Display screen defect detection method and system based on visual perception

Abstract

The present invention relates to the field of defect detection, and in particular to a display screen defect detection method and system based on visual perception. A display screen defect detection system based on visual perception comprises: a test image acquisition module, a standard image storage module, a scattered light analysis module, an image distortion repair module, a difference analysis module and a defect analysis module. The image distortion repair model in the present invention can effectively eliminate distortion problems caused by factors such as camera settings and environmental influences, and ensure the quality of the display image to be tested; by analyzing the corresponding difference feature map between the display image to be tested and its corresponding standard image, the defects in the display performance of the display screen can be effectively detected.

Description

Technical Field

[0001] The present invention relates to the field of defect detection, and in particular to a display screen defect detection method and system based on visual perception. Background Art

[0002] In today's digital age, display screens, as the core carrier of information presentation, play a vital role in all walks of life. Whether it is a smartphone, tablet, TV, computer monitor or advertising screen, high-quality display performance is essential to ensure the superiority of the end-user experience. With the continuous advancement of technology and the improvement of consumer expectations, the display quality requirements for display screens are also getting higher and higher, including color accuracy, brightness, contrast and defect-free image presentation.

[0003] However, during the production process, due to problems such as material inconsistency, errors in the manufacturing process, and assembly accuracy, various defects may occur on the display screen, such as bright spots, dark spots, and uneven colors. These defects seriously affect the overall quality of the display screen and user experience.

[0004] Traditional defect detection methods mainly rely on manual visual inspection, but this method is not only time-consuming and labor-intensive, but the detection results are also affected by the operator's experience and subjective judgment, resulting in poor consistency and repeatability of detection. In addition, as the resolution of display screens increases and the size increases, the difficulty and complexity of manual inspection also increase. Summary of the invention

[0005] The present invention uses a high-resolution camera to take a picture of a display screen to obtain the response of the display screen to an input standard image; next, by performing scattered light analysis on the image to be tested, the generated scattering feature map of the image to be tested can enhance the image features of the image to be tested in the display screen; secondly, the image distortion repair model can effectively eliminate the distortion problems caused by factors such as camera settings and environmental influences, ensure the quality of the displayed image to be tested, and thus provide a more accurate data basis for subsequent defect detection; finally, by analyzing the corresponding difference feature map between the displayed image to be tested and its corresponding standard image, the defects in the display performance of the display screen can be effectively detected.

[0006] A display screen defect detection method based on visual perception, comprising:

[0007] Acquire the image to be tested presented by the display screen in response to the standard image;

[0008] Performing scattered light analysis on the image to be tested to construct a scattering characteristic map of the image to be tested;

[0009] The image to be tested and the scattering characteristic map of the image to be tested are sent to the image distortion restoration model for processing, and the display image to be tested is output;

[0010] Perform difference calculation and normalization operation on the display image to be tested and its corresponding standard image, and output a difference feature map;

[0011] All difference feature maps are spliced ​​according to channels to construct the display feature map to be tested, the display feature map to be tested is subjected to defect analysis, and the display defect detection result is output.

[0012] As a preferred embodiment of the present invention, the image distortion restoration model is established based on the U-net network, including five encoding blocks, respectively recorded as the first encoding block, the second encoding block, the third encoding block, the fourth encoding block and the fifth encoding block, four decoding blocks, respectively recorded as the first decoding block, the second decoding block, the third decoding block and the fourth decoding block, four multi-scale fusion blocks, respectively recorded as the first multi-scale fusion block, the second multi-scale fusion block, the third multi-scale fusion block and the fourth multi-scale fusion block, and a self-attention mechanism block, and the encoding block is used to extract features of the image to be tested, and the decoding block is used to restore the image resolution; the multi-scale fusion block is used to perform feature fusion; the self-attention mechanism block is used to perform a self-attention mechanism operation through the scattering feature map of the image to be tested.

[0013] As a preferred embodiment of the present invention, the image to be tested is sent to the image distortion restoration model for processing, and the displayed image to be tested is output, which specifically includes the following steps:

[0014] The image to be tested is processed by the first encoding block, the second encoding block, the third encoding block, the fourth encoding block and the fifth encoding block in sequence, and the first feature map, the second feature map, the third feature map, the fourth feature map and the fifth feature map are output. The fourth feature map is processed by the fourth multi-scale fusion block to obtain the sixth feature map, and the fifth feature map is fused with the sixth feature map and processed by the fourth decoding block to obtain the seventh feature map; the third feature map is processed by the third multi-scale fusion block to obtain the eighth feature map, and the seventh feature map is fused with the eighth feature map and processed by the third decoding block to obtain the ninth feature map; the second feature map is processed by the second multi-scale fusion block to obtain the tenth feature map, and the ninth feature map is fused with the tenth feature map and processed by the second decoding block to obtain the eleventh feature map; the first feature map is processed by the first multi-scale fusion block to obtain the twelfth feature map, and the twelfth feature map is fused with the thirteenth feature map and processed by the first decoding block to obtain the thirteenth feature map; the thirteenth feature map and the scattering feature map of the image to be tested are processed by the self-attention mechanism block to output the display image to be tested.

