TFT panel defect identification method, computer program product and system
By using multi-state light-based image acquisition and an improved U-NET architecture, combined with image preprocessing and feature extraction techniques, the accuracy fluctuation problem of the automated inspection system for TFT-LCD panels when identifying complex defects such as Mura was solved, achieving efficient and accurate defect identification.
Patent Information
- Application Number
- CN202510960009.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-11
- Publication Date
- 2025-11-21
AI Technical Summary
Existing automated inspection systems for TFT-LCD panels exhibit significant accuracy fluctuations when identifying complex defect types such as Mura, and traditional manual visual inspection is inefficient and susceptible to human factors, making it difficult to identify minute defects.
By employing image acquisition under various lighting conditions, combined with an improved U-NET architecture and a lightweight defect classification model, and through visual acquisition, image preprocessing, defect classification, and feature extraction, and utilizing techniques such as Gabor filtering and B-spline surface fitting, point defects, line defects, and Mura-type defects are identified, thereby enhancing detection accuracy.
It improves the comprehensiveness and accuracy of defect detection in TFT-LCD panels, reduces human error, enhances the ability to identify complex defect patterns, especially the detection accuracy of Mura-type defects, and solves the problem of large fluctuations in detection accuracy.
Smart Images
Figure CN120997131A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of TFT panel detection, and in particular to a TFT panel defect identification method, a computer program product and a TFT panel defect identification system. BACKGROUND
[0002] With the rapid development of global information industry, the update and replacement of products in the consumer electronics field, especially smart phones, tablet computers, televisions and the like, have brought about rapid development of the TFT-LCD panel industry. The demand for TFT-LCD panels in emerging fields such as vehicle-mounted displays, industrial control and medical equipment is also rising. As a key component of modern electronic devices, the quality of TFT-LCD panels directly affects the performance and user experience of terminal products. However, various defects can easily occur in TFT-LCD panels during production, so related detection must be performed before shipment to ensure product quality.
[0003] Traditional TFT-LCD panel detection usually relies on manual visual inspection, which is inefficient and susceptible to human factors, leading to missed detection or misjudgment. Moreover, manual visual inspection cannot accurately identify defects that are too small or even pixel-level. With the development of TFT-LCD panels towards high resolution and flexible display, traditional manual visual inspection methods cannot meet higher quality standards.
[0004] Currently, automatic detection systems centered on machine vision technology have gradually become the main means of TFT-LCD panel detection. Automatic detection systems use high-precision image acquisition equipment and advanced image processing algorithms to achieve non-contact, fast and accurate defect detection of TFT-LCD panels. Through feature extraction and intelligent analysis, automatic detection systems can quickly scan TFT-LCD panels and identify most obvious physical defects such as scratches, foreign matter contamination and line abnormalities, thereby improving detection efficiency and accuracy, and effectively reducing the misjudgment rate and missed detection rate caused by manual detection. However, for defects such as Mura (luminance unevenness) that require more complex image processing methods for identification, and for different defects that have large differences in recognition rate when different image processing methods are used, existing automatic detection systems have the problem of large detection precision fluctuations. SUMMARY
[0005] The technical problem to be solved by the present application is to provide a TFT panel defect identification method, a computer program product storing a computer program that implements the steps of the method when executed, and a TFT panel defect identification system, which can accurately identify different types of panel defects.
[0006] In order to solve the above technical problems, the first aspect of the present application provides a TFT panel defect recognition method, comprising the following steps:
[0007] A visual acquisition step is performed to acquire images of the TFT-LCD panel under multiple state lights;
[0008] A defect classification step is performed to call a trained defect classification model to identify the defect type of the acquired image, wherein the defect type includes multiple types of point defects, line defects and Mura defects; if a Mura defect is identified, an image enhancement step is performed, wherein the image enhancement step includes background suppression and B-spline surface fitting to segment the Mura area in the image;
[0009] A defect extraction step is performed to extract defect features of the corresponding type for each defect type image, and the extracted defect features are output as a defect set of the current TFT-LCD panel.
