A fast detection method for out-of-control points of display screen pixels

Abstract

The present invention discloses a method for quickly detecting pixel out-of-control points of a display screen, which relates to the technical field of quickly detecting pixel out-of-control points of a display screen. The present invention divides an image by using the periodic parameters of the pixel arrangement of the display screen, and adopts third-order polynomial fitting based on local brightness statistics in each region to realize adaptive background reconstruction and ensure the retention of details in each region; the multi-scale saliency detection introduces additional scales and local structure tensor weighted fusion to improve the insufficient capture of abnormal regions caused by fixed scale combinations and enhance the edge clarity; further, time-domain filtering is combined to eliminate accidental noise, and a pre-trained CNN is used to refine and classify abnormal textures, which overall improves the detection robustness and accuracy and effectively makes up for the disadvantages of traditional methods in background reconstruction and multi-scale processing.

Description

Technical Field

[0001] The present invention relates to the technical field of rapid detection of out-of-control points of display screen pixels, and particularly to a method for rapid detection of out-of-control points of display screen pixels. Background Art

[0002] During the industrial production process of a display screen, the display screen image will have a large local change in background brightness due to uneven lighting conditions and the inherent pixel arrangement characteristics of the device; the traditional background smoothing process has a poor reduction effect on local details, and the contrast of some out-of-control points is insufficient, making it difficult to clearly distinguish from normal areas; in this case, even after adjusting a certain empirical threshold, background noise will still interfere with the detection result, increasing the risk of missed detection; at the same time, most common multi-scale saliency detection methods extract features within a preset scale range and cannot cover abnormal areas with large size changes or local breaks caused by pixel gap distribution; some detection schemes optimize detection by fusing results of different scales, but in local areas with complex backgrounds and weak detail information, it is difficult to balance comprehensiveness and fineness and cannot fully meet the accuracy and real-time requirements of industrial on-line detection.

[0003] In summary, the traditional method for rapid detection of out-of-control points of display screen pixels has problems of incomplete background separation and poor scale matching when dealing with local uneven illumination and variable defect sizes. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] The present invention provides a method for rapid detection of out-of-control points of display screen pixels to solve the problems of incomplete background separation and poor scale matching in the traditional method for rapid detection of out-of-control points of display screen pixels when dealing with local uneven illumination and variable defect sizes.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] An embodiment of the present invention provides a method for rapid detection of out-of-control points of display screen pixels, which includes:

[0008] Step S1, preprocessing the collected display screen image, and the preprocessing steps include:

[0009] Dividing the image into several local sub-regions according to the periodic parameters of the display screen pixel arrangement;

[0010] Extracting low-frequency coefficients from the pixel data in each local sub-region by discrete cosine transform, and performing background reconstruction by using third-order polynomial surface fitting based on the local brightness mean and standard deviation to obtain a background reconstruction image;

[0011] Step S2, perform multi-scale saliency detection on the background reconstructed image; the multi-scale saliency detection includes calculating local pixel saliency values at scales of at least 1 times, 1.5 times, 2 times, 2.5 times, and 3 times respectively, and fusing the saliency maps at each scale into a comprehensive saliency map;

[0012] Step S3, perform local connectivity analysis on the comprehensive saliency map, and the analysis steps include using edge detection and convex hull fitting techniques to perform connectivity processing on abnormal regions to determine the complete abnormal regions formed by pixel runaway points on the display screen;

[0013] Step S4, perform time-domain filtering on multiple frames of display screen images collected continuously. The time-domain filtering corrects the detection results based on the inter-frame difference and local consistency verification strategy, so as to filter out accidental noise and obtain stable detection results.

[0014] As a preferred solution of the fast detection method for pixel runaway points on a display screen according to the present invention, wherein: the division of the local sub-regions is carried out according to the periodic characteristic parameters of the pixel arrangement on the display screen, and the fitting parameters of the polynomial surface fitting are dynamically determined by the pixel brightness statistical data within the local sub-regions.

