A method for detecting defects in liquid crystal display screen

Abstract

The invention discloses a method for detecting defects in a liquid crystal display screen, and relates to the technical field of defect detection in liquid crystal display screens. The invention adopts a bidirectional polarized transmission structure light source to project polarized light of 0° and 90°, reduces interference of parasitic reflection on defect detection, introduces an adaptive exposure control unit, adjusts exposure parameters in real time according to the transmittance and reflection characteristics of a glass substrate, optimizes imaging quality, and avoids detection effects affected by overexposure or underexposure; in an image processing stage, parasitic reflection information is extracted by combining a polarization difference algorithm of double-frame subtraction, and high-frequency interference is removed by using a fast Fourier transform (FFT), so as to obtain an optimized defect image after parasitic reflection suppression, so as to make real defect information more prominent and improve defect contrast; multi-scale Gaussian fitting filtering, a gradient lifting method, and super-resolution reconstruction are adopted to enhance defect features in an image, so as to make low-contrast defect areas easier to identify.

Description

Technical Field

[0001] The invention relates to the technical field of liquid crystal display screen defect detection, and in particular to a liquid crystal display screen defect detection method. Background Art

[0002] Defect detection is extremely important in the manufacturing process of LCD screens. As electronic devices develop towards high resolution, large screens and ultra-thinness, the process precision of screen manufacturing continues to improve. Any tiny defect may become a key factor affecting the qualification rate of the entire batch of products.

[0003] In the production process of TFTLCD displays, defect detection of glass substrates mostly uses reflective optical methods, that is, using high-resolution industrial cameras and specific lighting systems to photograph the screen surface, and identifying defect features through image processing algorithms. However, as glass substrates develop towards ultra-thinness and high transmittance, their optical properties cause light to be reflected multiple times inside the glass during the detection process, thereby generating "virtual images" or parasitic stripes in the detection image, causing real defect information to be interfered with or misjudged.

[0004] To address this problem, some traditional methods use multi-angle lighting, illuminating the glass surface with light from different directions, taking images at multiple angles, and then using algorithms to synthesize a high-confidence defect image. However, since light from different angles may still undergo complex refraction and reflection inside the glass, parasitic images are difficult to completely eliminate. In addition, polarizers can use the selective transmission characteristics of polarized light to weaken some reflected light, but in high-curvature glass inspection, the filtering effect of the polarizer is limited due to the multiple scattering and refraction of light, and parasitic reflections cannot be completely removed.

[0005] It can be seen that under the current detection requirements, the false detection rate and missed detection rate of traditional methods are relatively high; therefore, how to further optimize the optical detection scheme without increasing too much computing cost so that it can more effectively suppress parasitic reflections while retaining the true defect information is still an urgent problem to be solved in the field of LCD display defect detection. Summary of the invention

[0006] In view of the above existing problems, the present invention is proposed.

[0007] The present invention provides a method for detecting defects of a liquid crystal display screen, which solves the problem that conventional glass substrate detection is affected by parasitic reflection, resulting in misjudgment of defects, and the existing filtering and polarization methods are still unable to completely eliminate virtual image interference.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0009] An embodiment of the present invention provides a method for detecting defects of a liquid crystal display screen, which comprises:

[0010] Step S1, building an optical imaging system for transparent glass substrate defect detection, and acquiring an original detection image after polarization modulation;

[0011] Step S2, performing image acquisition based on the original detection image obtained in step S1, recording the detection results under 0° polarized light and 90° polarized light respectively, and recording multiple frames of detection images under different polarized lights;

[0012] Step S3, performing image fusion and parasitic reflection suppression on the multiple frames of detection images in step S2 to obtain an optimized defect detection image after parasitic reflection suppression;

[0013] Step S4, performing defect enhancement processing on the optimized defect image to obtain an enhanced high-resolution defect image;

[0014] Defect enhancement processing includes multi-scale Gaussian fitting filtering, gradient boosting method and defect feature amplification based on super-resolution reconstruction;

[0015] Step S5, based on the high-resolution defect image, using a deep learning-based classifier to perform automatic defect recognition and classification to obtain defect detection results and classification information;

[0016] Step S6, performing intelligent analysis and quality assessment on the defect detection results and classification information of step S5, and generating a detection report.

