Liquid crystal display screen defect detection method, system, equipment and medium

Abstract

The invention relates to the technical field of display, and discloses a liquid crystal display screen defect detection method which comprises the following steps: collecting defect image data of a liquid crystal display screen, and marking the defect image data; carrying out denoising, normalization and size adjustment processing on the defect image; a multi-scale convolutional neural network is used to extract image features, the features comprise local features and global features, the invention also provides a liquid crystal display screen defect detection system, and the system comprises a data acquisition module used for collecting defect image data of a liquid crystal display screen; and the image preprocessing module is used for carrying out denoising, normalization and size adjustment on the defect image. According to the method, the multi-scale convolutional neural network, variational reasoning, Bayesian optimization and other technologies are combined, the accuracy and efficiency of defect detection of the liquid crystal display screen are improved, particularly, the method has excellent performance in recognition of complex backgrounds and fine defects, and efficient and accurate automatic detection is ensured.

Description

Technical Field

[0001] The present invention relates to the field of display technology, and in particular to a method, system, equipment and medium for detecting defects in a liquid crystal display screen. Background Art

[0002] With the continuous development of modern electronic devices, liquid crystal displays (LCDs), as one of the main forms of display technology, have been widely used in televisions, smart phones, computers and various consumer electronic products. Consumers' requirements for display effects are increasing day by day, and the quality inspection of LCD screens is particularly important. Especially in the production process, how to efficiently and accurately detect various defects in the display screen (such as bright spots, dark spots, color difference, scratches, etc.) is the key to ensuring product quality.

[0003] Existing LCD defect detection technologies mostly rely on manual visual inspection and automated methods based on simple image processing technology. Manual inspection is intuitive and flexible, and can detect most obvious defects in a timely manner, but the work intensity is high and it is easily affected by fatigue and subjective factors, and the efficiency is low. The automated method based on image processing technology identifies defects through traditional algorithms such as image segmentation, edge detection, and threshold setting. This method can effectively improve the detection speed and is suitable for large-scale production line applications.

[0004] There are still some shortcomings in the existing technology, which are mainly reflected in the accuracy and adaptability of defect detection; the existing solutions based on traditional image processing mostly rely on fixed threshold methods, which cannot flexibly deal with various types of defects, especially for small or hidden defects, false detection and missed detection are more serious; in addition, these methods usually have limited processing capabilities and cannot fully utilize complex features for accurate judgment; traditional methods have poor low-contrast defect recognition capabilities under complex backgrounds, and cannot dynamically adjust the detection standards, so it is easy to ignore some subtle defects. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method, system, device and medium for detecting defects in a liquid crystal display screen, which solve the problems in the prior art of low defect detection accuracy and poor recognition of complex backgrounds and tiny defects.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for detecting defects in a liquid crystal display screen, comprising the following steps: Collect defect image data of LCD screens and annotate them; Performing denoising, normalization and size adjustment on the defective image; Using a multi-scale convolutional neural network to extract image features, the features include local features and global features; Perform weighted fusion of feature maps from different scales to obtain the final image feature representation; Infer defect distribution using variational inference methods; Optimize hyperparameters using Bayesian optimization methods; The defect characteristics are further analyzed through fractal analysis and complex network models.

[0007] Preferably, the collecting defective image data of the liquid crystal display screen includes: Acquire multiple defect images of the liquid crystal display screen by an image acquisition device; Annotate the image to indicate the type and location of defects.

[0008] Preferably, the denoising, normalization and resizing of the defective image includes: Use image denoising algorithms to remove noise from images; Normalize the brightness and contrast of the image to ensure image consistency; Resize the image to a fixed size to meet the input requirements of subsequent processing modules.

[0009] Preferably, the multi-scale convolutional neural network extracts image features including: Use multiple convolution kernels of different sizes to extract local features of the image; The feature maps extracted at each scale are weighted and fused to form the final comprehensive feature representation.

[0010] Preferably, the inferring defect distribution using a variational inference method comprises: Modeling defective areas in LCD displays as latent variables; The inference process is optimized by the variational inference algorithm to obtain the inference results of defect distribution.

[0011] Preferably, the method of optimizing hyperparameters using the Bayesian optimization method comprises: Select the objective function to be optimized and automatically adjust the hyperparameters through the Bayesian optimization algorithm; The Bayesian optimization method is used to maximize the expected value and variance of the objective function, thereby automatically selecting the optimal network hyperparameters.

[0012] Preferably, the further analysis of defect characteristics by fractal analysis and complex network model includes: Use fractal analysis to calculate the self-similarity of defects and obtain the fractal dimension of defects; Defects are viewed as graph networks, the correlation between defect nodes is calculated, and the relationship between defects is further analyzed through complex network models.

[0013] The present invention also provides a liquid crystal display screen defect detection system, comprising: A data acquisition module, used for collecting defect image data of the liquid crystal display screen; An image preprocessing module, used for denoising, normalizing and resizing the defective image; Feature extraction module, used to extract defect features through multi-scale convolutional neural network; An optimization module for inferring defect distributions through variational inference methods and optimizing hyperparameters using Bayesian optimization methods; The analysis module is used to further analyze defect characteristics through fractal analysis and complex network models to improve detection accuracy.