[0015] As a preferred solution of the fast detection method for pixel runaway points on a display screen according to the present invention, wherein: in step S1, the preprocessing of the display screen image adopts a region division method based on the periodic parameters of pixel arrangement. This method extracts the horizontal direction period parameter and the vertical direction period parameter of the image according to the periodic characteristics of the pixel arrangement in the image, and combines the dynamic adaptation factor and to calculate the size of the local sub-regions, and the formula is:

[0016] ,

[0017] wherein, represents the pixel width in the horizontal direction of the local sub-region, represents the pixel height in the vertical direction of the local sub-region, represents the horizontal direction period parameter of the pixel arrangement on the display screen, represents the vertical direction period parameter of the pixel arrangement on the display screen, represents the dynamic adaptation factor in the horizontal direction, represents the dynamic adaptation factor in the vertical direction, represents the ceiling function,

[0018] Use the overall size of the image and to calculate the number of local sub-regions, and the formula is:

[0019] ,

[0020] wherein, represents the number of local sub-regions in the horizontal direction of the image, represents the number of local sub-regions in the vertical direction of the image, represents the total width of the image, represents the total height of the image, and are respectively the sizes of the local sub-regions, represents the floor function;

[0021] Obtain the dynamic adjustment factor based on the local frequency domain energy distribution, and estimate the main frequency through local Fourier transform and After that, the dynamic factor correction formula is:

[0022] ,

[0023] wherein, represents the dynamic adaptation factor in the horizontal direction, represents the dynamic adaptation factor in the vertical direction, represents the main frequency of the local frequency domain in the horizontal direction, represents the main frequency of the local frequency domain in the vertical direction, and are respectively the horizontal and vertical periodic parameters of the pixel arrangement of the display screen, represents the minimum value function.

[0024] As a preferred solution of the method for quickly detecting pixel out-of-control points of a display screen according to the present invention, wherein: in the multi-scale saliency detection, the fusion of saliency maps at each scale adopts a weighted average method, and the weighting coefficients are determined based on local structure tensor analysis to enhance the clarity of the edges of out-of-control points.

[0025] As a preferred solution of the method for quickly detecting pixel out-of-control points of a display screen according to the present invention, wherein: in step S2, multi-scale saliency detection is performed on the background reconstructed image, local pixel saliency values are calculated using a multi-scale window, and then the weighting coefficients at each scale are determined by combining local structure tensor analysis, and finally fused into a comprehensive saliency map.

[0026] As a preferred solution of the method for quickly detecting pixel out-of-control points of a display screen according to the present invention, wherein: in step S2, the multi-scale saliency detection method is:

[0027] At each scale under, calculate the local contrast based on the background reconstructed image , and the calculation formula of the local saliency value is:

[0028] ,

[0029] wherein, represents the local pixel saliency value at scale , represents the pixel value of the background reconstruction image at position , represents the mean value of the pixels within a window centered at with a size of , represents the standard deviation of the pixels within the same window, is a small constant introduced to avoid division by zero, is the scale , and the local window size is defined as: , wherein, is the basic window size, is the scale, reflecting the linear expansion relationship of the window with the scale change;

[0030] Local structure tensor analysis is introduced to determine the weighting coefficient. For scale , the local structure tensor is defined as:

[0031] ,

[0032] wherein, represents the local structure tensor at scale at position , and respectively represent the gradient values of the image at position along the horizontal and vertical directions, respectively represent the horizontal and vertical pixel indices within the local neighborhood region;

[0033] is a neighborhood region centered at with a size of ;

[0034] For , solve for the eigenvalues. Let and be its eigenvalues, satisfying , and the anisotropy measure is defined as:

[0035] ,

[0036] wherein, represents the anisotropy measure at scale .