[0017] As a preferred solution of the method for detecting defects in a liquid crystal display screen described in the present invention, the optical imaging system includes an industrial camera, a bidirectional polarized transmission structure light source and a multi-angle lighting module.

[0018] Wherein, the bidirectional polarization transmission structure light source is used to project polarized light of 0° and 90°.

[0019] As a preferred solution of the method for detecting defects in a liquid crystal display screen described in the present invention, the image acquisition process of step S2 adopts a CMOS sensor and an adaptive exposure control unit, wherein the adaptive exposure control unit adjusts the exposure parameters according to the transmittance and surface reflection characteristics of the glass substrate.

[0020] As a preferred solution of the method for detecting defects in a liquid crystal display screen of the present invention, the step of the adaptive exposure control unit adjusting the exposure parameters according to the light transmittance and surface reflection characteristics of the glass substrate is as follows:

[0021] Calculate the average light transmittance of the glass substrate using the following formula:

[0022] ,

[0023] in, Represents the average light transmittance of the glass substrate, Represents the total number of sampling points in the detection area, Indicates The transmittance of each sampling point,

[0024] Calculate the average reflectivity of the glass substrate using the following formula:

[0025] ,

[0026] in, represents the average reflectivity of the glass substrate, represents the number of measurement points within the sampling area, Indicates The surface reflectivity of the measuring points,

[0027] The optimal exposure time is calculated based on the average transmittance and average reflectance. The calculation formula is:

[0028] ,

[0029] in, represents the calculated optimal exposure time, Indicates the base exposure time set by the system.

[0030] Calculate the exposure time after adaptive adjustment, the calculation formula is:

[0031] ,

[0032] in, Represents the exposure time after adaptive adjustment. and To adjust the system parameters adaptively, according to historical test data and current environmental conditions,

[0033] Adjust the exposure gain of the CMOS sensor. The adjustment formula is:

[0034] ,

[0035] in, Represents the adjusted exposure gain, This is the initial gain setting.

[0036] As a preferred solution of the method for detecting defects in a liquid crystal display screen described in the present invention, wherein:

[0037] In step S3, a polarization difference algorithm of double-frame subtraction is used to extract parasitic reflection information;

[0038] In step S3, the high-frequency interference component is removed by using the fast Fourier transform (FFT).

[0039] As a preferred solution of the method for detecting defects in a liquid crystal display screen described in the present invention, the step of performing image fusion and parasitic reflection suppression on the multiple frames of detection images in step S2 to obtain an optimized defect detection image after parasitic reflection suppression is as follows:

[0040] Acquire dual-frame polarization images: and The detection images are obtained under 0° and 90° polarized light respectively, then:

[0041] ,

[0042] ,

[0043] in, represents the surface scattered light component, represents the parasitic reflection component,

[0044] The parasitic reflection information is extracted using the double-frame subtraction method:

[0045] ,

[0046] in, represents the parasitic reflection component,

[0047] Construct a polarization-differential dereflected image:

[0048] ,

[0049] in, represents the image after removing parasitic reflections;

[0050] Apply Fast Fourier Transform FFT to remove high frequency interference:

[0051] Assume that the frequency domain representation of the original image is , the high frequency filter function is ,but:

[0052] ,

[0053] in, represents the optimized defect detection image after removing high-frequency interference, represents the inverse Fourier transform.

[0054] As a preferred solution of the method for detecting defects in a liquid crystal display screen described in the present invention, the step of performing defect enhancement processing on the optimized defect image to obtain an enhanced high-resolution defect image is as follows:

[0055] Apply a multiscale Gaussian fitting filter:

[0056] ,

[0057] in, represents the enhanced image after Gaussian filtering, Represents standard deviation The Gaussian filter is applied to the image ,

[0058] The gradient boosting method is used to enhance the defect edge, and the formula is:

[0059] ,

[0060] in, represents the enhanced defect edge image, and Respectively indicate along and Directional gradient calculation;

[0061] Amplify defect features based on super-resolution reconstruction:

[0062] Assume that the reconstructed image output by the deep learning model is ,but:

[0063] ,

[0064] in, is the high-resolution defect image after super-resolution enhancement. is the enhancement factor.