[0014] The present invention also provides a liquid crystal display screen defect detection device, comprising: An image acquisition device, used to acquire defect images of a liquid crystal display screen; A data processing unit for performing image preprocessing, feature extraction and optimization processes; The display unit is used to display the detection results and output the location information of the defects.

[0015] The present invention also provides a liquid crystal display screen defect detection medium, comprising a computer-readable storage medium storing program code, wherein the program code can implement a liquid crystal display screen defect detection method when executed by a computer.

[0016] The present invention provides a method, system, device and medium for detecting defects in a liquid crystal display screen. The invention has the following beneficial effects: 1. The present invention uses a multi-scale convolutional neural network (MS-CNN) to perform deep learning processing on LCD screen defects, providing more accurate and detailed defect detection than traditional methods. Compared with the processing method that relies on a single scale in the prior art, this method can effectively capture defect features at different scales, improve the comprehensiveness and robustness of detection, and is particularly suitable for complex and subtle defects.

[0017] 2. The present invention introduces a variational reasoning method to derive the probability distribution of defect areas, thereby improving the adaptability of defect detection. Compared with the traditional thresholding method, this technology based on probability reasoning can adaptively adjust the judgment criteria, reduce the occurrence of false detection and missed detection, and performs well in low contrast and complex backgrounds.

[0018] 3. The present invention automatically adjusts hyperparameters through Bayesian optimization, avoiding the time waste and computational cost caused by traditional manual parameter adjustment. Compared with the common grid search method, Bayesian optimization can find the optimal parameter configuration within a limited number of experiments, improving the overall performance of the model, especially showing higher efficiency and accuracy in complex tasks.

[0019] 4. The present invention deeply explores the geometric and topological characteristics of defects by combining fractal analysis and complex network models. This innovative analysis method can not only reveal the self-similarity and complexity of defects, but also effectively model the spatial relationship between defects, making defect classification more accurate, especially for defects with complex or similar morphology. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the method of the present invention; Figure 2 is a system structure diagram of the present invention; Figure 3 It is a device module diagram of the present invention; Figure 4 This is the medium control diagram of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] Please see attached Figure 1 The embodiment of the present invention provides a method for detecting defects of a liquid crystal display screen, comprising the following steps: S1. Collect defect image data of the LCD screen and mark it; The image data collection and annotation of LCD screen defects provide essential high-quality input for subsequent model training, ensuring that the inspection system can accurately identify and classify various types of defects. Therefore, the quality, type and annotation accuracy of image data are crucial to the final inspection accuracy.

[0023] In order to cover different types of defects, the data collection process needs to be carried out at multiple stages of LCD production, including display production, factory inspection, and possible damage during transportation. High-definition camera equipment (such as high-resolution digital cameras, machine vision cameras) will capture the surface of the LCD and capture various defect characteristics. These images should have sufficient resolution, usually at least 1080p or higher resolution images, to ensure that defects can be clearly captured and accurately calibrated.

[0024] Specifically, the collected image data may include the following types: Bright spot: An area of ​​abnormal brightness that appears on the display screen, usually due to a problem with the arrangement of the liquid crystal molecules in the display panel.

[0025] Dark spot: Due to the failure of the circuit or liquid crystal unit in some areas of the LCD screen, the area cannot display normal light.

[0026] Pixel failure: Some pixels on the display do not respond or show wrong colors, appearing as dead pixels or stuck pixels.

[0027] Color difference: The LCD screen displays inconsistent colors, which may be due to manufacturing errors or aging of the display panel.

[0028] The process of collecting these images needs to ensure the diversity of images, including different backgrounds, different display conditions, and different defect types and locations. In addition, the collection of image data should avoid interference from ambient light to maintain image clarity.

[0029] The collected defect images need to be annotated so that each defect can be identified during subsequent training. The purpose of annotation is to clearly indicate the type and location of the defect in the image. Specifically, image annotation usually includes the following: Defect type annotation: Classify each defect area in the image and annotate it as bright spot, dark spot, pixel failure, color difference, etc.

[0030] Position marking: Use a rectangular box or other marking tools to specify the exact location of the defect area. The coordinates of the rectangular box are usually expressed in pixel coordinates. For example, the coordinates of the upper left corner of the rectangular box are , the coordinates of the lower right corner are ,in To calibrate the position coordinates of the defect area.

[0031] Size annotation: In some embodiments, the size of the defect (e.g., the diameter of a bright spot or the side length of a dark spot) is also annotated to help subsequent models identify defects of different sizes.

[0032] Generally, labeling can be done manually or with semi-automatic labeling tools. Manual labeling is usually done by experts using professional software, while semi-automatic labeling is done manually after preliminary labeling using a preliminarily trained machine learning model. Commonly used labeling tools include LabelImg and RectLabel, which can automatically display the boundaries of defect areas and provide labeling tools. Manual labelers only need to adjust the position and size of the labeling box.

[0033] In a possible implementation, the image is annotated using a rectangular frame.

[0034] For example, for a bright spot defect on an LCD screen, if the bright spot is located in the image =150 to =180 position, =120 to =150, the coordinates of the upper left corner of the annotation box are (150,120), and the coordinates of the lower right corner are (180,150), which clearly points out the position and range of the bright spot defect in the image.

[0035] For other types of defects, such as color difference, the annotation may include the color difference value range and corresponding pixel value of the defect area.

[0036] All annotated images and their corresponding label information will be stored in the data management system. The data management system is responsible for organizing and storing image data, while recording the annotation information of each image (such as defect type, location, size, etc.). The system should also support batch import and export of annotated data for use in training and validating models.

