, since the global backlight unevenness in the low-frequency sub-band overlaps with the low-frequency features of the real defect, directly processing the entire feature map may lead to the mis-enhancement of background interference. By extracting local feature vectors, the local brightness gradient pattern and background noise can be selectively separated while preserving the overall morphology of the defect (such as the gradient characteristics of the Mura defect). In other words, through localization operations, a precise target for enhancement is provided for subsequent spatial-semantic modulation, avoiding excessive smoothing or information loss in non-defect regions by global operations, while maintaining the spatial continuity of the feature map, laying the foundation for cross-regional context mining.

[0049] Specifically, in this embodiment of the application, step 321-2 includes: performing n random scans on the low-frequency approximate sub-band feature map to obtain n low-frequency approximate sub-band feature channel feature vectors as a sparse set of low-frequency approximate sub-band feature reference vectors, which can be expressed by the following formula:

[0050]

[0051] in, , , and These are the 1st, 2nd, and 3rd elements in the sparse set of low-frequency approximate subband feature reference vectors, respectively. The and the first A low-frequency approximate sub-band feature reference vector, and .

[0052] It can be understood that generating a sparse set of low-frequency approximate sub-band feature reference vectors by randomly sampling the low-frequency sub-band feature map n times essentially constructs a dynamic feature memory covering different brightness distribution patterns. Specifically, the randomness design overcomes the sensitivity of fixed sampling (such as a sliding window) to local overfitting. By introducing spatial diversity, it ensures that the reference set can cover typical brightness gradient patterns in the feature map (including potential defect areas and normal background areas). The sparse sampling strategy strikes a balance between computational efficiency and information coverage, simulating the "selective attention" mechanism of biological vision, selecting representative local patterns from global features, and avoiding interference from noise or redundant features. In other words, this step provides multi-scale, multi-location contextual reference benchmarks for subsequent modulation, enabling the enhancement process to adapt to the brightness distribution characteristics of different regions.

[0053] Specifically, Figure 6 is a flowchart of step S321-3 in the image analysis-based LCD system fault detection method according to an embodiment of this application. As shown in Figure 6, step S321-3 includes: S321-31, calculating the Poincaré distance between the feature vector of the low-frequency approximate sub-band feature channel to be enhanced and each low-frequency approximate sub-band feature reference vector in the sparse set of the low-frequency approximate sub-band feature reference vectors to obtain the low-frequency approximate sub-band feature space modulation parameter matrix; S321-32, calculating the implicit semantic association between the feature vector of the low-frequency approximate sub-band feature channel to be enhanced and each low-frequency approximate sub-band feature reference vector in the sparse set of the low-frequency approximate sub-band feature reference vectors to obtain the low-frequency approximate sub-band feature semantic association encoding. A set of matrices; S321-33, using each low-frequency approximate sub-band feature semantic association coding matrix in the set of low-frequency approximate sub-band feature semantic association coding matrices as a basic mask control component and using the low-frequency approximate sub-band feature space modulation parameter matrix as a secondary mask control component, to explicitly model and control the low-frequency approximate sub-band feature information compensation coding vector between each low-frequency approximate sub-band feature reference vector in the sparse set of low-frequency approximate sub-band feature reference vectors and the feature vector of the low-frequency approximate sub-band feature channel to be enhanced, so as to obtain the latent coding vector of the low-frequency approximate sub-band feature enhancement component.

[0054] More specifically, in the embodiments of this application, step S321-31 includes: calculating the Poincaré distance between the feature vector of the low-frequency approximate sub-band feature channel to be enhanced and each low-frequency approximate sub-band feature reference vector in the sparse set of the low-frequency approximate sub-band feature reference vectors to obtain the low-frequency approximate sub-band feature space modulation parameter matrix, which can be expressed by the following formula:

[0055]

[0056]

[0057] in, To calculate the square of the Euclidean norm of a vector It is an inverse hyperbolic cosine function. for and The Poincaré distance between them These are the eigenvalues ​​in the modulation parameter matrix of the low-frequency approximate subband feature space. It is the low-frequency approximate subband characteristic space modulation parameter matrix.