[0010] Further, an image matching step is performed after the defect classification step to identify the corresponding relationship between the state light and the defect type of the acquired image; if the corresponding relationship does not meet the preset corresponding relationship, the defect classification step is performed again until the preset corresponding relationship is met or the defect classification step is performed for a preset number of times.
[0011] Further, the state light includes R, G, B, White, Black, Gray and H; the defect corresponding to the R / G / B state light includes a point defect or a line defect; the defect corresponding to the White / Black state light includes a large-area defect or a contrast abnormality defect; and the defect corresponding to the Gray / H state light includes a Mura defect.
[0012] Further, an image preprocessing step is performed before the defect classification step, comprising the following steps:
[0013] A weighted template difference image processing step is performed to select a preset small size image at the background position of the acquired image as an ideal texture template, divide the acquired image into a plurality of single units of the same size, subtract each unit from the ideal texture template after weighted processing, and then synthesize all units to remove the texture background;
[0014] A bilateral filter denoising step is performed to calculate the weight of all pixels in the neighborhood based on the spatial distance and color difference for all pixels processed by the weighted template difference image processing step, and to calculate the weighted average of the pixel values in the neighborhood according to the calculated weight, so as to obtain the pixel neighborhood as the filtered pixel value in the neighborhood;
[0015] The maximum entropy segmentation step calculates foreground and background entropy values after threshold segmentation of the image processed by the bilateral filter denoising step, and selects a threshold with the maximum total entropy value as the threshold standard for segmenting the foreground and the background.
[0016] Further, the preset small size image of the background position of the acquired image in the weighted template difference image processing step refers to an image with a size of l x k at the image marker position.
[0017] Further, the method further comprises a Gabor filter texture enhancement step or a brightness equalization step executed after the maximum entropy segmentation step, and specifically: if the state light of the current image is R, G or B light, the Gabor filter texture enhancement step is executed to enhance the visibility of a single liquid crystal tube, so as to extract texture information in different directions and sizes; if the state light of the current image is White, Black, Gray or H light, the brightness equalization step is adopted to adjust histogram equalization, so as to enhance the overall features of the image and improve the defect contrast.
[0018] Further, the background suppression in the image enhancement step refers to contrast enhancement processing of the extracted background image.
[0019] Further, the B-spline surface fitting refers to step-by-step solving of a bicubic B-spline surface function by a product algorithm to perform image smoothing processing.
[0020] The second aspect also provides a computer program product storing a computer program, which can realize the steps of the TFT panel defect identification method when executed by a processor.
[0021] The third aspect also provides a TFT panel defect identification system, which comprises a processor and a controller, an image taking lens and a plurality of state light sources connected to the controller respectively, the plurality of state light sources are switched to align with the TFT-LCD panel under the control of the controller, and the computer program product is also provided, and the computer program on the computer program product can be executed by the processor.
[0022] The TFT panel defect identification system provided by the application executes the steps of the TFT panel defect identification method, acquires images of the TFT-LCD panel under a plurality of state lights, more comprehensively captures different types of defects, and thus improves the comprehensiveness and accuracy of detection; adopts a defect classification model to automatically identify and classify a plurality of types of defects such as point defects, line defects and Mura defects, improves the detection speed, enhances the recognition ability of complex defect patterns, reduces errors caused by human factors, and separately performs background suppression and B-spline surface fitting on Mura defects, effectively highlights these defect regions, improves the detection accuracy of Mura defects, and effectively solves the problem of large fluctuation of detection accuracy of the automatic detection system. BRIEF DESCRIPTION OF DRAWINGS
[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced.
[0024] Figure 1 is a structural schematic diagram of a TFT panel defect recognition system;
[0025] Figure 2 is a flowchart of a TFT panel defect recognition method. DETAILED DESCRIPTION
[0026] The present application will be further described in detail below in combination with specific embodiments.