[0037] In some embodiments, in order to improve the consistency and accuracy of image annotation, a quality inspection process may be designed to ensure the correctness of each annotation. This process can review the annotated images to ensure that all annotations meet the preset standards and avoid incorrect annotations caused by human errors.

[0038] In the process of marking LCD screen defects, the coordinates of each marked area are one of the key parameters. For example, the coordinates of the rectangular box Indicates the upper left and lower right corners of the annotation box. For each defect in the annotated image, the parameters of the annotation box are defined as follows: : The horizontal coordinate of the upper left corner of the rectangle, in pixels.

[0039] : The vertical coordinate of the upper left corner of the rectangle, in pixels.

[0040] : The horizontal coordinate of the lower right corner of the rectangle, in pixels.

[0041] : The vertical coordinate of the lower right corner of the rectangle, in pixels.

[0042] In addition, the type of defect is related to its annotation category, which can be one of the following: Bright Spot: A small, bright area, usually caused by irregularities in the liquid crystal arrangement on a display panel or a circuit problem.

[0043] Dark spot: An area in a display panel that is unable to emit light, usually due to the inability of some liquid crystal cells to activate.

[0044] Defective pixels: Some pixels in the display have abnormal colors or are not showing.

[0045] Color difference: Due to inconsistent color of the LCD panel or other reasons, the display screen will have color difference.

[0046] These annotated data provide the necessary label information for the subsequent training of deep learning models, ensuring that the models can learn the characteristics of different defect types and accurately identify them.

[0047] High-resolution images are collected through high-definition imaging equipment, and the type, location and size of defects are calibrated through manual or semi-automatic annotation tools to ensure the accuracy and comprehensiveness of the training data.

[0048] S2, denoising, normalizing and resizing the defect image; Image preprocessing can significantly improve image quality, reduce interference and improve detection accuracy through effective denoising, normalization and resizing.

[0049] Generally speaking, during the image acquisition process, the display image may be affected by noise. Common noise types include Gaussian noise, salt and pepper noise, etc. Noise will reduce the quality of the image and affect subsequent feature extraction and detection tasks. Therefore, denoising is a necessary step.

[0050] Specifically, the commonly used method for denoising is filtering, among which Gaussian filtering is a common image denoising method that can effectively smooth images and remove noise. The Gaussian filter convolves the image with a Gaussian function to smooth the high-frequency noise in the image while retaining the low-frequency information and edge details of the image.

[0051] In one possible implementation, Gaussian filtering can be implemented by the following formula: ; in: Indicates the denoised image at position The pixel value of is the original image at position The pixel value of is the Gaussian kernel function, which represents the weight of the Gaussian distribution. The calculation formula is: ; in: is the standard deviation of the Gaussian kernel, which determines the smoothness of the filter; and is the horizontal and vertical offset relative to the center point in the image coordinate system; is the normalizing constant of the Gaussian distribution; As an option, the Gaussian filter can adjust the standard deviation based on the noise intensity of the image and the processing requirements. The smaller the value. Values ​​are suitable for slight noise, larger The lower value is suitable for removing strong noise.

[0052] Normally, the brightness value of an image may vary due to changes in different devices, environments, and lighting conditions, which can lead to certain brightness and contrast differences between images. In order to eliminate this difference and make the subsequent feature extraction process more consistent, image normalization is particularly important.

[0053] Specifically, normalization is to scale the image pixel values ​​to a uniform range. Usually, the pixel value range of an image is 0 to 255. After normalization, all pixel values ​​can be scaled to the range of 0 to 1, which is convenient for the neural network model to process.

[0054] In one possible implementation, the normalization of the image can be achieved by the following formula: ; in: Indicates the normalized image at position The pixel value of is the original image at position The pixel value of is the minimum value of all pixels in the original image; is the maximum value of all pixels in the original image.

[0055] After normalization, the pixel values ​​in the image are standardized to between 0 and 1, thereby avoiding the impact of changes in image brightness and contrast on the model training process.

[0056] As an option, in addition to the standard linear normalization method, you can also consider performing methods such as histogram equalization or contrast stretching on the image to further optimize the contrast and brightness distribution of the image.

[0057] Generally, the size of LCD screen images is not consistent, which may cause inconsistencies in subsequent processing. In order to ensure that the subsequent deep learning model can receive input images of consistent size, all images need to be resized.

[0058] Specifically, the resizing operation is usually implemented by scaling or cropping. In this embodiment, all images are resized to a uniform size to meet the input requirements of a deep convolutional neural network (CNN).

[0059] In one possible implementation, if the image is large and the defect is located in the central area of ​​the image, the area containing the defect can be extracted by cropping. The cropping area can be determined based on the annotation information. For example, if the defect is located in the middle of the image, the position and size of the cropping box can be determined by calculating the coordinates of the annotation box. The formula for the cropping operation can be expressed as: ; in: Indicates that the cropped image is at position The pixel value of Indicates the starting coordinates of the cropping box, which is determined based on the defect's annotation box.

[0060] As another option, if the image size is small or the defect information will be lost after cropping, a scaling method can be used. Image scaling calculates new pixel values ​​through an interpolation algorithm. Common interpolation algorithms include nearest neighbor interpolation, bilinear interpolation, and bicubic interpolation. In this embodiment, a bilinear interpolation algorithm is used, which estimates the new pixel value by calculating the weighted average of the surrounding four pixels.