[0065] As a preferred solution of the liquid crystal display defect detection method described in the present invention, in which: in step S5, a pre-trained CNN model is used in combination with an improved Attention mechanism to extract and classify features of the defect area, and a dynamic gated convolution method is used to optimize the edge detection of defects to obtain defect detection results and classification information.

[0066] As a preferred solution of the method for detecting defects in a liquid crystal display screen described in the present invention, the step of using a deep learning-based classifier to automatically identify and classify defects to obtain defect detection results and classification information is as follows:

[0067] Construct deep learning input feature matrix: The high-resolution defect image after super-resolution enhancement is , then the feature matrix of the input deep learning model is:

[0068] ,

[0069] in, represents the input feature matrix, It is a feature extraction function responsible for transforming the image Mapped to a high-dimensional feature space,

[0070] The CNN model is used to extract defect features. The extraction process is as follows:

[0071] ,

[0072] in, Represents the input feature matrix The extracted deep features, Represents a convolutional neural network, which includes convolutional layers, pooling layers, and fully connected layers.

[0073] An improved Attention mechanism is introduced to enhance feature representation. The enhancement process is expressed as:

[0074] ,

[0075] in, is the Attention weight matrix, and are the weight matrix and bias vector of the Attention layer respectively. is the activation function,

[0076] Calculate the Attention weighted feature representation, the calculation formula is:

[0077] ,

[0078] in, is the feature matrix optimized by the Attention mechanism, represents element-wise multiplication,

[0079] Dynamic gated convolution is used to optimize edge features. The optimization process is as follows:

[0080] ,

[0081] in, represents the optimized edge features, is the dynamic gated convolution kernel, is the bias term, represents the convolution operation,

[0082] The defect classifier calculates the final defect class probability:

[0083] ,

[0084] in, is the probability distribution of defect categories, is the weight matrix of the classification layer, is the bias vector, The function is used to map features to category probability space.

[0085] Output the final defect category based on the maximum probability category :

[0086] ,

[0087] in, is the final defect category.

[0088] As a preferred solution of the method for detecting defects in a liquid crystal display screen described in the present invention, the step of performing intelligent analysis and quality assessment on the defect detection results and classification information of step S5 and generating a detection report is as follows:

[0089] The statistical formula for the categories and quantities of defects is:

[0090] ,

[0091] in, Indicates category The number of defects, For the The classification results of defects are is the indicator function, when The value is 1 when , otherwise it is 0.

[0092] Calculate the defect coverage rate using the following formula:

[0093] ,

[0094] in, For Category Defect coverage, is the total area of ​​defects of this category, is the total area of ​​the detection area,

[0095] Calculate the quality score using the formula:

[0096] ,

[0097] in, Rate the quality, For Category The weight coefficient of

[0098] Generate test report :

[0099] ,

[0100] in, Represents the final inspection report, which includes the number of defect categories, coverage, and quality scores.

[0101] The beneficial effects of the present invention are as follows: the present invention adopts a bidirectional polarized transmission structure light source to project 0° and 90° polarized light, reduces the interference of parasitic reflection on defect detection, introduces an adaptive exposure control unit, adjusts exposure parameters in real time according to the transmittance and reflection characteristics of the glass substrate, optimizes imaging quality, and avoids the detection effect being affected by overexposure or underexposure; in the image processing stage, the parasitic reflection information is extracted by combining the polarization difference algorithm of dual-frame subtraction, and the fast Fourier transform FFT is used to remove high-frequency interference, so as to obtain an optimized defect image after parasitic reflection suppression, so as to make the real defect information more prominent and improve the defect contrast.

[0102] The present invention adopts multi-scale Gaussian fitting filtering, gradient boosting method and super-resolution reconstruction to enhance the defect features in the image, making the low-contrast defect area easier to identify. In the defect identification link, a CNN-based deep learning classifier is introduced, combined with the Attention mechanism to optimize feature extraction, and dynamic gated convolution is used to improve the defect edge detection capability, so that the classifier can more accurately identify the defect type and reduce the classification errors caused by the diversity of defect morphology.