[0037] As a preferred solution of the method for quickly detecting out-of-control points of display screen pixels according to the present invention, wherein: in step S2, the method of fusing the saliency maps of each scale into a comprehensive saliency map is as follows:

[0038] Based on the anisotropic measure, the scale weighting coefficient is determined using an exponential function, and the formula is:

[0039] ,

[0040] wherein, represents the weighting coefficient at scale , is a parameter for controlling the weighting sensitivity, and its value adjusts the steepness of the exponential response;

[0041] The fusion of the saliency maps of each scale adopts weighted average, and the formula for the comprehensive saliency map is:

[0042] ,

[0043] wherein, is the fused comprehensive saliency map;

[0044] After background reconstruction, insufficient local contrast may mask abnormal features. Through the above multi-scale detection, windows of different scales capture local fine and global features respectively. Larger windows make up for the lack of information in low-contrast regions, while smaller windows retain local details. The weighting coefficients provided by local structure tensor analysis further strengthen edge information, making abnormal regions show higher responses in the comprehensive saliency map, and fully extracting and enhancing abnormal features in the image at multiple scales.

[0045] As a preferred solution of the method for quickly detecting out-of-control points of display screen pixels according to the present invention, wherein: the detection method further includes using a pre-trained convolutional neural network to classify the local texture features of the abnormal region obtained in step S3 through transfer learning to further correct the detection result.

[0046] As a preferred solution of the method for quickly detecting out-of-control points of display screen pixels according to the present invention, wherein: in step S4, the pre-trained convolutional neural network classifies the local texture features of the abnormal region obtained in step S3 through transfer learning;

[0047] A network structure combining a feature extraction module and a classifier is adopted. The pre-trained weights are from the ImageNet dataset, and after fine-tuning on the display screen abnormal texture data, the texture of the abnormal region is discriminated and corrected.

[0048] As a preferred solution of the method for quickly detecting out-of-control points of display screen pixels according to the present invention, wherein: in step S4, the steps of classifying the local texture features are as follows:

[0049] Abnormal area image As input, it is processed by a feature extraction module composed of multiple layers of convolution and pooling. The formula is:

[0050] ,

[0051] where represents the feature map output by the th layer, represents the activation function, represents the convolution kernel weights of the th layer, represents the bias of the th layer, represents the convolution operation, represents the input feature map of the th layer, where ;

[0052] After layers of feature extraction, the final feature map is flattened to form a vector and then input into a fully connected classifier. The formula is:

[0053] ,

[0054] where represents the classification vector output by the fully connected layer, represents the weight matrix of the fully connected layer, represents the vector obtained by flattening the feature map output by the th layer, represents the bias of the fully connected layer;

[0055] The classification vector is converted into a class probability distribution using the softmax function. The formula is:

[0056] ,

[0057] where represents the probability of the th class, represents the th element in the classification vector , represents the exponential function;

[0058] The pre-trained weights are from the ImageNet dataset, which has accumulated rich feature representation capabilities in large-scale image classification tasks. Through transfer learning, the feature extraction module of the pre-trained model is introduced into the abnormal area texture feature classification task and fine-tuned for display screen abnormal data, so that the network is more suitable for the discriminant task of local texture;

[0059] The saliency map fused previously is corrected using the anomaly class probability output by the CNN, and the correction formula is:

[0060] , where represents the corrected saliency map, represents the comprehensive saliency map fused in step S2, represents the anomaly class probability output by the CNN.

[0061] The beneficial effects of the present invention are as follows: The present invention divides an image using the periodic parameters of the display screen pixel arrangement, and adopts third-order polynomial fitting based on local brightness statistics in each region to achieve adaptive background reconstruction, ensuring the retention of details in each region; The multi-scale saliency detection introduces additional scales and local structure tensor weighted fusion, improving the insufficient capture of abnormal regions caused by fixed scale combinations and enhancing the edge sharpness; Further combined with time-domain filtering to eliminate accidental noise, and using a pre-trained CNN to refine and classify abnormal textures, the overall detection robustness and accuracy are improved, effectively making up for the disadvantages of traditional methods in background reconstruction and multi-scale processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.

[0063] Figure 1 It is a flowchart of the method for quickly detecting pixel out-of-control points on the display screen in Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0064] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings of the specification.

[0065] Many specific details are set forth in the following description to facilitate a thorough understanding of the present invention, but the present invention may be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0066] Second, the "one embodiment" or "embodiment" referred to herein means a specific feature, structure, or characteristic that may be included in at least one implementation manner of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that exclude each other with other embodiments.