[0061] The image is scaled to × ,in and are the width and height of the target size respectively.

[0062] Through image denoising, normalization and resizing, this step provides standardized, denoised and uniform-sized input data for the subsequent deep learning model, greatly improving the accuracy of subsequent feature extraction and defect classification.

[0063] S3, use multi-scale convolutional neural network to extract image features, including local features and global features; After the image denoising, normalization and resizing are completed, the image has a standardized format suitable for feature extraction. At this point, the next step is to use a multi-scale convolutional neural network (MS-CNN) to extract useful features from the image, including local features and global features. These features will provide key information for defect detection and classification.

[0064] In this embodiment, the multi-scale convolutional neural network (MS-CNN) extracts multi-level features in the image, including local detail information and global structural information, through convolution kernels of different sizes. Local features focus on small defects in the image, while global features help the network understand the overall structure of the image and the distribution of defects. By combining multi-scale information, MS-CNN can extract features at multiple scales simultaneously, enhancing the detection capability of the network.

[0065] Generally speaking, defects in LCD screens often manifest as abnormalities in brightness, color or structure in local areas, so the extraction of local features is crucial for identifying small-scale defects. Local features usually involve small areas in the image and can help the network identify small defect information, such as bright spots and dark spots.

[0066] Specifically, the multi-scale convolutional neural network performs convolution operations on the image through multiple convolution kernels of different sizes to extract local features. Small-sized convolution kernels (such as 3×3 or 5×5) are suitable for capturing detailed information in the image, such as small bright spots, dark spots or pixel failures, while larger-sized convolution kernels can capture a wider range of local structural features.

[0067] In this embodiment, the local features are extracted by the following convolution operation formula: ; in: Indicates that the image is at position Extracted local features; is the original image at position The pixel value of It is the weight matrix of the convolution kernel, which represents the weighted relationship between the convolution kernel and the image in the local area; is the radius of the convolution kernel, which determines the size of the convolution kernel. For example, for a 3×3 convolution kernel, =1; , Indicates the location of the offset during the convolution operation.

[0068] The function of this formula is to extract the features of the local area by weighted summing each pixel in the image and its surrounding neighborhood through convolution operation. This local convolution can effectively identify small-scale defects in the image and provide support for subsequent global feature extraction and defect classification.

[0069] Alternatively, defects in LCD screens can be not only local, but also involve structural problems in larger areas or the entire screen. Therefore, extracting global features is crucial to understand the overall structure of the image and large-scale defects.

[0070] Specifically, multi-scale convolutional neural networks capture global features in images by using larger convolution kernels, such as 7 × 7 or 11 × 11. Larger convolution kernels can cover a wider area, thus helping the network understand large-scale defects in images, such as large dark spots, color aberrations, or other types of global structural anomalies.

[0071] In this embodiment, the global feature extraction is also performed through a similar convolution operation, but the convolution kernel size used is larger. By increasing the size of the convolution kernel, the network can extract features in a larger range and identify the overall morphology of large-scale defects. The formula for this operation is similar to the local feature extraction formula, except that the size of the convolution kernel is different, and the formula is: ; in: Represents the global feature extraction result; is the convolution kernel weight used to extract global features; is the radius of the convolution kernel, which is usually large, such as 3 or 5, corresponding to a convolution kernel size of 7×7 or 11×11.

[0072] The function of this formula is to process each pixel in the image and its surrounding area by using a larger convolution kernel to extract large-scale image features. This operation helps to detect large-scale defects in the image, such as large areas of bright spots or color differences.

[0073] Generally speaking, through multi-scale convolutional neural networks, local and global features in images are extracted at multiple scales. These features at different scales can provide multi-level information of the image, thereby improving the accuracy and robustness of the network.

[0074] Specifically, in a multi-scale convolutional neural network, images are processed at different scales through multiple convolution kernels. The feature maps of each scale are assigned different weights according to their importance in defect detection, and finally a comprehensive feature representation is obtained through weighted fusion. By fusing features from different scales, the network can better understand the local details and global structures in the image.

[0075] In this embodiment, features of different scales are fused by weighted summation, and the fused features are used as input of the model for further processing. The fusion formula is as follows: ; in: Represents the fused image features; It is a scale The extracted feature map below; It is a scale The weight coefficients under are dynamically adjusted through training.

[0076] Through this weighted fusion approach, the network is able to extract information from both local and global scales to obtain a comprehensive description of the image and improve the accuracy of defect detection.

[0077] As an alternative, local feature extraction and global feature extraction of multi-scale convolutional neural network effectively captures details and structural information in the image through multi-scale convolution operations. The extraction of local features helps to identify small-scale defects, while the extraction of global features helps the network understand the structure of the entire image, especially when facing large-scale defects. By weighted fusion of features at different scales, MS-CNN can integrate multi-level information and further improve detection accuracy.

[0078] Generally speaking, through this multi-scale feature extraction method, the network can effectively extract diverse features in the image, not only identifying small-scale subtle defects, but also identifying large-scale anomalies in the image.

[0079] S4, weighted fusion of feature maps from different scales to obtain the final image feature representation; The goal of weighted fusion is to combine feature maps extracted at multiple scales so that the information at each scale can be weighted according to its importance, thus ultimately obtaining an image representation with strong feature expression capabilities. By fusing features at different scales, the network can obtain richer feature information at different levels, thereby improving the recognition capability and detection accuracy of different defect types.