[0058] It is understandable that using the Poincaré distance to measure the hierarchical relationship between the feature vectors of the low-frequency approximate sub-band feature channels to be enhanced and the low-frequency approximate sub-band feature reference vectors in hyperbolic space is better at capturing tree-like or hierarchical data distribution characteristics than Euclidean distance. Specifically, the brightness gradation patterns in the low-frequency sub-band (such as the intensity changes between defects and the background) often exhibit non-linear hierarchical relationships (e.g., strong gradation from the center to weak gradation at the edges). The Poincaré distance, through the geometric properties of hyperbolic space, models such relationships as spatial correlation strength. The generated low-frequency approximate sub-band feature spatial modulation parameter matrix not only quantifies the spatial similarity between the region to be enhanced and the reference region, but also implicitly encodes the brightness distribution patterns across regions (e.g., the difference between the local gradation trend of defects and the global changes in the background), providing a basis for differentiated weight allocation for subsequent spatial modulation, thereby enhancing defects while suppressing interference from irrelevant regions.

[0059] More specifically, in the embodiments of this application, steps S321-32 include: calculating the implicit semantic association between the feature vector of the low-frequency approximate sub-band feature channel to be enhanced and each low-frequency approximate sub-band feature reference vector in the sparse set of the low-frequency approximate sub-band feature reference vectors to obtain a set of low-frequency approximate sub-band feature semantic association coding matrices, which can be expressed by the following formula:

[0060]

[0061]

[0062] in, For matrix multiplication, for The corresponding weight matrix, It is a transpose operation. yes and The length of the vectors after multiplication yes function, yes and The semantic association coding matrix of low-frequency approximate sub-band features between, that is, the set of semantic association coding matrices of low-frequency approximate sub-band features. A low-frequency approximate sub-band feature semantic association coding matrix It is a set of semantic association coding matrices for low-frequency approximate sub-band features. , and These are the 1st, 2nd, and 3rd elements in the set of low-frequency approximate subband feature semantic association coding matrices, respectively. A low-frequency approximate sub-band feature semantic association coding matrix.

[0063] It can be understood that by mining the abstract semantic similarity (such as the commonalities and differences in brightness gradient patterns) between the feature vectors of the low-frequency approximate sub-band feature channels to be enhanced and the reference vectors of the low-frequency approximate sub-band features, a semantic association coding matrix for low-frequency approximate sub-band features is generated. This step goes beyond pixel-level surface feature matching, focusing on higher-order semantic patterns (such as "center-edge gradient" and "multi-scale texture similarity"), which can effectively solve the problem of misassociation caused by traditional methods relying on explicit feature similarity. For example, when there are weak defects in the region to be enhanced, semantic association can filter out reference regions with similar gradient directions but different intensities, even if they are spatially far apart. That is, the obtained semantic association coding matrix for low-frequency approximate sub-band features serves as a basic mask, realizing semantic-level filtering of reference vectors, ensuring that subsequent information compensation only injects feature components strongly correlated with the defect pattern, avoiding interference from irrelevant background information.

[0064] More specifically, in the embodiments of this application, steps S321-33 include: calculating the positional difference vector between each low-frequency approximate sub-band feature reference vector and the feature vector of the low-frequency approximate sub-band feature channel to be enhanced to obtain a set of low-frequency approximate sub-band feature information compensation coding vectors. This process can be expressed as:

[0065]

[0066] in, yes and The low-frequency approximate sub-band feature information compensation encoding vector between them;

[0067] Based on the basic masking control component and the secondary masking control component, mask-compensation optimization based on dual-field cooperative modulation is performed on each low-frequency approximate sub-band feature information compensation coding vector in the set of low-frequency approximate sub-band feature information compensation coding vectors to obtain a set of low-frequency approximate sub-band feature information compensation optimized coding vectors. This process can be expressed as:

[0068]

[0069]

[0070] in, yes The low-frequency approximate sub-band feature information is used to compensate the modulation coding vector after modulation. yes Optimized low-frequency approximate sub-band feature information compensation optimization coding vector;

[0071] Based on the basic masking control component and the secondary masking control component, a global explicit modeling modulation aggregation is performed on the set of low-frequency approximate subband feature information compensation optimized coding vectors to obtain the low-frequency approximate subband feature enhancement component implicit coding vector. The above process can be expressed as a whole by the following formula:

[0072]

[0073] in, It is the number of vectors in the sparse set of low-frequency approximate subband feature reference vectors. It is the latent coding vector of the low-frequency approximate subband feature enhancement component.