[0103] The present invention adopts an intelligent analysis method for statistical defect categories and coverage and calculating quality scores. Through quantitative calculation of defect distribution, it improves the visual evaluation capability of the overall quality of liquid crystal display screens, and automatically generates test reports to support real-time feedback and quality optimization of the production line.

[0104] Compared with traditional methods, the present invention significantly reduces the detection interference caused by parasitic reflections, improves the stability of high-curvature glass and OLED screen detection, and enhances the ability to identify tiny defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0106] Figure 1 The figure is a schematic flow chart of the method for detecting defects in a liquid crystal display screen according to the present invention. DETAILED DESCRIPTION

[0107] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0108] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0109] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0110] Example 1, reference Figure 1 , this embodiment provides a method for detecting defects in a liquid crystal display screen, comprising the following steps:

[0111] Step S1, building an optical imaging system for transparent glass substrate defect detection, and acquiring an original detection image after polarization modulation;

[0112] The optical imaging system includes an industrial camera, a bidirectional polarized transmission structured light source and a multi-angle lighting module.

[0113] Wherein, the bidirectional polarized transmission structure light source is used to project polarized light of 0° and 90°;

[0114] Step S2, performing image acquisition based on the original detection image obtained in step S1, recording the detection results under 0° polarized light and 90° polarized light respectively, and recording multiple frames of detection images under different polarized lights;

[0115] The image acquisition process of step S2 uses a CMOS sensor and an adaptive exposure control unit, wherein the adaptive exposure control unit adjusts exposure parameters according to the light transmittance and surface reflection characteristics of the glass substrate;

[0116] The step of the adaptive exposure control unit adjusting the exposure parameters according to the light transmittance and surface reflection characteristics of the glass substrate is as follows:

[0117] Calculate the average light transmittance of the glass substrate using the following formula:

[0118] ,

[0119] in, Represents the average light transmittance of the glass substrate, Represents the total number of sampling points in the detection area, Indicates The transmittance of each sampling point,

[0120] Calculate the average reflectivity of the glass substrate using the following formula:

[0121] ,

[0122] in, represents the average reflectivity of the glass substrate, represents the number of measurement points within the sampling area, Indicates The surface reflectivity of the measuring points,

[0123] The optimal exposure time is calculated based on the average transmittance and average reflectance. The calculation formula is:

[0124] ,

[0125] in, represents the calculated optimal exposure time, Indicates the base exposure time set by the system.

[0126] Calculate the exposure time after adaptive adjustment, the calculation formula is:

[0127] ,

[0128] in, Represents the exposure time after adaptive adjustment. and To adjust the system parameters adaptively, according to historical test data and current environmental conditions,

[0129] Adjust the exposure gain of the CMOS sensor. The adjustment formula is:

[0130] ,

[0131] in, Represents the adjusted exposure gain, is the initial setting gain,

[0132] Specifically, by calculating the average transmittance and reflectivity of the glass substrate and dynamically adjusting the exposure time and gain, the imaging process always obtains a clear original detection image;

[0133] Since the transmittance of the glass substrate may be affected by process changes and the reflectivity is affected by the angle of the light source, dynamic exposure time adjustment is used to eliminate overexposure and underexposure, improve the contrast of the detection image, and adaptively adjust the model to optimize according to the ambient light intensity and glass characteristics to improve the stability of the detection system;

[0134] Step S3, performing image fusion and parasitic reflection suppression on the multiple frames of detection images in step S2 to obtain an optimized defect detection image after parasitic reflection suppression;

[0135] In step S3, a polarization difference algorithm of double-frame subtraction is used to extract parasitic reflection information;

[0136] In step S3, the high frequency interference component is removed by using the fast Fourier transform FFT;

[0137] The step of performing image fusion and parasitic reflection suppression on the multi-frame detection images of step S2 to obtain the optimized defect detection image after parasitic reflection suppression is as follows:

[0138] Acquire dual-frame polarization images: and The detection images are obtained under 0° and 90° polarized light respectively, then:

[0139] ,

[0140] ,

[0141] in, represents the surface scattered light component, represents the parasitic reflection component,