[0080] In general, defects in LCD screens are not limited to a specific scale, and defects may show different characteristics at multiple scales. Local features usually reflect small-scale defects in the image, such as bright spots or dark spots, while global features reveal the overall structure of the image and large-scale defects. By weighted fusion of feature maps from different scales, the network can effectively combine local and global information and avoid information loss due to the insufficiency of a single scale.

[0081] Specifically, multi-scale convolutional neural networks extract features by using convolution kernels of different sizes, which makes the feature maps of each scale play a different role in defect recognition. Smaller convolution kernels can capture details, while larger convolution kernels are better at capturing large-scale image information. Weighted fusion of these feature maps from different scales can retain detailed information while enhancing the recognition ability of large-scale structures.

[0082] In this embodiment, weighted fusion is achieved by weighted summation of feature maps of different scales. A weight coefficient is assigned to the feature map of each scale according to its contribution to the defect detection task, and these weight coefficients are usually automatically optimized by the back propagation algorithm during the training process. The weighted fused feature map contains multi-level information from different scales, which provides rich feature expressions for subsequent defect classification and location.

[0083] As an option, the weight factor It can be dynamically adjusted according to the contribution of each scale feature map. Usually, the network learns the importance of different scales in different tasks through training, thereby assigning different weights to each scale. Smaller scale features may be more important for details and local defect detection, while larger scale features may be more important for global structure and large-scale defect recognition.

[0084] Specifically, during the training process, the network calculates the contribution of each scale feature map through the loss function and automatically learns the weight coefficient of each scale through gradient update. These weight coefficients reflect the relative importance of different scale features in the defect detection task.

[0085] In one possible implementation, these weight coefficients can be updated according to the loss value or confidence level. The specific update formula is: ; in: It is a scale The weight coefficient of Is with scale The associated loss value or confidence reflects the performance of the scale feature in the detection task; is a normalization constant used to ensure that the sum of the weights of all scales is 1, and the calculation formula is: ; in: is a normalization constant used to ensure that the sum of all scale weights is 1; Is with scale The associated loss value or confidence indicates the contribution of the scale feature to the inference, which is usually learned through back-propagation; is the total number of scales, indicating the number of scales used in the network; The function of this formula is to calculate the weight coefficient of each scale by exponentially processing the loss or confidence of each scale, and ensure that the sum of all weights is 1 through the normalization constant. In this way, the network can dynamically adjust the importance of each scale in the final feature fusion, further improving the detection ability of the model.

[0086] As an option, by weighted fusion of feature maps from different scales, the network finally obtains a fused feature map, which contains multi-level information from various scales. This information can cover various features from small defects to large-scale structures. Therefore, the fused feature map is richer and more comprehensive than the feature map of a single scale.

[0087] Specifically, the fused feature map can provide multi-angle information support in defect detection tasks. For example, local features help detect bright spots, dark spots, and small-scale pixel failures, while global features help detect large-scale defects such as color difference and surface scratches. Through the weighted fusion of multi-scale feature maps, the network can focus on local and global information in the image at the same time, thereby improving the recognition ability and accuracy of the model.

[0088] In general, through the weighted fusion of multi-scale feature maps, the network can integrate information of different scales, which includes not only detailed local features, but also the global structure of the image and large-scale defects. The fused feature map can provide a richer and more comprehensive feature representation for subsequent defect classification and location, allowing the network to maintain high accuracy and robustness when dealing with complex scenes and diverse defects.

[0089] S5. Infer the defect distribution using variational inference methods; Variational inference is used to estimate the probability distribution of defect regions in an image. Specifically, variational inference derives the posterior distribution of latent variables (i.e., defect regions) through an optimization process. This method can not only infer the location of defects in an image, but also quantify the probability of each region being a defect, thus providing a basis for subsequent defect classification and location.

[0090] In general, variational inference methods are often used to process high-dimensional and complex probabilistic models, especially for inferring the distribution of latent variables. The basic idea is to approximate the true posterior distribution by introducing an approximate variational distribution, thereby converting complex inference problems into more computationally tractable optimization problems. Compared with traditional inference methods, variational inference can be more stable and efficient in processing defect images, especially in the presence of noise, complex backgrounds, and low contrast, and has higher robustness.

[0091] Specifically, the goal of the variational inference method is to minimize the variational free energy to find the approximate distribution that is closest to the true posterior distribution. Variational inference assumes that the defect area (latent variable) and the observed data (image features) are associated through a certain probability model. Therefore, by minimizing the variational free energy, the optimal probability distribution of the defect area can be derived.

[0092] In this embodiment, variational free energy is the core of variational inference, which measures the gap between the variational distribution and the true posterior distribution. By minimizing the variational free energy, we can get the optimal estimate of the defect area in the image.

[0093] The calculation formula of variational free energy is: ; in: is the variational free energy, which represents the approximate distribution The gap between the true posterior distribution and is a variational distribution, which represents the probability distribution of the defect area (latent variable); is the joint probability distribution, representing the observed value (i.e., image features) and latent variables (i.e., defect regions); In the variational distribution Next, the expectation of the joint probability represents the correlation between the image features and the defect area; is a variational distribution The amount of self-information below represents the estimation of the defect area.