[0074] It can be understood that a hierarchical modulation mechanism is formed by using the semantic association coding matrix of low-frequency approximate sub-band features as the primary mask and the spatial modulation parameter matrix of low-frequency approximate sub-band features as the secondary mask. Specifically, the primary mask selects low-frequency approximate sub-band feature reference vectors that match the defect pattern based on semantic similarity (such as regions with similar gradient directions), while the secondary mask strengthens the contribution weight of low-frequency approximate sub-band feature reference vectors in the local neighborhood based on spatial correlation. For example, for a suspected defect region, the semantic mask prioritizes reference vectors with similar gradient patterns, while the spatial mask assigns higher weights to neighboring regions, thereby dynamically fusing cross-regional contextual information. This explicit modeling strategy breaks through the dependence of traditional enhancement methods on implicit feature learning. Through an interpretable two-layer control mechanism, it accurately injects enhancement components that conform to the physical characteristics of the defect (such as supplementing weak gradient signals), while suppressing pseudo-enhancement caused by noise or background unevenness.

[0075] Specifically, in the information coding framework for low-frequency approximate subband feature enhancement, the implementation of the field density modulation configuration relies on the synergistic effect of the low-frequency approximate subband feature semantic association coding matrix and the low-frequency approximate subband feature spatial modulation parameter matrix. Specifically, the low-frequency approximate subband feature semantic association coding matrix, as the standard field, encodes the implicit semantic associations between feature vectors (such as the abstract similarity of brightness gradient patterns), which is achieved through… Maintaining transformation invariance constraints. This operation essentially projects the information compensation component under semantic masking (i.e., the difference between the feature to be enhanced and the reference vector) into the geometric measure space of the standard field, ensuring the enhancement process's robustness to a globally varying brightness background. The low-frequency approximate sub-band feature space modulation parameter matrix serves as a covariant adjustment field, dynamically adjusting the local geometry of the standard field through the spatial correlation strength quantized by the Poincaré distance (i.e., the hierarchical brightness distribution pattern). Based on... The covariant adjustment mechanism applies spatial modulation parameters to the curved spatial topology, enabling the geometric measure of the derivative hierarchy of the standard field to adapt to the spatial distribution characteristics of the feature map. For example, in the defect edge region, the covariant adjustment field enhances the contribution weight of neighboring reference vectors, correcting the deviation of the standard field in local curvature, thereby improving the spatial continuity of the defect gradient signal. In other words, the strength coupling of the inter-field interaction is quantified through conventional partial differential equations, the core of which lies in establishing the differential manifold relationship between semantic association and spatial modulation. Specifically, the partial differential coupling terms of the standard field and the covariant adjustment field reflect the joint gradient direction of semantic filtering and spatial enhancement, allowing the low-frequency approximate subband feature information compensation optimization encoding vector to simultaneously satisfy semantic consistency and spatial smoothness constraints. This coupling mechanism, while suppressing high-frequency noise interference (such as artifacts caused by camera noise), enhances the characterization ability of low-frequency subband features for weak defects through cross-regional contextual information mining (such as the long-range correlation between the local gradient pattern of the defect and the reference vector).

[0076] Specifically, in this embodiment, step S321-4 is used to: fuse the latent coding vector of the low-frequency approximate subband feature enhancement component and the feature vector of the low-frequency approximate subband feature channel to be enhanced to obtain the enhanced low-frequency approximate subband feature channel feature vector, wherein the enhanced low-frequency approximate subband feature channel feature vector is the channel feature vector of the pixel position (i,j) of the enhanced low-frequency approximate subband feature map, which can be expressed by the following formula:

[0077]

[0078] in, and It is a weighted hyperparameter. yes The enhanced low-frequency approximation subband feature channel feature vector is the channel feature vector at the (i,j)th pixel position of the enhanced low-frequency approximation subband feature map.

[0079] It is understandable that fusing the latent coding vector of the low-frequency approximate sub-band feature enhancement component with the feature vector of the low-frequency approximate sub-band feature channel to be enhanced can achieve synergistic optimization of local feature representation and global contextual information. Specifically, the feature vector of the original low-frequency approximate sub-band feature channel to be enhanced retains the initial shape and low-frequency characteristics of the defect, while the latent coding vector of the low-frequency approximate sub-band feature enhancement component supplements the weak signals lost due to background interference or noise suppression (such as the local gradient details of the Mura defect) through cross-regional semantic association. The fusion process is not a simple weighted superposition, but rather a reconstruction of the feature space distribution through nonlinear mapping, so that the enhanced features are both sensitive to small defects and robust to background changes. For example, in areas with uneven backlighting, the fused features can highlight the continuous gradient pattern of the defect while weakening the interference of global brightness changes, ultimately improving the ability of the low-frequency sub-band features to distinguish between defects and the background.