[0015] As a preferred embodiment of the present invention, in the multi-scale fusion block, the following operations are specifically performed:

[0016] The feature map input to the multi-scale fusion block is recorded as the input feature map; the feature map output by the multi-scale fusion block is recorded as the output feature map;

[0017] The input feature map undergoes a dilation factor of n 1 The first feature map to be fused is obtained by performing a dilated convolution operation with a coefficient of n. The first fused feature map is subjected to a global average pooling operation to obtain a second feature map to be fused. The first feature map to be fused is subjected to a dilation operation with a coefficient of n. 2 The third feature map to be fused is fused with the first feature map to be fused, and a convolution operation is performed, the convolution kernel can be "1×1", and the fourth feature map to be fused is obtained; the fourth feature map to be fused is expanded once with a coefficient of n 3 The fifth feature map to be fused is obtained by performing a dilated convolution operation with a coefficient of n. The fifth feature map to be fused is fused with the first feature map to be fused, and a convolution operation is performed to obtain a sixth feature map to be fused. The sixth feature map to be fused is subjected to a dilation operation with a coefficient of n. 4 The hole convolution operation is performed to obtain the seventh feature map to be fused; the second feature map to be fused, the third feature map to be fused, the fifth feature map to be fused and the seventh feature map to be fused are spliced ​​according to the channel, and a convolution operation is performed once to obtain an output feature map; and {n 1 , n 2 , n 3 , n 4} is the expansion coefficient set, which is obtained by simulation calculation through the swarm optimization algorithm.

[0018] As a preferred embodiment of the present invention, in the self-attention mechanism block, the following operations are specifically performed:

[0019] The thirteenth feature map is multiplied by the key weight matrix and the value weight matrix respectively to obtain the key feature map K and the value feature map V; the scattering feature map of the image to be tested is multiplied by the query weight matrix to obtain the query feature map Q; the attention weight matrix ATT is calculated based on the key feature map K and the query feature map Q, ATT=softmax[QK T / (d) 0.5 ], where d is the dimension of the query feature map Q; finally, the attention weight matrix ATT is multiplied by the value feature map V to obtain the display image to be tested.

[0020] As a preferred embodiment of the present invention, the training of the image distortion restoration model includes the following steps:

[0021] Obtain several images to be tested and their corresponding scattering feature maps of the images to be tested, and annotate the images to be tested and their corresponding scattering feature maps of the images to be tested through standard images. All the annotated images to be tested and their corresponding scattering feature maps of the images to be tested form a training set, and then send the training set to the image distortion restoration model with parameter initialization for training. Take the standard image as the target and calculate the loss value. If the training conditions are met, output the trained image distortion restoration model; if the training conditions are not met, continue to train the image distortion restoration model through the training set.

[0022] A display screen defect detection system based on visual perception, comprising:

[0023] The test image acquisition module is used to acquire the test image presented by the display screen in response to the standard image through a high-resolution camera device;

[0024] A standard image storage module, used for storing standard images;

[0025] A scattered light analysis module is used to perform scattered light analysis on the image to be tested to construct a scattering characteristic map of the image to be tested;

[0026] An image distortion restoration module is used to send the image to be tested and the scattering characteristic map of the image to be tested into the image distortion restoration model for processing, and output the display image to be tested;

[0027] A difference analysis module is used to perform difference calculation and normalization operations on the display image to be tested and its corresponding standard image, and output a difference feature map;

[0028] The defect analysis module is used to splice all the difference feature maps according to the channels, construct the display feature map to be tested, perform defect analysis on the display feature map to be tested, and output the display defect detection results.

[0029] The present invention has the following advantages:

[0030] 1. The present invention uses a high-resolution camera to take a picture of a display screen to obtain the response of the display screen to an input standard image; next, by performing scattered light analysis on the image to be tested, the generated scattering feature map of the image to be tested can enhance the image features of the image to be tested in the display screen; secondly, the image distortion repair model can effectively eliminate the distortion problems caused by factors such as camera settings and environmental influences, thereby ensuring the quality of the displayed image to be tested, thereby providing a more accurate data basis for subsequent defect detection; finally, by analyzing the corresponding difference feature map between the displayed image to be tested and its corresponding standard image, the defects in the display performance of the display screen can be effectively detected.

[0031] 2. The present invention extracts features of the image to be tested through the encoding block, can learn the deep-level features corresponding to the image in the display screen, and then analyzes the deep-level features corresponding to the image in the display screen through the decoding block, and outputs the actual image corresponding to the image to be tested, thereby eliminating distortion problems caused by factors such as camera settings and environmental influences, thereby improving the accuracy of subsequent display screen defect detection.