[0142] The parasitic reflection information is extracted using the double-frame subtraction method:

[0143] ,

[0144] in, represents the parasitic reflection component,

[0145] Construct a polarization-differential dereflected image:

[0146] ,

[0147] in, represents the image after removing parasitic reflections;

[0148] Apply Fast Fourier Transform FFT to remove high frequency interference:

[0149] Assume that the frequency domain representation of the original image is , the high frequency filter function is ,but:

[0150] ,

[0151] in, represents the optimized defect detection image after removing high-frequency interference, represents the inverse Fourier transform,

[0152] Specifically, based on the images of 0° and 90° polarized light, the parasitic reflection information is extracted by double-frame subtraction, the parasitic reflection components are calculated, and unnecessary reflections are removed from the original image to improve the accuracy of defect detection. In addition, the fast Fourier transform is used to remove high-frequency interference to make the edge information of the image clearer.

[0153] Step S4, performing defect enhancement processing on the optimized defect image to obtain an enhanced high-resolution defect image;

[0154] Defect enhancement processing includes multi-scale Gaussian fitting filtering, gradient boosting method and defect feature amplification based on super-resolution reconstruction;

[0155] The step of performing defect enhancement processing on the optimized defect image to obtain an enhanced high-resolution defect image is:

[0156] Apply a multiscale Gaussian fitting filter:

[0157] ,

[0158] in, represents the enhanced image after Gaussian filtering, Represents standard deviation The Gaussian filter is applied to the image ,

[0159] The gradient boosting method is used to enhance the defect edge, and the formula is:

[0160] ,

[0161] in, represents the enhanced defect edge image, and Respectively indicate along and Directional gradient calculation;

[0162] Amplify defect features based on super-resolution reconstruction:

[0163] Assume that the reconstructed image output by the deep learning model is ,but:

[0164] ,

[0165] in, is the high-resolution defect image after super-resolution enhancement. is the enhancement factor,

[0166] Specifically, the scaled Gaussian filter is used to smooth the background while retaining the detailed information of the defect area and reducing noise interference; the gradient boosting enhances the edge information of the defect area and improves the clarity of detection; the super-resolution reconstruction uses deep learning methods to further amplify the defect features, thereby distinguishing smaller defects and improving detection accuracy;

[0167] Step S5, based on the high-resolution defect image, using a deep learning-based classifier to perform automatic defect recognition and classification to obtain defect detection results and classification information;

[0168] In step S5, the pre-trained CNN model is combined with the improved Attention mechanism to extract and classify the defect area, and the dynamic gated convolution method is used to optimize the edge detection of the defect to obtain the defect detection result and classification information;

[0169] The step of using a deep learning-based classifier to automatically identify and classify defects to obtain defect detection results and classification information is as follows:

[0170] Construct deep learning input feature matrix: The high-resolution defect image after super-resolution enhancement is , then the feature matrix of the input deep learning model is:

[0171] ,

[0172] in, represents the input feature matrix, It is a feature extraction function responsible for transforming the image Mapped to a high-dimensional feature space,

[0173] The CNN model is used to extract defect features. The extraction process is as follows:

[0174] ,

[0175] in, Represents the input feature matrix The extracted deep features, Represents a convolutional neural network, which includes convolutional layers, pooling layers, and fully connected layers.

[0176] An improved Attention mechanism is introduced to enhance feature representation. The enhancement process is expressed as:

[0177] ,

[0178] in, is the Attention weight matrix, and are the weight matrix and bias vector of the Attention layer respectively. is the activation function,

[0179] Calculate the Attention weighted feature representation, the calculation formula is:

[0180] ,

[0181] in, is the feature matrix optimized by the Attention mechanism, represents element-wise multiplication,

[0182] Dynamic gated convolution is used to optimize edge features. The optimization process is as follows:

[0183] ,

[0184] in, represents the optimized edge features, is the dynamic gated convolution kernel, is the bias term, represents the convolution operation,

[0185] The defect classifier calculates the final defect class probability:

[0186] ,

[0187] in, is the probability distribution of defect categories, is the weight matrix of the classification layer, is the bias vector, The function is used to map features to category probability space.