[0080] In step S322, the enhanced low-frequency approximation sub-band feature map, the enhanced horizontal detail sub-band feature map, the enhanced vertical detail sub-band feature map, and the enhanced diagonal detail sub-band feature map are fused to obtain a multi-scale fused feature map of the LCD panel image. Correspondingly, considering that although sub-band feature maps of different frequency bands and directions enhance the defect features of specific dimensions through enhancement processing, their isolated analysis still has the limitation of information fragmentation. While the enhanced low-frequency approximation sub-band feature map can characterize the overall brightness distribution of defects, it loses edge detail information; and while the high-frequency detail enhanced sub-band feature maps in the horizontal, vertical, and diagonal directions capture local gradient changes, they lack contextual awareness of the global morphology of defects. Therefore, in order to fully utilize the advantages of each sub-band, integrate the global and local information of the image, and more comprehensively describe the features of the LCD panel image, this application obtains a multi-scale fused feature map of the LCD panel image by fusing the enhanced low-frequency approximation sub-band feature map, the enhanced horizontal detail sub-band feature map, the enhanced vertical detail sub-band feature map, and the enhanced diagonal detail sub-band feature map. In this way, the resulting multi-scale fused feature map of the LCD panel image can provide a comprehensive feature representation for subsequent defect detection and analysis. It contains both the overall structural information of the image and detailed texture information, organically combining features of different scales and directions. This allows the model to analyze and understand the image from multiple perspectives, thereby more accurately locating and describing the location, shape, size, and texture of Mura defects, providing a more sufficient basis for defect classification and evaluation.

[0081] In step S4, statistical analysis based on Mura segmentation regions is performed on the multi-scale fusion feature map of the LCD panel image to obtain the Mura region statistical feature encoding vector. Specifically, Figure 7 is a flowchart of step S4 in the LCD system fault detection method based on image analysis according to an embodiment of this application. As shown in Figure 7, step S4 includes: S41, performing image semantic segmentation on the multi-scale fusion feature map of the LCD panel image to obtain the Mura segmentation region image; S42, extracting the statistical features of the Mura segmentation region image to obtain the Mura region statistical feature encoding vector.

[0082] In step S41, image semantic segmentation is performed on the multi-scale fusion feature map of the LCD panel image to obtain a Mura segmentation region image. Correspondingly, considering that although the multi-scale fusion feature map of the LCD panel image integrates global brightness distribution and multi-directional edge details, the problem of accurate boundary delineation of defect areas still needs to be solved. Image semantic segmentation can classify each pixel in the image based on the information in the feature map, determine whether it belongs to the Mura defect area, and thus accurately delineate the boundary and range of the Mura defect, providing accurate location information for subsequent defect analysis and processing. Based on this, in the technical solution of this application, image semantic segmentation is performed on the multi-scale fusion feature map of the LCD panel image to obtain a Mura segmentation region image. That is, the Mura segmentation region image obtained through image semantic segmentation can accurately identify the Mura defect area on the LCD panel. Classifying each pixel in the image and marking the pixels belonging to the Mura defect forms a clear segmentation region, making the Mura defect clearly visible in the image, facilitating direct location and observation of the defect by operators or subsequent analysis algorithms, and providing a basis for defect evaluation and repair.

[0083] Specifically, firstly, a multi-scale fusion feature map of the LCD panel image is used as input data. This feature map has already integrated multi-scale information of low-frequency brightness distribution and high-frequency edge details through previous steps. To ensure that the feature values ​​of each channel are in a uniform dimension, the input feature map needs to be normalized, for example, by using a Min-Max scaling method, to avoid subsequent segmentation deviations due to differences in numerical ranges.

[0084] In the semantic segmentation network architecture design phase, an encoder-decoder structure network model, such as U-Net or DeepLabv3+, is typically adopted to balance the fusion of spatial details and high-level semantic information. The encoder part progressively downsamples through convolution and pooling operations to extract high-level semantic features, including the texture patterns and brightness distribution of Mura defects. The decoder part restores spatial resolution through transposed convolution or interpolation upsampling and combines this with the skip connection features from the encoder to refine the segmentation boundaries. To further improve segmentation accuracy, spatial or channel attention mechanisms, such as the CBAM module, can be introduced in the decoding stage to dynamically weight key regions related to Mura defects in multi-scale feature maps while suppressing background noise interference.