[0027] The TFT panel defect recognition system of the present embodiment, as shown in Figure 1 , includes a processor and a controller, and an image taking lens and a plurality of state light sources connected to the controller respectively. The plurality of state light sources include R, G, B, White, Black, Gray, and H seven state lights. These state lights are switched under the control of the controller to align the TFT-LCD panel. The image taking lens acquires images of the TFT-LCD panel under the seven state lights under the control of the controller, so as to more comprehensively capture different types of defects. The TFT panel defect recognition system of the present embodiment performs image preprocessing, defect detection, and classification on these images by executing the TFT panel defect recognition method of the present embodiment, automatically identifies and classifies a plurality of types of defects such as point defects, line defects, and Mura type defects by using a defect classification model, improves the detection speed, enhances the recognition ability for complex defect patterns, reduces errors caused by human factors, and separately performs background suppression and B-spline surface fitting for Mura type defects, effectively highlights these defect regions, improves the detection accuracy for Mura defects, and effectively solves the problem of large fluctuation of detection accuracy of the automatic detection system.
[0028] The TFT panel defect recognition method of the present embodiment, as shown in Figure 2 , includes the following steps.
[0029] 1. Visual acquisition step, acquiring images of the TFT-LCD panel under a plurality of state lights. The quality evaluation of the TFT-LCD panel mainly refers to the display effect under R, G, B, White, Black, Gray, and H seven state lights. Under different states, the texture characteristics and defect performance of the panel are different, and the image acquisition scheme is optimized accordingly. The specific description is as follows.
[0030] R / G / B state light optimization: In R, G, B states, the periodic texture background of the panel is most obvious, and individual liquid crystal tubes are clearly visible. A high-frequency LED array light source is used, combined with a high-resolution CCD camera (such as 2 million pixels), to ensure that the image can capture tiny point defects and line defects. The exposure time is set to 1-10 ms, automatically adjusted according to the panel brightness; the resolution is set to 80%-100% of the original panel resolution, balancing image quality and processing speed. In these states, mainly detect point defects (such as bright spots, dark spots) and line defects (such as short lines, open lines).
[0031] White / Black state light optimization: In White and Black states, the texture characteristics of the panel are weakened, but the overall characteristics are enhanced, such as scratches, foreign matter, and other defects with improved contrast. A low-frequency LED array light source is used, combined with a large field of view telephoto lens, to ensure that the entire panel can be covered and large-area defects can be captured. The exposure time is set to 10-100 ms, automatically adjusted according to the ambient light; the resolution is set to 50%-80% of the original panel resolution, to improve processing speed. In these states, mainly detect large-area defects and contrast abnormalities.
[0032] Gray / H state light optimization: In Gray and H states, Mura-type defects of the panel are most obvious. A uniform illumination LED array light source is used, combined with a high-sensitivity CCD camera, to ensure that low-contrast cloud spot-type defects can be captured. The exposure time is set to 100-500 ms, automatically adjusted according to the panel brightness and Mura severity; the resolution is set to 30%-50% of the original panel resolution, to improve processing speed and reduce computational load. In these states, mainly detect Mura-type defects (such as brightness unevenness, color spots).
[0033] Multi-light source switching strategy: Through intelligent light source switching algorithm, the best state light is automatically selected according to the panel size and defect type. For small size high resolution panels, R / G / B state light is preferred for detailed detection; for large size panels, White / Black state light is preferred to improve detection efficiency; for panels suspected to have Mura-type defects, switch to Gray / H state light for key detection. The light source switching time is controlled within 100 ms, ensuring that the overall real-time performance of the system is not affected.
[0034] Image quality evaluation: After collecting images under each state light, the system automatically evaluates the image quality (such as signal-to-noise ratio, contrast, uniformity, etc.), ensuring that the image meets the detection requirements. For images that do not meet the quality standards, the system will automatically re-collect, avoiding false positives or missed detections due to image quality problems. Image quality evaluation standards include: brightness uniformity error ≤ ± 2.5%, chroma error ≤ ± 0.01, signal-to-noise ratio ≥ 30 dB, etc.