[0094] As an alternative, by introducing the variational distribution and the joint probability , variational inference can use the relationship between image features and defect areas to derive the probability distribution of defects during training. represents our initial estimate of the defect area, and minimizing the variational free energy is to optimize this estimate as close as possible to the true posterior distribution.

[0095] Specifically, when performing variational inference, we first need to build a joint probability model ,in is the observed data of the image (i.e., the feature map extracted from the previous step), is the potential variable of the defect area. Then, the optimal distribution of the defect area is derived by optimizing the variational free energy.

[0096] The process of optimizing variational free energy can be carried out by following the steps below: Initialize variational distribution: In order to perform inference, you first need to select an initial variational distribution , usually we choose a simple probability distribution (such as Gaussian distribution) as a preliminary estimate.

[0097] Calculate the expected value: Based on the current variational distribution , calculate the expected value and The two expected values ​​measure the relationship between the image features and the defect area, and the estimation quality of the defect area by the variational distribution, respectively.

[0098] Minimize variational free energy: Use an optimization algorithm such as gradient descent or variational EM to minimize the variational free energy , thereby updating the variational distribution , and obtain the optimal estimate of the defect area.

[0099] Deriving defect distribution: Finally, the optimized variational distribution That is the optimal probability distribution of the defect area, which indicates the possibility that each area in the image is a defect.

[0100] In general, the defect distribution obtained by variational inference method A defect probability is provided for each image region. These probabilities can be post-processed by setting thresholds, and regions with probability values ​​higher than a certain threshold are considered defective regions, thereby achieving defect detection and location.

[0101] Specifically, the inferred defect distribution It can be displayed through image visualization tools. The probability of defect areas can be mapped to the original image to form a visual defect map, which helps to intuitively present the detected defect areas.

[0102] As an option, further post-processing methods (such as morphological operations) can be used to refine the boundaries of the defects and optimize the detection results. In this way, the final detected defect area will be displayed with high accuracy, which is convenient for subsequent analysis and repair.

[0103] Specifically, the variational inference method can effectively derive the probability distribution of defect areas in the image by optimizing the variational free energy.

[0104] S6. Optimize hyperparameters using Bayesian optimization methods; Bayesian optimization predicts the performance of different hyperparameter configurations by building a proxy model (usually using a Gaussian process) and combining known hyperparameter configurations and model performance, and selects the next set of hyperparameters for experimentation based on this prediction. The advantage of this method is that it can find the optimal hyperparameter combination with fewer trials, thereby improving optimization efficiency.

[0105] In general, the hyperparameters of deep learning models, such as learning rate, batch size, network depth, convolution kernel size, etc., will significantly affect the performance of the model. Traditional hyperparameter tuning methods, such as grid search or random search, usually require a lot of computing resources and time, and it is difficult to guarantee the optimal hyperparameter combination. Bayesian optimization greatly improves the optimization efficiency by using known test results to infer the best hyperparameter configuration.

[0106] Specifically, the basic process of Bayesian optimization can be divided into the following steps: Surrogate model construction: Bayesian optimization uses a surrogate model to estimate the relationship between hyperparameter configuration and model performance. Gaussian Process (GP) is usually used as a surrogate model.

[0107] Sampling strategy: Based on the predictions of the current surrogate model, Bayesian optimization selects hyperparameter configurations with high expected improvement (EI) for the next round of experiments.

[0108] Model update: Based on the experimental results, the proxy model is updated and the hyperparameter configuration is continuously optimized through an iterative process.

[0109] Bayesian optimization often uses Gaussian process (GP) as a proxy model. Gaussian process is a non-parametric probability model that can describe the relationship between the objective function (model performance) and the hyperparameter configuration without explicit parameters.

[0110] The Gaussian process is defined as: ; in: Represents the objective function, that is, given the hyperparameter configuration The performance of the model when (e.g., the accuracy of the validation set); is the mean function, usually assumed to be 0, indicating the mean of the objective function; is the covariance function (or kernel function) used to describe the hyperparameter configuration and The correlation between them. Commonly used covariance functions include RBF kernel (radial basis function) and so on.

[0111] As an alternative, the advantage of Gaussian processes is that they can provide the standard deviation (uncertainty) of the predictions, which enables Bayesian optimization to select those hyperparameter configurations with large uncertainties for experimentation, thereby accelerating convergence to the optimal solution.

[0112] Bayesian optimization selects the hyperparameter configuration with the largest expected improvement (EI) for the next round of experiments. The expected improvement measures the potential improvement of a hyperparameter configuration relative to the current optimal solution.

[0113] The formula for expected improvement is: ; in: Indicates in the hyperparameter configuration The objective function value under ; is the currently known optimal objective function value (i.e., the best performance of the model); is the expected improvement, which indicates how much the current hyperparameter configuration improves relative to the optimal configuration.

[0114] Specifically, EI guides the next hyperparameter selection by calculating the potential improvement of the current hyperparameter configuration relative to the known optimal solution. Bayesian optimization selects the hyperparameter configuration that is most likely to provide performance improvement based on the predicted expected improvement value.

[0115] In general, after selecting the hyperparameter configuration with the largest EI for experimentation, Bayesian optimization will obtain new experimental results, which will help update the surrogate model (Gaussian process). The addition of new data enables the surrogate model to more accurately describe the relationship between the hyperparameter configuration and the objective function, thereby providing better predictions for the next round of experiments.