[0085] During the training phase, labeled data is needed as supervision signals. Typically, manually labeled or simulated binary masks of Mura defects are used, where defect regions are labeled as 1 and the background as 0. The loss function design must consider the class imbalance problem, where Mura defects usually account for a small proportion. Therefore, the main loss function can be Dice Loss or Focal Loss to enhance the model's focus on defect regions. Furthermore, boundary-aware loss, such as Boundary Loss, can be introduced to strengthen the network's ability to recognize the gradual transition regions at the edges of Mura defects.

[0086] During the inference phase, the preprocessed multi-scale fused feature map of the LCD panel image is input into the trained semantic segmentation network, outputting a probability map of each pixel belonging to a Mura defect, with values ​​ranging from 0 to 1. Subsequently, the probability map is converted into a binary segmentation mask through thresholding. Adaptive thresholding methods (such as the Otsu algorithm) or a fixed threshold (such as 0.5) can be used for binarization to generate a preliminary Mura segmentation region image. To optimize the segmentation results, morphological post-processing is required, including using opening operations to eliminate isolated noise points and using connected component analysis to retain regions with areas larger than a preset threshold to filter out minor false defects.

[0087] Finally, the optimized binary mask is aligned with the original image to ensure accurate matching of the defect region boundaries, outputting a Mura segmentation region image, where the defect region is represented in white and the background in black. To verify the segmentation accuracy, metrics such as IoU (Intersection over Union) or pixel-level accuracy can be used to evaluate the consistency between the segmentation results and the actual annotations, ensuring that it meets the accuracy requirements of industrial inspection.

[0088] In step S42, statistical features of the Mura segmentation region image are extracted to obtain the statistical feature encoding vector of the Mura region. Accordingly, while semantic segmentation can accurately locate defect regions, simply relying on the segmentation results cannot meet the needs of defect classification and quantitative evaluation. If defect evaluation is performed directly based on pixel-level brightness values ​​or simple morphological parameters, it is difficult to characterize the complex visual characteristics of defects. For example, the human eye's perception of Mura defects depends not only on the absolute value of brightness differences but also on abstract features such as distribution uniformity and texture coherence. Furthermore, the distribution patterns of different Mura defects (such as dot-like, line-like, and cloud-like defects) in the statistical feature space differ significantly, but existing methods lack systematic modeling of multi-dimensional features, making it difficult for classification models to establish a mapping relationship between defect morphology and quality level. Therefore, this application extracts statistical features of the Mura segmentation region image to obtain the statistical feature encoding vector of the Mura region. Specifically, in a concrete example of this application, for the segmented binary mask region, the brightness distribution features in the original enhanced image within its coverage area are first extracted, including indicators such as gray-level mean, variance, and gradient histogram entropy, to characterize the overall contrast and local fluctuation characteristics of the defect. Simultaneously, the morphological features of the defect region, such as area-to-perimeter ratio, compactness, and Fourier descriptors, are calculated to quantify its geometric distribution pattern. Furthermore, a texture analysis algorithm is used to extract the contrast and correlation parameters of the gray-level co-occurrence matrix, capturing the microscopic texture characteristics of the defect. These features, after normalization and feature selection, are fused into a fixed-dimensional statistical feature encoding vector according to preset weights, forming a multi-dimensional defect fingerprint that takes into account brightness, morphology, and texture.