[0067] Embodiment 1, referring to Figure 1 , this embodiment provides a method for quickly detecting out-of-control points of display screen pixels, including the following steps:

[0068] Step S1, preprocess the collected display screen image, and the preprocessing steps include:

[0069] Divide the image into several local sub-regions according to the periodic parameters of the display screen pixel arrangement;

[0070] Extract the low-frequency coefficients of the pixel data in each local sub-region by using the discrete cosine transform, and perform background reconstruction by using the third-order polynomial surface fitting based on the local brightness mean and standard deviation to obtain the background reconstruction image;

[0071] The division of the local sub-regions is carried out according to the periodic characteristic parameters of the display screen pixel arrangement, and the fitting parameters of the polynomial surface fitting are dynamically determined by the pixel brightness statistical data in the local sub-regions;

[0072] In step S1, the preprocessing of the display screen image adopts a region division method based on the periodic parameters of the pixel arrangement. This method extracts the horizontal direction period parameter and the vertical direction period parameter of the image according to the periodic characteristics of the pixel arrangement in the image, and combines the dynamic adaptation factor and to calculate the size of the local sub-region. The formula is:

[0073] ,

[0074] where, represents the pixel width of the local sub-region in the horizontal direction, represents the pixel height of the local sub-region in the vertical direction, represents the horizontal direction period parameter of the display screen pixel arrangement, represents the vertical direction period parameter of the display screen pixel arrangement, represents the dynamic adaptation factor in the horizontal direction, represents the dynamic adaptation factor in the vertical direction, represents the ceiling function,

[0075] Use the overall size of the image and to calculate the number of local sub-regions. The formula is:

[0076] ,

[0077] Among them, represents the number of local sub-regions in the horizontal direction of the image, represents the number of local sub-regions in the vertical direction of the image, represents the total width of the image, represents the total height of the image, and are respectively the sizes of the local sub-regions, represents the floor function;

[0078] Obtain the dynamic adjustment factor based on the local frequency domain energy distribution, and estimate the main frequency through local Fourier transform and After that, the dynamic factor correction formula is:

[0079] ,

[0080] Among them, represents the dynamic adaptation factor in the horizontal direction, represents the dynamic adaptation factor in the vertical direction, represents the main frequency of the local frequency domain in the horizontal direction, represents the main frequency of the local frequency domain in the vertical direction, and are respectively the horizontal and vertical periodic parameters of the pixel arrangement of the display screen, represents the minimum value function;

[0081] Specifically, use the periodic parameters of the pixel arrangement of the display screen to divide the image into regions, and construct a mathematical model reflecting the local structural characteristics;

[0082] When calculating the size of the local sub-region, introduce local frequency domain information through the dynamic adaptation factor to achieve adaptive division of different display screen structures, reasonably allocate image resources for the number of regions, so that each local region can maintain sufficient details while covering the overall characteristics; the adjustment of the dynamic factor improves the compatibility of the model with various display screens;

[0083] Step S2, perform multi-scale saliency detection on the background reconstructed image; the multi-scale saliency detection includes calculating the local pixel saliency values at least at scales of 1 times, 1.5 times, 2 times, 2.5 times, and 3 times respectively, and fusing the saliency maps at each scale into a comprehensive saliency map;

[0084] In the multi-scale saliency detection, the fusion of the saliency maps at each scale adopts a weighted average method, and the weighting coefficients are determined based on local structure tensor analysis to enhance the clarity of the edges of the runaway points;

[0085] In step S2, multi-scale saliency detection is performed on the background reconstructed image. The local pixel saliency value is calculated using a multi-scale window, and then the weighted coefficients at each scale are determined by combining local structure tensor analysis. Finally, they are fused into a comprehensive saliency map.