[0116] The updated proxy model can be expressed by the following formula: ; in: is the predicted value of the objective function, that is, the model under a given hyperparameter configuration Performance prediction under is the covariance matrix of the objective function values, which represents the correlation between the objective functions under different hyperparameter configurations.

[0117] As an option, after each experiment, Bayesian optimization updates the surrogate model by minimizing variational free energy and other methods, and gradually converges to the optimal hyperparameter combination by combining the latest hyperparameter configuration information.

[0118] In practical applications, Bayesian optimization is often used to tune the following key hyperparameters of deep learning models: Learning rate: controls the weight update step. Too high a learning rate may lead to unstable training, while too low a learning rate may cause the model to converge too slowly.

[0119] Batch size: The number of samples used in each iteration. A larger batch size can speed up training but may affect the model's generalization ability.

[0120] Convolution kernel size: controls the receptive field of the convolution layer. Smaller convolution kernels help capture detailed information, while larger convolution kernels help capture broader structural information.

[0121] Network depth: controls the number of network layers. Deeper networks can learn more complex features, but may also lead to overfitting.

[0122] In this embodiment, Bayesian optimization can automatically adjust these hyperparameters, thereby improving the detection accuracy of the model under different types of defects.

[0123] Bayesian optimization can significantly improve the performance of deep learning models by effectively adjusting the hyperparameters of the models.

[0124] S7, further analyze the defect characteristics through fractal analysis and complex network model; Fractal analysis evaluates the geometry and self-similarity of defect areas, while complex network models allow us to explore the interrelationships, spatial distribution patterns, and interaction characteristics between defects. These two complementary analysis methods can help us understand defect characteristics more comprehensively and improve the performance of the inspection system through further analysis and optimization.

[0125] Generally speaking, fractal analysis is used to describe and quantify geometric shapes with self-similarity and complex boundaries. In the defect detection task of LCD screens, the defect area usually exhibits complex and irregular shapes, and traditional geometric feature extraction methods are difficult to effectively capture this complexity. Fractal analysis can provide a method to quantify defect morphology and reveal the geometric complexity of the defect area.

[0126] Specifically, the fractal dimension is a measure of the complexity of graphics, describing the geometry and roughness of the defect area. The larger the fractal dimension, the more complex the boundary of the defect area and the more irregular the shape. The box-counting method is usually used to calculate the fractal dimension. The box-counting method covers the defect area with grid boxes of different sizes, calculates the minimum number of boxes required, and analyzes the relationship between the number of boxes and the box size.

[0127] In this embodiment, the fractal dimension is calculated by the following formula: ; in: is the grid size The number of boxes required to cover the defect area; is a constant that represents the proportional relationship between the number of grids and the grid size; is the size of the grid; is the fractal dimension, which represents the geometric complexity of the defect area.

[0128] By calculating the fractal dimension , we can quantify the geometric complexity of the defect area, thus providing effective geometric features for defect classification. This information helps to distinguish different types of defects, especially when the defect types are similar or the morphology is subtle, fractal analysis can provide important geometric distinction basis.

[0129] In general, complex network models are used to describe and analyze systems consisting of nodes and edges, where nodes represent objects and edges represent the relationships between these objects. In defect detection, complex network models can effectively describe the spatial relationships between defect areas and reveal the topological structure and interactions between defects. By modeling defect areas as nodes in the network, the relationships between defects can be represented by edges in the network.

[0130] Specifically, we construct a complex network in which each node represents a defect area and the edges between nodes represent the relationship between defect areas. The relationship between defect areas can be determined by calculating the spatial distance, similarity or other characteristics between them. By analyzing the structural characteristics of the network (such as node degree, clustering coefficient and connectivity), we can gain a deeper understanding of the distribution pattern of defect areas and their interactions.

[0131] In this embodiment, the node degree is the number of connections of each node in the network, which indicates the "influence" of the defect area or the degree of its relationship with other defect areas. The degree calculation formula is: ; in: Is a node The degree of defect area the number of spatial relationships with other defect areas; is the element of the adjacency matrix, representing the defect area and defect areas When the defect area With defect area When there is a relationship between =1, otherwise 0; nnn is the total number of nodes in the network, indicating the number of all defective areas.

[0132] By calculating the topological characteristics of each node in the network, such as degree distribution, clustering coefficient, and network density, we can analyze the correlation and spatial distribution between defect areas, thereby improving the accuracy of defect classification.

[0133] As an option, in some applications, we can also analyze the interaction between defective areas by calculating the shortest path length (ShortestPathLength) of the network. The shortest path length can reflect the connectivity between two defective areas, thereby revealing the closeness of the relationship between the defects.

[0134] Generally speaking, fractal analysis focuses on the geometric complexity of defects, while complex network models focus on the topological relationships and interactions between defects. By combining these two methods, we can deeply analyze defect characteristics from multiple perspectives. This comprehensive analysis helps to improve the recognition accuracy of complex defect types, especially for defects with complex morphology or special spatial distribution.

[0135] Specifically, by combining fractal dimension with structural features such as degree distribution and clustering coefficient of complex networks, we can comprehensively analyze the geometric and topological characteristics of defects. For example, fractal dimension can help us quantify the geometric complexity of defect areas, while degree distribution and clustering coefficient in the network can reveal the spatial correlation between defect areas. By combining these two features, the defect detection model can obtain richer feature information and improve the accuracy of classification and detection.