[0089] In step S5, the Mura defect identification result is obtained based on the statistical feature encoding vector of the Mura region. Specifically, in this embodiment, step S5 includes: inputting the statistical feature encoding vector of the Mura region into a defect classifier based on an SVM model to obtain the Mura defect identification result. In particular, the Mura defect identification result is divided into "no defect," "minor defect," "moderate defect," and "serious defect" levels. It should be understood that Support Vector Machine (SVM) is a supervised machine learning algorithm that performs well in pattern recognition and classification tasks. It has good performance for processing a limited amount of high-dimensional data, and the statistical feature encoding vector of the Mura region is precisely the high-dimensional data representation obtained after feature extraction from the Mura segmentation region image. SVM can find an optimal hyperplane in high-dimensional space to separate data points of different categories, and has effective processing methods for both linearly separable and non-linearly separable data (mapping the data to a higher-dimensional space through a kernel function to make it linearly separable). In the Mura defect classification task, the statistical feature encoding vectors corresponding to different types of Mura defects can be regarded as data points of different categories, and SVM can use the features of these data points to construct a classification model. By analyzing and judging the input feature vectors using an SVM model, and classifying Mura defects into corresponding categories based on a pre-trained classification model, such as different types of Mura defects corresponding to different shapes, sizes, and grayscale distributions, this helps production personnel understand the specifics of the defects so that they can take targeted measures to repair or improve the production process.

[0090] In summary, the image analysis-based LCD system fault detection method based on the embodiments of this application is explained. First, an LCD panel image is acquired using a high-resolution industrial camera. Then, homomorphic filtering and CLAHE contrast enhancement processing are applied to the image to obtain an enhanced LCD panel image. Next, multi-scale feature enhancement fusion is performed on the enhanced image to obtain a multi-scale fused feature map of the LCD panel image. Then, statistical analysis is performed on this feature map based on Mura segmentation regions to obtain statistical feature encoding vectors for the Mura regions. Finally, based on these encoding vectors, the Mura defect identification result is obtained. This method can more accurately capture the features of Mura defects, reduce misjudgments and missed detections caused by interference factors such as light sources and noise, improve the detection accuracy of Mura defects, and ensure the stability of the detection results by adapting to different production conditions.

[0091] Figure 8 is a system block diagram of an LCD system fault detection system based on image analysis according to an embodiment of this application. As shown in Figure 8, the LCD system fault detection system 100 based on image analysis according to an embodiment of this application includes: an image acquisition module 110 for acquiring an LCD panel image acquired by a high-resolution industrial camera; an image preprocessing module 120 for performing homomorphic filtering and CLAHE contrast enhancement on the LCD panel image to obtain an enhanced LCD panel image; an image encoding module 130 for performing multi-scale feature enhancement fusion on the enhanced LCD panel image to obtain a multi-scale fusion feature map of the LCD panel image; a Mura region analysis module 140 for performing statistical analysis based on Mura segmentation regions on the multi-scale fusion feature map of the LCD panel image to obtain a Mura region statistical feature encoding vector; and a defect identification module 150 for obtaining a Mura defect identification result based on the Mura region statistical feature encoding vector.

[0092] Here, those skilled in the art will understand that the specific functions and operations of each unit and module in the above-described image analysis-based LCD system fault detection system 100 have been described in detail in the description of the image analysis-based LCD system fault detection method with reference to Figures 1 to 7, and therefore, their repeated description will be omitted.

[0093] In summary, the image analysis-based LCD system fault detection system 100 based on the embodiments of this application is explained. It first acquires an LCD panel image captured by a high-resolution industrial camera, then performs homomorphic filtering and CLAHE contrast enhancement processing on the image to obtain an enhanced LCD panel image. Next, it performs multi-scale feature enhancement fusion on the enhanced image to obtain a multi-scale fused feature map of the LCD panel image. Then, it performs statistical analysis on the feature map based on Mura segmentation regions to obtain a statistical feature encoding vector for the Mura region. Finally, based on this encoding vector, it obtains the Mura defect identification result. This approach can more accurately capture the features of Mura defects, reduce misjudgments and missed detections caused by interference factors such as light sources and noise, improve the detection accuracy of Mura defects, and ensure the stability of the detection results by adapting to different production conditions.

Claims

1. A fault detection method for an LCD system based on image analysis, characterized in that, include: Acquire images of LCD panels captured by a high-resolution industrial camera; The LCD panel image is subjected to homomorphic filtering and CLAHE contrast enhancement to obtain an enhanced LCD panel image; The enhanced LCD panel image is subjected to multi-scale feature enhancement and fusion to obtain a multi-scale fused feature map of the LCD panel image, including: performing wavelet decomposition and feature extraction on the enhanced LCD panel image to obtain a low-frequency approximate sub-band feature map, a horizontal detail sub-band feature map, a vertical detail sub-band feature map, and a diagonal detail sub-band feature map; and performing spatial-semantic dual-dimensional feature enhancement and fusion processing on the low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map to obtain the multi-scale fused feature map of the LCD panel image. Statistical analysis based on Mura segmentation regions is performed on the multi-scale fusion feature map of the LCD panel image to obtain the statistical feature encoding vector of the Mura region; Based on the statistical feature encoding vector of the Mura region, the Mura defect identification result is obtained.