[0188] Output the final defect category based on the maximum probability category :

[0189] ,

[0190] in, is the final defect category;

[0191] Specifically, this step constructs the input feature matrix of deep learning, extracts defect features through CNN, introduces an improved Attention mechanism to enhance feature representation and improve classification accuracy; dynamic gated convolution is used to optimize the edge features of defects, making the classifier learn defect morphology more accurately, calculates category probabilities using the Softmax function, and outputs defect classification results based on the maximum probability, thereby achieving automated defect recognition and classification;

[0192] Step S6, performing intelligent analysis and quality assessment on the defect detection results and classification information of step S5, and generating a detection report;

[0193] The step of performing intelligent analysis and quality assessment on the defect detection results and classification information of step S5 and generating a detection report is as follows:

[0194] The statistical formula for the categories and quantities of defects is:

[0195] ,

[0196] in, Indicates category The number of defects, For the The classification results of defects are is the indicator function, when The value is 1 when , otherwise it is 0.

[0197] Calculate the defect coverage rate using the following formula:

[0198] ,

[0199] in, For Category Defect coverage, is the total area of ​​defects of this category, is the total area of ​​the detection area,

[0200] Calculate the quality score using the formula:

[0201] ,

[0202] in, Rate the quality, For Category The weight coefficient of

[0203] Generate test report :

[0204] ,

[0205] in, Represents the final inspection report, including the number of defect categories, coverage, and quality scores;

[0206] Specifically, the defect detection results are statistically analyzed, the number and coverage of defects in each category are calculated, and the quality score is calculated by weighting to quantify the overall quality level of the LCD screen.

[0207] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for detecting defects in a liquid crystal display screen, characterized in that: include, Step S1, building an optical imaging system for transparent glass substrate defect detection, and acquiring an original detection image after polarization modulation; Step S2, performing image acquisition based on the original detection image obtained in step S1, recording the detection results under 0° polarized light and 90° polarized light respectively, and recording multiple frames of detection images under different polarized lights; Step S3, performing image fusion and parasitic reflection suppression on the multiple frames of detection images in step S2 to obtain an optimized defect detection image after parasitic reflection suppression; In step S3, a polarization difference algorithm of double-frame subtraction is used to extract parasitic reflection information; In step S3, the high frequency interference component is removed by using the fast Fourier transform FFT; The step of performing image fusion and parasitic reflection suppression on the multi-frame detection images of step S2 to obtain the optimized defect detection image after parasitic reflection suppression is as follows: Acquire dual-frame polarization images: and The detection images are obtained under 0° and 90° polarized light respectively, then: ， ， in, represents the surface scattered light component, represents the parasitic reflection component, The parasitic reflection information is extracted using the double-frame subtraction method: ， in, represents the parasitic reflection component, Construct a polarization-differential dereflected image: ， in, represents the image after removing parasitic reflections; Apply Fast Fourier Transform FFT to remove high frequency interference: Assume that the frequency domain representation of the original image is , the high frequency filter function is ,but: ， in, represents the optimized defect detection image after removing high-frequency interference, represents inverse Fourier transform; Step S4, performing defect enhancement processing on the optimized defect image to obtain an enhanced high-resolution defect image; Defect enhancement processing includes multi-scale Gaussian fitting filtering, gradient boosting method and defect feature amplification based on super-resolution reconstruction; The step of performing defect enhancement processing on the optimized defect image to obtain an enhanced high-resolution defect image is: Apply a multiscale Gaussian fitting filter: ， in, represents the enhanced image after Gaussian filtering, Represents standard deviation The Gaussian filter is applied to the image , The gradient boosting method is used to enhance the defect edge, and the formula is: ， in, represents the enhanced defect edge image, and Respectively indicate along and Directional gradient calculation; Amplify defect features based on super-resolution reconstruction: Assume that the reconstructed image output by the deep learning model is ,but: ， in, is the high-resolution defect image after super-resolution enhancement. is the enhancement coefficient; Step S5, based on the high-resolution defect image, using a deep learning-based classifier to perform automatic defect recognition and classification to obtain defect detection results and classification information; In step S5, the pre-trained CNN model is combined with the improved Attention mechanism to extract and classify the defect area, and the dynamic gated convolution method is used to optimize the edge detection of the defect to obtain the defect detection result and classification information; Step S6, performing intelligent analysis and quality assessment on the defect detection results and classification information of step S5, and generating a detection report.