[0136] By combining fractal analysis and complex network models, this step provides a more detailed and comprehensive analysis method for LCD display defect detection.

[0137] The liquid crystal display defect detection system described below and the liquid crystal display defect detection method described above can refer to each other.

[0138] Please see attached Figure 2 The present invention also provides a liquid crystal display screen defect detection system, comprising: A data acquisition module, used for collecting defect image data of the liquid crystal display screen; Image preprocessing module, used to denoise, normalize and resize defective images; Feature extraction module, used to extract defect features through multi-scale convolutional neural network; An optimization module for inferring defect distributions through variational inference methods and optimizing hyperparameters using Bayesian optimization methods; The analysis module is used to further analyze defect characteristics through fractal analysis and complex network models to improve detection accuracy.

[0139] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, which will not be repeated here.

[0140] The liquid crystal display defect detection device described below and the liquid crystal display defect detection method described above can refer to each other.

[0141] Please see attached Figure 3 The present invention also provides a liquid crystal display screen defect detection device, comprising: An image acquisition device, used to acquire defect images of a liquid crystal display screen; A data processing unit for performing image preprocessing, feature extraction and optimization processes; The display unit is used to display the detection results and output the location information of the defects.

[0142] The device of this embodiment can be used to execute the above method embodiment, and its principles and technical effects are similar, which will not be repeated here.

[0143] The liquid crystal display defect detection medium described below and the liquid crystal display defect detection method described above can be referred to each other.

[0144] Please see attached Figure 4 The present invention also provides a liquid crystal display defect detection medium, comprising a computer-readable storage medium storing a program code, which can implement a liquid crystal display defect detection method when executed by a computer.

[0145] The medium of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, which will not be described in detail here.

[0146] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting defects in a liquid crystal display screen, characterized in that: The following steps are involved: Collect defect image data of LCD screens and annotate them; Performing denoising, normalization and size adjustment on the defective image; Using a multi-scale convolutional neural network to extract image features, the features include local features and global features; Perform weighted fusion of feature maps from different scales to obtain the final image feature representation; Infer defect distribution using variational inference methods; Optimize hyperparameters using Bayesian optimization methods; The defect characteristics are further analyzed through fractal analysis and complex network models.

2. A method for detecting defects in a liquid crystal display according to claim 1, characterized in that: The collecting of defective image data of the liquid crystal display screen comprises: Acquire multiple defect images of the liquid crystal display screen by an image acquisition device; Annotate the image to indicate the type and location of defects.

3. The method for detecting defects in a liquid crystal display screen according to claim 1, characterized in that: The denoising, normalizing and resizing of the defective image comprises: Use image denoising algorithms to remove noise from images; Normalize the brightness and contrast of the image to ensure image consistency; Resize the image to a fixed size to meet the input requirements of subsequent processing modules.

4. The method for detecting defects in a liquid crystal display screen according to claim 1, characterized in that: The multi-scale convolutional neural network extracts image features including: Use multiple convolution kernels of different sizes to extract local features of the image; The feature maps extracted at each scale are weighted and fused to form the final comprehensive feature representation.

5. The method for detecting defects in a liquid crystal display screen according to claim 1, characterized in that: The method of inferring defect distribution using a variational inference method includes: Modeling defective areas in LCD displays as latent variables; The inference process is optimized by the variational inference algorithm to obtain the inference results of defect distribution.

6. The method for detecting defects in a liquid crystal display screen according to claim 1, characterized in that: The use of the Bayesian optimization method to optimize hyperparameters includes: Select the objective function to be optimized and automatically adjust the hyperparameters through the Bayesian optimization algorithm; The Bayesian optimization method is used to maximize the expected value and variance of the objective function, thereby automatically selecting the optimal network hyperparameters.

7. The method for detecting defects in a liquid crystal display screen according to claim 1, characterized in that: The further analysis of defect characteristics by fractal analysis and complex network model includes: Use fractal analysis to calculate the self-similarity of defects and obtain the fractal dimension of defects; Defects are viewed as graph networks, the correlation between defect nodes is calculated, and the relationship between defects is further analyzed through complex network models.

8. A liquid crystal display screen defect detection system, characterized in that: A method for detecting defects in a liquid crystal display screen according to any one of claims 1 to 7, comprising: A data acquisition module, used for collecting defect image data of the liquid crystal display screen; An image preprocessing module, used for denoising, normalizing and resizing the defective image; Feature extraction module, used to extract defect features through multi-scale convolutional neural network; An optimization module for inferring defect distributions through variational inference methods and optimizing hyperparameters using Bayesian optimization methods; The analysis module is used to further analyze defect characteristics through fractal analysis and complex network models to improve detection accuracy.

9. A liquid crystal display screen defect detection device, characterized in that: A method for detecting defects in a liquid crystal display screen according to any one of claims 1 to 7, comprising: An image acquisition device, used to acquire defect images of a liquid crystal display screen; A data processing unit for performing image preprocessing, feature extraction and optimization processes; The display unit is used to display the detection results and output the location information of the defects.

10. A liquid crystal display defect detection medium, characterized in that: The invention comprises a computer-readable storage medium storing program codes, and when the program codes are executed by a computer, a method for detecting defects of a liquid crystal display screen can be realized.

Images