[0035] In the image acquisition process of TFT-LCD panel, the image quality is often seriously affected due to the periodic texture background on the panel surface and environmental interference. The system performs the following image preprocessing steps before the defect classification step, realizes the special image preprocessing and enhancement strategy, and effectively improves the defect detection effect. The image preprocessing steps include the following steps.
[0036] The weighted template difference image processing step selects a preset small size image at the background position of the acquired image as an ideal texture template, divides the acquired image into a plurality of single units of the same size, subtracts each unit from the ideal texture template after weighted processing, and then synthesizes all the units to remove the texture background. The preset small size image at the background position of the acquired image refers to an image of size l x k at the image marker position. The weighting factor w is calculated by the formula w = mean_ij / mean_cell to ensure the accuracy of brightness correction.
[0037] The bilateral filter denoising step still has a small amount of residual texture background and noise in the residual image processed by the difference image method. For all pixels processed by the weighted template difference image processing step, the weights of all pixels in the neighborhood are calculated based on the spatial distance and color difference, the pixel values in the neighborhood are weighted and averaged according to the calculated weights, and the calculated pixel neighborhood is taken as the pixel value after neighborhood filtering. The bilateral filter algorithm is adopted to consider the weighted average of the pixel domain and the spatial domain, which effectively suppresses the residual noise. The two Gaussian kernel functions of the bilateral filter control the spatial distance and the pixel value difference, and by adjusting σ_s (spatial standard deviation) and σ_r (pixel value standard deviation), the texture characteristics of different panels can be flexibly adapted.
[0038] The maximum entropy segmentation step calculates the foreground and background entropy values after threshold segmentation of the image processed by the bilateral filter denoising step, and selects the threshold with the maximum total entropy value as the threshold standard for segmenting the foreground and background. For the processed image, the threshold segmentation algorithm based on maximum entropy is adopted, which is less affected by the size of the target defect and can realize high-precision defect segmentation.
[0039] It also includes a Gabor filter texture enhancement step or a brightness equalization step performed after the maximum entropy segmentation step, specifically:
[0040] If the state light of the current image is R, G or B light, the Gabor filter texture enhancement step is performed to enhance the visibility of individual liquid crystal tubes, so as to extract texture information of different directions and sizes; the Gabor filter can extract texture information of different directions and scales by adjusting the spatial frequency, direction and bandwidth, which is particularly suitable for detecting point defects and line defects;
[0041] If the state light of the current image is White, Black, Gray or H light, the brightness equalization step is adopted to adjust the histogram equalization to enhance the overall characteristics of the image and improve the defect contrast; by adjusting the gray scale distribution of the image, the contrast between the defect area and the background area is maximized, and the detection effect is improved.
[0042] 2. The defect classification step calls the trained defect classification model to identify the defect type of the acquired image, which includes multiple types of point defects, line defects and Mura defects; if Mura defects are identified, the image enhancement step is performed, which includes background suppression and B-spline surface fitting to segment the Mura area in the image; in the image enhancement step, background suppression refers to contrast enhancement processing of the extracted background image.
[0043] The traditional U-NET network has problems such as large number of model parameters, insufficient real-time performance, poor detection effect on low-contrast Mura defects, etc. in TFT-LCD defect detection. Therefore, based on the improved U-NET architecture, combined with attention mechanism and lightweight design, high-precision and real-time defect detection is realized. Specifically, MobileNetV2 is used as the backbone network, and depth separable convolution is used to decompose ordinary convolution into channel-wise convolution and 1x1 point convolution, which greatly reduces the calculation amount.
[0044] Lightweight U-NET architecture: multi-task classification design: the improved U-NET adopts an encoder-decoder multi-head output architecture, and designs a special processing path for point defects (high texture), line defects (edge features) and Mura defects (low-contrast areas). The point defect detection branch uses high-resolution feature maps to enhance the sensitivity to small defects; the line defect detection branch uses edge detection and morphological processing to extract clear linear features; the Mura defect detection branch combines background suppression and low-contrast feature enhancement technology to improve the recognition ability of cloud spot defects. Among them, the defects corresponding to the R / G / B state light of the embodiment include point defects or line defects; the defects corresponding to the White / Black state light include large-area defects or contrast abnormal defects; the defects corresponding to the Gray / H state light include Mura defects.