[0086] In step S2, the multi-scale saliency detection method is as follows:

[0087] At each scale based on the background reconstructed image calculate the local contrast. The formula for the local saliency value is:

[0088] ,

[0089] where represents the local pixel saliency value at scale , represents the pixel value of the background reconstructed image at position , represents the mean value of the pixels within a window centered at with a size of , represents the standard deviation of the pixels within the same window, is a small constant introduced to avoid division by zero, is the scale and the local window size at scale is defined as: where is the basic window size,

[0090] Local structure tensor analysis is introduced to determine the weighted coefficients. For scale , the local structure tensor is defined as:

[0091] ,

[0092] where represents the local structure tensor at scale at position , and respectively represent the gradient values of the image at position along the horizontal and vertical directions, respectively represent the horizontal and vertical pixel indices within the local neighborhood region;

[0093] is the neighborhood region centered at with a size of ;

[0094] For solving the eigenvalues, let and be its eigenvalues, satisfying , and define the anisotropy measure as:

[0095] ,

[0096] where represents the anisotropy measure at scale ;

[0097] In step S2, the method of fusing the saliency maps at each scale into a comprehensive saliency map is:

[0098] Based on the anisotropy measure, use the exponential function to determine the scale weighting coefficient, and the formula is:

[0099] ,

[0100] where represents the weighting coefficient at scale , is a parameter for controlling the weighting sensitivity, and its value adjusts the steepness of the exponential response;

[0101] The fusion of the saliency maps at each scale adopts weighted averaging, and the formula for the comprehensive saliency map is:

[0102] ,

[0103] where is the fused comprehensive saliency map;

[0104] After background reconstruction, insufficient local contrast may mask abnormal features. Through the above multi-scale detection, windows at different scales capture local fine and global features respectively. Larger windows make up for the lack of information in low-contrast areas, while smaller windows retain local details. The weighting coefficients provided by local structure tensor analysis further enhance edge information, making abnormal regions show higher responses in the comprehensive saliency map, and fully extracting and enhancing abnormal features in the image at multiple scales;

[0105] Specifically, through multi-scale saliency detection, the abnormal features in the background-reconstructed image are magnified layer by layer. At each scale, the local mean and standard deviation are used to reflect the pixel contrast situation, and subtle abnormalities can be detected in low-contrast areas. Introducing local structure tensors improves the problem of edge blurring, enabling effective highlighting of abnormal edges after weighting at each scale. Multi-scale weighted fusion makes the detection results take into account both overall information and local details, thus having higher adaptability and robustness in images on different displays;

[0106] Step S3: Perform local connectivity analysis on the comprehensive saliency map. The analysis steps include using edge detection and convex hull fitting techniques to perform connectivity processing on the abnormal region to determine the complete abnormal region formed by the out-of-control points of the display screen pixels;

[0107] Step S4: Perform time-domain filtering on multiple consecutive frames of display screen images. The time-domain filtering corrects the detection results based on the inter-frame difference and local consistency verification strategy, so as to filter out accidental noises and obtain stable detection results;

[0108] The detection method further includes using a pre-trained convolutional neural network to perform local texture feature classification on the abnormal region obtained in step S3 through transfer learning to further correct the detection results;

[0109] In step S4, the pre-trained convolutional neural network performs local texture feature classification on the abnormal region obtained in step S3 through transfer learning;

[0110] Adopt a network structure combining a feature extraction module and a classifier. The pre-trained weights are from the ImageNet dataset and are fine-tuned on the display screen abnormal texture data to discriminate and correct the abnormal region texture;

[0111] In step S4, the steps for performing local texture feature classification are as follows:

[0112] The abnormal region image is used as the input and processed through a feature extraction module composed of multiple layers of convolution and pooling. The formula is:

[0113] ,

[0114] where, represents the feature map output by the th layer, represents the activation function, represents the convolution kernel weights of the th layer, represents the bias of the th layer, represents the convolution operation, represents the input feature map of the th layer, where ;

[0115] After layers of feature extraction, the final feature map is flattened to form a vector and then input into a fully connected classifier. The formula is:

[0116] ,

[0117] where, Represents the classification vector output by the fully connected layer, Represents the weight matrix of the fully connected layer, Represents the Vector obtained by flattening the feature map output by the layer;

[0118] The classification vector is converted into a class probability distribution using the softmax function. The formula is:

[0119] ,

[0120] where, Represents the probability of the th class, Represents the classification vector and the th element in it, Represents the exponential function;