2. The LCD system fault detection method based on image analysis according to claim 1, characterized in that, Wavelet decomposition and feature extraction are performed on the enhanced LCD panel image to obtain low-frequency approximate sub-band feature maps, horizontal detail sub-band feature maps, vertical detail sub-band feature maps, and diagonal detail sub-band feature maps, including: Wavelet decomposition is performed on the enhanced LCD panel image to obtain a low-frequency approximate subband matrix, a horizontal detail subband matrix, a vertical detail subband matrix, and a diagonal detail subband matrix. Feature extraction based on dilated convolutional coding is performed on the low-frequency approximate subband matrix, the horizontal detail subband matrix, the vertical detail subband matrix, and the diagonal detail subband matrix to obtain the low-frequency approximate subband feature map, the horizontal detail subband feature map, the vertical detail subband feature map, and the diagonal detail subband feature map.

3. The LCD system fault detection method based on image analysis according to claim 1, characterized in that, The low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map are subjected to spatial-semantic dual-dimensional feature enhancement fusion processing to obtain the multi-scale fusion feature map of the LCD panel image, including: The low-frequency approximate sub-band feature map, the horizontal direction detail sub-band feature map, the vertical direction detail sub-band feature map, and the diagonal direction detail sub-band feature map are subjected to spatial-semantic dual-dimensional feature enhancement to obtain enhanced low-frequency approximate sub-band feature map, enhanced horizontal direction detail sub-band feature map, enhanced vertical direction detail sub-band feature map, and enhanced diagonal direction detail sub-band feature map; The enhanced low-frequency approximate sub-band feature map, the enhanced horizontal detail sub-band feature map, the enhanced vertical detail sub-band feature map, and the enhanced diagonal detail sub-band feature map are fused to obtain the multi-scale fused feature map of the LCD panel image.

4. The LCD system fault detection method based on image analysis according to claim 3, characterized in that, The low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map are subjected to spatial-semantic dual-dimensional feature enhancement to obtain enhanced low-frequency approximate sub-band feature maps, enhanced horizontal detail sub-band feature maps, enhanced vertical detail sub-band feature maps, and enhanced diagonal detail sub-band feature maps, including: Extract the channel feature vector at the (i,j)th pixel position from the low-frequency approximate sub-band feature map as the low-frequency approximate sub-band feature channel feature vector to be enhanced; The low-frequency approximate sub-band feature map is randomly scanned n times to obtain n low-frequency approximate sub-band feature channel feature vectors as a sparse set of low-frequency approximate sub-band feature reference vectors; The sparse set of the low-frequency approximate subband feature reference vector and the feature vector of the low-frequency approximate subband feature channel to be enhanced are implicitly compensated and enhanced based on the mask control component to obtain the implicit coding vector of the low-frequency approximate subband feature enhancement component. The latent coding vector of the low-frequency approximate subband feature enhancement component and the feature vector of the low-frequency approximate subband feature channel to be enhanced are fused to obtain the enhanced low-frequency approximate subband feature channel feature vector, wherein the enhanced low-frequency approximate subband feature channel feature vector is the channel feature vector of the pixel position (i,j) of the enhanced low-frequency approximate subband feature map.

5. The LCD system fault detection method based on image analysis according to claim 4, characterized in that, The sparse set of the low-frequency approximate sub-band feature reference vectors and the feature vectors of the low-frequency approximate sub-band feature channels to be enhanced are subjected to implicit compensation enhancement based on a mask modulation component to obtain the implicit coding vector of the low-frequency approximate sub-band feature enhancement component, including: Calculate the Poincaré distance between the feature vector of the low-frequency approximate subband feature channel to be enhanced and each low-frequency approximate subband feature reference vector in the sparse set of the low-frequency approximate subband feature reference vector to obtain the low-frequency approximate subband feature space modulation parameter matrix; The implicit semantic association between the feature vector of the low-frequency approximate sub-band feature channel to be enhanced and each low-frequency approximate sub-band feature reference vector in the sparse set of the low-frequency approximate sub-band feature reference vector is calculated to obtain a set of low-frequency approximate sub-band feature semantic association coding matrices. Using each low-frequency approximate subband feature semantic association coding matrix in the set of low-frequency approximate subband feature semantic association coding matrices as the basic mask control component and the low-frequency approximate subband feature space modulation parameter matrix as the secondary mask control component, explicit modeling and control are performed on the low-frequency approximate subband feature information compensation coding vector between each low-frequency approximate subband feature reference vector in the sparse set of low-frequency approximate subband feature reference vectors and the feature vector of the low-frequency approximate subband feature channel to be enhanced, so as to obtain the latent coding vector of the low-frequency approximate subband feature enhancement component.