[0045] Among them, the attention mechanism fusion of the embodiment is to embed convolution block attention module (CBAM) and focus linear attention module (FLA) in the network to enhance the attention degree of key features. CBAM adjusts the feature weight dynamically through attention mechanism in channel and space dimensions; FLA focuses on small target feature extraction, which is particularly suitable for detecting small point defects and low-contrast Mura defects. These attention mechanisms combined with the lightweight design of the improved U-NET not only improve the detection accuracy, but also maintain the real-time processing capability.
[0046] The embodiment also improves the segmentation accuracy by optimizing the mixed loss function: a mixed loss function combining Focal and Dice is used to solve the class imbalance problem in small sample defect detection. The Focal loss function effectively reduces the weight of easy-to-classify samples and focuses on difficult-to-classify samples; the Dice loss function improves the segmentation accuracy by calculating the overlap rate of the predicted mask and the real mask. The combination of the two loss functions enables the model to maintain high accuracy under limited training samples.
[0047] Further, through Mura detection enhancement technology, B-spline surface fitting and background suppression technology are introduced in the Mura defect detection branch. The background suppression performs contrast enhancement processing on the extracted background image, and the B-spline surface fitting solves the problem of Mura defects being weakened due to high fitting accuracy in traditional methods by step-by-step solving of bicubic B-spline surface function through product algorithm for image smoothing processing. At the same time, through adaptive initial curve and level set function, accurate segmentation of Mura regions with uncertain shape and quantity is realized, and data compression and block fitting method is used to improve the algorithm efficiency.
[0048] 3. An image matching step performed after the defect classification step, identifying the correspondence between the state light of the acquired image and the defect type, and if the preset correspondence is not met, the defect classification step is performed again until the preset correspondence is met or the defect classification step is performed for a preset number of times.
[0049] 4. A defect extraction step, extracting the defect features of the corresponding type for each defect type image, and outputting the extracted defect features as a set of defects of the current TFT-LCD panel.
[0050] The TFT panel defect recognition system of the embodiment realizes high-precision and real-time detection of point defects, line defects and Mura defects through the combination of a double-station visual acquisition system and a lightweight U-NET algorithm. The system has adaptive detection parameter configuration capability for multi-specification LCDs and can effectively meet the detection needs of panels of different sizes and resolutions. The double-station visual acquisition system includes a high-resolution CCD camera and a telecentric lens. The CCD camera has 1.3 million pixels, and the telecentric lens has a depth of field of 3 mm. The system realizes accurate alignment and focusing of panels of different sizes by cooperating with a manually adjustable X, Y and Z axis displacement platform. Based on the double-station imaging lens, the system can automatically identify LCD panels of different sizes and resolutions, dynamically adjust detection parameters, and ensure that image acquisition is not affected by changes in panel thickness and position fluctuations, thereby maintaining high-precision detection effects on panels of various specifications. Specifically, the system integrates a parameter configuration database to store the optimal acquisition parameters (such as exposure time, color temperature, resolution, etc.) of LCD panels of different sizes and resolutions. The system automatically identifies the size of the panel through a 3D camera and a deep learning algorithm (such as YOLOv8 lightweight version), triggers parameter library query, and realizes dynamic parameter adjustment. For example, for large-size panels of the 10.5 generation line (3370x2940mm), the system will automatically extend the exposure time and adjust the light source distance; while for small-size panels, shorter exposure time and closer light source distance are used.
[0051] The embodiment realizes the TFT panel defect recognition method described above through a computer program, which is stored in a computer program product and executed by a computer processor to realize the TFT panel defect recognition method described above. The TFT panel defect recognition system embodiment described above is only illustrative, and the modules described as separate components can or can not be physically separated, and the components displayed as modules can or can not be physical modules, i.e., they can be located in one place or distributed on multiple network modules. Some or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. Those skilled in the art can understand and implement without creative labor.