[0121] The pre-trained weights are from the ImageNet dataset, which has accumulated rich feature representation capabilities in large-scale image classification tasks. Through transfer learning, the feature extraction module of the pre-trained model is introduced into the abnormal region texture feature classification task, and fine-tuning is performed on the abnormal data of the display screen, so that the network is more suitable for the discriminant task of local texture;

[0122] The abnormal class probability output by the CNN is used to correct the previously fused saliency map. The correction formula is:

[0123] , where, Represents the corrected saliency map, Represents the comprehensive saliency map fused in step S2, Represents the abnormal class probability output by the CNN;

[0124] Specifically, a network composed of multiple convolutional and fully connected layers is constructed to achieve refined classification of the local texture of the abnormal region. The weights obtained by the pre-trained model on the large-scale data provide a good initial basis for feature extraction. After fine-tuning, the model is more suitable for the characteristics of the abnormal texture of the display screen. The class probability generated by the softmax function can be used to quantify the credibility of the abnormal region, and then multiplied by the previously fused saliency map to correct possible false detections and missed detections.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limitations. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for quickly detecting out-of-control points of display screen pixels, characterized in that: Including, Step S1, preprocess the collected display screen images. The preprocessing steps include: Period parameters in the horizontal direction of the arrangement of the image according to the display screen pixels and period parameters in the vertical direction are divided into a number of local sub-regions; Extract low-frequency coefficients from the pixel data in each local region using discrete cosine transform, and perform background reconstruction using third-order polynomial surface fitting based on the local brightness mean and standard deviation to obtain a background reconstruction image; Step S2, perform multi-scale saliency detection on the background reconstruction image; the multi-scale saliency detection includes calculating local pixel saliency values at least at scales of 1 times, 1.5 times, 2 times, 2.5 times, and 3 times respectively, and fusing the saliency maps at each scale into a comprehensive saliency map; Step S3, perform local connectivity analysis on the comprehensive saliency map. The analysis steps include using edge detection and convex hull fitting techniques to perform connectivity processing on the abnormal region to determine the complete abnormal region formed by the pixel out-of-control points on the display screen; Step S4, perform temporal filtering on multiple consecutive frames of display screen images collected. The temporal filtering corrects the classification results of the abnormal regions based on the classification results of a pre-trained convolutional neural network CNN.

2. The rapid detection method for out-of-control points of display screen pixels according to claim 1, characterized in that: The division of the local sub-regions is based on the horizontal direction period parameter of the display screen pixel arrangement and the vertical direction period parameter to carry out, and the fitting parameters of the polynomial surface fitting are dynamically determined by the pixel brightness statistical data within the local sub-regions.

3. The rapid detection method for out-of-control points of display screen pixels according to claim 2, wherein: In step S1, the preprocessing of the display screen image adopts a region division method based on the periodic parameters of pixel arrangement. This method extracts the horizontal direction period parameter and the vertical direction period parameter , and combines the dynamic adaptation factor and to calculate the local sub-region size. The formula is: ， Among them, represents the pixel width in the horizontal direction of the local sub-region, represents the pixel height in the vertical direction of the local sub-region, represents the horizontal periodic parameter of the pixel arrangement of the display screen, represents the vertical periodic parameter of the pixel arrangement of the display screen, represents the dynamic adaptation factor in the horizontal direction, represents the dynamic adaptation factor in the vertical direction, represents the ceiling function, Using the overall size of the image and , calculate the number of local sub-regions. The formula is as follows: ， Among them, represents the number of local sub-regions in the horizontal direction of the image, represents the number of local sub-regions in the vertical direction of the image, represents the total width of the image, represents the total height of the image, and are the sizes of the local sub-regions respectively, represents the floor function; Obtain a dynamic adjustment factor based on the local frequency domain energy distribution, and estimate the main frequency through local Fourier transform and After that, the dynamic factor correction formula is as follows: ， Among them, represents the dynamic adaptation factor in the horizontal direction, represents the dynamic adaptation factor in the vertical direction, represents the main frequency of the local frequency domain in the horizontal direction, represents the main frequency of the local frequency domain in the vertical direction, and are respectively the period parameters in the horizontal and vertical directions of the pixel arrangement on the display screen, represents the minimum value function.