6. The LCD system fault detection method based on image analysis according to claim 5, characterized in that, Using each low-frequency approximate sub-band feature semantic association coding matrix in the set of low-frequency approximate sub-band feature semantic association coding matrices as a basic masking control component and the low-frequency approximate sub-band feature spatial modulation parameter matrix as a secondary masking control component, explicit modeling and control are performed on the low-frequency approximate sub-band feature information compensation coding vector between each low-frequency approximate sub-band feature reference vector in the sparse set of low-frequency approximate sub-band feature reference vectors and the feature vector of the low-frequency approximate sub-band feature channel to be enhanced, to obtain the implicit coding vector of the low-frequency approximate sub-band feature enhancement component, including: Calculate the positional difference vector between each low-frequency approximate sub-band feature reference vector and the feature vector of the low-frequency approximate sub-band feature channel to be enhanced to obtain a set of low-frequency approximate sub-band feature information compensation coding vectors; Based on the basic mask control component and the secondary mask control component, mask-compensation optimization based on dual-field cooperative modulation is performed on each low-frequency approximate sub-band feature information compensation coding vector in the set of low-frequency approximate sub-band feature information compensation coding vectors to obtain a set of low-frequency approximate sub-band feature information compensation optimized coding vectors. Based on the basic masking control component and the secondary masking control component, the set of low-frequency approximate subband feature information compensation optimized coding vectors is subjected to global explicit modeling modulation aggregation to obtain the low-frequency approximate subband feature enhancement component implicit coding vector.

7. The LCD system fault detection method based on image analysis according to claim 6, characterized in that, Statistical analysis based on Mura segmentation regions is performed on the multi-scale fusion feature map of the LCD panel image to obtain the statistical feature encoding vector of the Mura region, including: Image semantic segmentation is performed on the multi-scale fusion feature map of the LCD panel image to obtain the Mura segmentation region image; Statistical features of the Mura segmentation region image are extracted to obtain the statistical feature encoding vector of the Mura region.

8. The LCD system fault detection method based on image analysis according to claim 7, characterized in that, Based on the statistical feature encoding vector of the Mura region, the Mura defect identification result is obtained, including: inputting the statistical feature encoding vector of the Mura region into a defect classifier based on an SVM model to obtain the Mura defect identification result.

9. A fault detection system for an LCD system based on image analysis, characterized in that, include: Image acquisition module, used to acquire LCD panel images captured by a high-resolution industrial camera; An image preprocessing module is used to perform homomorphic filtering and CLAHE contrast enhancement on the LCD panel image to obtain an enhanced LCD panel image. An image encoding module is used to perform multi-scale feature enhancement and fusion on the enhanced LCD panel image to obtain a multi-scale fused feature map of the LCD panel image. The image encoding module is used to: perform wavelet decomposition and feature extraction on the enhanced LCD panel image to obtain a low-frequency approximate sub-band feature map, a horizontal detail sub-band feature map, a vertical detail sub-band feature map, and a diagonal detail sub-band feature map; and perform spatial-semantic dual-dimensional feature enhancement and fusion processing on the low-frequency approximate sub-band feature map, the horizontal detail sub-band feature map, the vertical detail sub-band feature map, and the diagonal detail sub-band feature map to obtain the multi-scale fused feature map of the LCD panel image. The Mura region analysis module is used to perform statistical analysis on the multi-scale fusion feature map of the LCD panel image based on Mura segmentation regions to obtain the Mura region statistical feature encoding vector. The defect identification module is used to obtain the Mura defect identification result based on the statistical feature encoding vector of the Mura region.

10. The LCD system fault detection system based on image analysis according to claim 9, characterized in that, The defect identification module is used to input the statistical feature encoding vector of the Mura region into a defect classifier based on an SVM model to obtain the Mura defect identification result.