[0052] Finally, it should be noted that the TFT panel defect recognition method disclosed in the embodiments of the present application is only the preferred embodiment of the present application and is used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions described in the foregoing embodiments can be modified or some technical features can be replaced by equivalents without departing from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A TFT panel defect identification method, comprising the following steps: a visual acquisition step of acquiring images of a TFT-LCD panel under multiple state lights; a defect classification step of calling a trained defect classification model to identify the defect type of the acquired images, the defect type including multiple types among point defects, line defects and Mura defects; if a Mura defect is identified, an image enhancement step is performed, the image enhancement step including background suppression and B-spline surface fitting to segment the Mura area in the image; a defect extraction step of extracting defect features of the corresponding type for the images of each defect type respectively, and outputting all the extracted defect features as a defect set of the current TFT-LCD panel.
2. The TFT panel defect identification method as described in claim 1, characterized in that, an image matching step performed after the defect classification step to identify the correspondence between the state light and the defect type of the acquired images; if the correspondence does not meet a preset correspondence, the defect classification step is performed again until the preset correspondence is met or the defect classification step is performed for a preset number of times.
3. The TFT panel defect recognition method according to claim 2, wherein the TFT panel defect recognition method is characterized by, The state lights include R, G, B, White, Black, Gray and H; the defects corresponding to the R / G / B state lights include point defects or line defects; the defects corresponding to the White / Black state lights include large-area defects or contrast abnormal defects; the defects corresponding to the Gray / H state lights include Mura defects.
4. The TFT panel defect identification method as described in claim 1, characterized in that, An image preprocessing step performed before the defect classification step includes the following steps: a weighted template difference image processing step of selecting a preset small-size image at the background position of the acquired image as an ideal texture template, segmenting the acquired image into a plurality of single units of the same size, subtracting each unit from the ideal texture template after weighted processing, and then synthesizing all the units to remove the texture background; a bilateral filter denoising step of calculating the weight of all pixels in the neighborhood of each pixel based on the spatial distance and color difference, and performing weighted average on the pixel values in the neighborhood according to the calculated weight, so as to obtain the pixel neighborhood as the filtered pixel value in the neighborhood; a maximum entropy segmentation step of calculating the foreground and background entropy values after threshold segmentation of the image processed by the bilateral filter denoising step, and selecting the threshold with the maximum total entropy value as the threshold standard for segmenting the foreground and background.
5. The TFT panel defect recognition method according to claim 4, wherein the TFT panel defect recognition method is characterized by, In the weighted template difference image processing step, the preset small-size image at the background position of the acquired image refers to an image of size l×k at the image marker position.
6. The TFT panel defect recognition method according to claim 4, wherein the TFT panel defect recognition method is characterized by, A Gabor filter texture enhancement step or a brightness equalization step is further performed after the maximum entropy segmentation step, specifically: if the state light of the current image is R, G or B light, the Gabor filter texture enhancement step is performed to enhance the visibility of a single liquid crystal tube, so as to extract texture information in different directions and sizes; if the state light of the current image is White, Black, Gray or H light, the brightness equalization step is adopted to adjust the histogram equalization to enhance the overall features of the image and improve the defect contrast.
7. The TFT panel defect identification method as described in claim 1, characterized in that, In the image enhancement step, the background suppression refers to contrast enhancement processing of the extracted background image.
8. The TFT panel defect identification method as described in claim 1, characterized in that, The B-spline surface fitting refers to step-by-step solving of a bi-cubic B-spline surface function by a product-type algorithm to smooth the image.
9. A computer program product storing a computer program, characterized in that, The computer program, when executed by a processor, can implement the steps of the TFT panel defect identification method in any one of claims 1-7.
10. A TFT panel defect recognition system, characterized by, The computer program product comprises a processor and a controller, and an image taking lens and a plurality of state light sources connected to the controller respectively, the plurality of state light sources are switched to align the TFT-LCD panel under the control of the controller, and the computer program product further comprises the computer program as claimed in claim 9, which can be executed by the processor.