4. A method for quickly detecting out-of-control points of display screen pixels according to claim 1, characterized in that: In the multi-scale saliency detection, the fusion of the saliency maps at each scale adopts a weighted average method, and the weighting coefficients are determined based on local structure tensor analysis.

5. A method for quickly detecting out-of-control points of display screen pixels according to claim 1, characterized in that: In step S2, perform multi-scale saliency detection on the background reconstruction image, calculate local pixel saliency values using a multi-scale window, then determine the weighting coefficients at each scale in combination with local structure tensor analysis, and finally fuse them into a comprehensive saliency map.

6. The rapid detection method for out-of-control points of display screen pixels according to claim 5, characterized in that: In step S2, the multi-scale saliency detection method is: At each scale Reconstruct the image according to the background Calculate the local contrast. The calculation formula for the local saliency value is as follows: ， Among them, represents the local pixel saliency value at scale , represents the pixel value of the background reconstruction image at position , represents the mean value of the pixels within the window centered at with a size of , represents the standard deviation of the pixels within the same window, is a small constant introduced to avoid division by zero, is the scale , and the local window size at scale is defined as: , where is the basic window size, and is the scale, reflecting the linear expansion relationship of the window with the change of scale; Introduce local structure tensor analysis to determine the weighting coefficient for the scale , and the local structure tensor is defined as: ， Among them, represents the scale at the position of the local structure tensor, and respectively represent the image gradient values in the horizontal and vertical directions at the position respectively represent the horizontal and vertical pixel indices within the local neighborhood region; centered at with a size of neighborhood region; For solving the eigenvalues, let and be its eigenvalues, satisfying , and define the anisotropy measure as: ， Among them, represents the anisotropy measure at the scale.

7. The rapid detection method for out-of-control points of display screen pixels according to claim 6, wherein: In step S2, the method of fusing the saliency maps at each scale into a comprehensive saliency map is: Based on the anisotropy measure, use the exponential function to determine the scale weighting coefficients. The formula is: ， Among them, represents the weighting coefficient at scale , is a parameter for controlling the weighting sensitivity, and its value adjusts the steepness of the exponential response; The fusion of the saliency maps at each scale adopts weighted average, and the formula for the comprehensive saliency map is: ， Among them, is the fused comprehensive saliency map.

8. The rapid detection method for out-of-control points of display screen pixels according to claim 1, wherein The detection method further includes classifying the local texture features of the abnormal region obtained in step S3 using a pre-trained convolutional neural network through transfer learning.

9. A method for quickly detecting out-of-control points of display screen pixels according to claim 8, characterized in that: In step S4, the pre-trained convolutional neural network classifies the local texture features of the abnormal region obtained in step S3 through transfer learning; Adopt a network structure combining a feature extraction module and a classifier. The pre-trained weights are from the ImageNet dataset and are fine-tuned on the display screen abnormal texture data to discriminate and correct the abnormal region texture.

10. A method for quickly detecting out-of-control points of display screen pixels according to claim 9, characterized in that: In step S4, the steps of classifying the local texture features are: Abnormal area image As input, it is processed by a feature extraction module composed of multiple layers of convolution and pooling, and the formula is: ， Among them, represents the feature map output by the th layer, represents the activation function, represents the convolutional kernel weights of the th layer, represents the bias of the th layer, represents the convolutional operation, represents the input feature map of the th layer, where ; After layer feature extraction, the final feature map is flattened, and after forming a vector, it is input into a fully connected classifier. The formula is as follows: ， Among them, represents the classification vector output by the fully connected layer, represents the weight matrix of the fully connected layer, represents the vector after flattening the feature map output by the layer, and represents the bias of the fully connected layer; Use the softmax function to convert the classification vector into a class probability distribution. The formula is: ， Among them, represents the probability of the th category, represents the th element in the classification vector; Use the abnormal class probability output by the CNN to correct the previously fused saliency map. The correction formula is: , where represents the corrected saliency map, represents the comprehensive saliency map after fusion in step S2, represents the probability of the abnormal category output by the CNN.

Images