2. A method for detecting defects in a liquid crystal display screen as claimed in claim 1, characterized in that: The optical imaging system includes an industrial camera, a bidirectional polarized transmission structured light source and a multi-angle lighting module. Wherein, the bidirectional polarization transmission structure light source is used to project polarized light of 0° and 90°.

3. A method for detecting defects in a liquid crystal display screen as claimed in claim 2, characterized in that: The image acquisition process of step S2 uses a CMOS sensor and an adaptive exposure control unit, wherein the adaptive exposure control unit adjusts exposure parameters according to the light transmittance and surface reflection characteristics of the glass substrate.

4. A method for detecting defects in a liquid crystal display screen as claimed in claim 3, characterized in that: The step of the adaptive exposure control unit adjusting the exposure parameters according to the light transmittance and surface reflection characteristics of the glass substrate is as follows: Calculate the average light transmittance of the glass substrate using the following formula: ， in, Represents the average light transmittance of the glass substrate, Represents the total number of sampling points in the detection area, Indicates The transmittance of each sampling point, Calculate the average reflectivity of the glass substrate using the following formula: ， in, represents the average reflectivity of the glass substrate, represents the number of measurement points within the sampling area, Indicates The surface reflectivity of the measuring points, The optimal exposure time is calculated based on the average transmittance and average reflectance. The calculation formula is: ， in, represents the calculated optimal exposure time, Indicates the base exposure time set by the system. Calculate the exposure time after adaptive adjustment, the calculation formula is: ， in, Represents the exposure time after adaptive adjustment. and To adjust the system parameters adaptively, according to historical test data and current environmental conditions, Adjust the exposure gain of the CMOS sensor. The adjustment formula is: ， in, Represents the adjusted exposure gain, This is the initial gain setting.

5. A method for detecting defects in a liquid crystal display screen as claimed in claim 4, characterized in that: The step of using a deep learning-based classifier to automatically identify and classify defects to obtain defect detection results and classification information is as follows: Construct deep learning input feature matrix: The high-resolution defect image after super-resolution enhancement is , then the feature matrix of the input deep learning model is: ， in, represents the input feature matrix, It is a feature extraction function responsible for transforming the image Mapped to a high-dimensional feature space, The CNN model is used to extract defect features. The extraction process is as follows: ， in, Represents the input feature matrix The extracted deep features, Represents a convolutional neural network, which includes convolutional layers, pooling layers, and fully connected layers. An improved Attention mechanism is introduced to enhance feature representation. The enhancement process is expressed as: ， in, is the Attention weight matrix, and are the weight matrix and bias vector of the Attention layer respectively. is the activation function, Calculate the Attention weighted feature representation, the calculation formula is: ， in, is the feature matrix optimized by the Attention mechanism, represents element-wise multiplication, Dynamic gated convolution is used to optimize edge features. The optimization process is as follows: ， in, represents the optimized edge features, is the dynamic gated convolution kernel, is the bias term, represents the convolution operation, The defect classifier calculates the final defect class probability: ， in, is the probability distribution of defect categories, is the weight matrix of the classification layer, is the bias vector, The function is used to map features to category probability space. Output the final defect category based on the maximum probability category : ， in, is the final defect category.

6. A method for detecting defects in a liquid crystal display screen as claimed in claim 5, characterized in that: The step of performing intelligent analysis and quality assessment on the defect detection results and classification information of step S5 and generating a detection report is as follows: The statistical formula for the categories and quantities of defects is: ， in, Indicates category The number of defects, For the The classification results of defects are is the indicator function, when The value is 1 when , otherwise it is 0. Calculate the defect coverage rate using the following formula: ， in, For Category Defect coverage, is the total area of ​​defects of this category, is the total area of ​​the detection area, Calculate the quality score using the formula: ， in, Rate the quality, For Category The weight coefficient of Generate test report : ， in, Represents the final inspection report, which includes the number of defect categories, coverage, and quality scores.

Images