[0032] 3. The multi-scale fusion block provided in the present invention can capture receptive field information of different scales, and fuse the receptive field information of different scales, which can effectively capture multi-scale contextual information with high coverage and improve the degree of learning in the feature extraction process.

[0033] 4. The present invention enhances the features of the thirteenth feature map through the self-attention mechanism and the scattering feature map of the image to be tested, and can enhance the image in the display screen according to the scattering optical characteristic information contained in the scattering feature map of the image to be tested, highlight the feature information corresponding to the image in the display screen, and thereby improve the effect of subsequent image distortion repair. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the structure of a display screen defect detection system based on visual perception adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0036] Example 1

[0037] A display screen defect detection method based on visual perception, comprising:

[0038] The image to be tested presented by the display screen in response to the standard image is obtained by a high-resolution camera device. It should be noted that in the process of performing display screen defect detection, the display screen will be started and a series of standard images will be input to the display screen. The standard image can be set by the operator. For example, a pure color image can be used to evaluate the color accuracy and saturation of the display screen. The smoothness and color performance of the color transition when the display screen is displayed can be detected by a gradient image. The display screen with the input standard image is photographed by a high-resolution camera device to obtain the image to be tested presented by the display screen in response to the standard image;

[0039] Perform scattered light analysis on the image to be tested to construct a scattering feature map of the image to be tested. The scattering feature map of the image to be tested can display the scattering characteristics of light in different areas of the image to be tested. The image features of the image to be tested on the display screen can be enhanced through the scattering feature map of the image to be tested, thereby improving the effectiveness of subsequent image distortion repair;

[0040] The image to be tested and the scattering feature map of the image to be tested are sent to the image distortion repair model for processing, and the display image to be tested is output; the image to be tested is processed by the image distortion repair model, so that the output display image to be tested can eliminate the image distortion problem caused by factors such as camera settings and environmental influences during the shooting process, thereby improving the accuracy of subsequent defect detection through the difference feature map;

[0041] Performing difference calculation and normalization operation on the display image to be tested and its corresponding standard image, and outputting a difference feature map; the difference feature map can characterize the defects of the display screen according to the difference between the display image to be tested and its corresponding standard image;

[0042] All difference feature maps are spliced ​​according to channels to construct the display feature map to be tested, the display feature map to be tested is subjected to defect analysis, and the display screen defect detection result is output; specifically, the display feature map to be tested can be subjected to defect analysis through a neural network model, and if there is a defect, the defect type and defect location are output, and the defect type includes bright spots, dark spots, and color deviation, etc.;

[0043] The present application uses a high-resolution camera to take a picture of a display screen to obtain the display screen's response to an input standard image; next, by performing a scattered light analysis on the image to be tested, the generated scattering feature map of the image to be tested can enhance the image features of the image to be tested in the display screen; secondly, the image distortion repair model can effectively eliminate distortion problems caused by factors such as camera settings and environmental influences, thereby ensuring the quality of the displayed image to be tested, thereby providing a more accurate data basis for subsequent defect detection; finally, by analyzing the corresponding difference feature map between the displayed image to be tested and its corresponding standard image, defects in the display performance of the display screen can be effectively detected.

[0044] The image distortion restoration model is established based on the U-net network, including five encoding blocks, which are respectively recorded as the first encoding block, the second encoding block, the third encoding block, the fourth encoding block and the fifth encoding block, four decoding blocks, which are respectively recorded as the first decoding block, the second decoding block, the third decoding block and the fourth decoding block, four multi-scale fusion blocks, which are respectively recorded as the first multi-scale fusion block, the second multi-scale fusion block, the third multi-scale fusion block and the fourth multi-scale fusion block, and a self-attention mechanism block, wherein the specific settings of the five encoding blocks and the four decoding blocks refer to the existing U-net network, and the encoding block is used to extract features of the image to be tested, and the decoding block is used to restore the image resolution so that the feature map is restored to its original size; the multi-scale fusion block is used to perform feature fusion; the self-attention mechanism block is used to perform self-attention mechanism operation through the scattering feature map of the image to be tested;

[0045] The image to be tested is sent to the image distortion restoration model for processing, and the display image to be tested is output, which specifically includes the following steps:

[0046] The image to be tested is processed by the first encoding block, the second encoding block, the third encoding block, the fourth encoding block and the fifth encoding block in sequence, and the first feature map, the second feature map, the third feature map, the fourth feature map and the fifth feature map are output; the fourth feature map is processed by the fourth multi-scale fusion block to obtain the sixth feature map, and the fifth feature map is fused with the sixth feature map and processed by the fourth decoding block to obtain the seventh feature map; the third feature map is processed by the third multi-scale fusion block to obtain the eighth feature map, and the seventh feature map is fused with the eighth feature map and processed by the third decoding block to obtain the ninth feature map; the second feature map is processed by the second multi-scale fusion block to obtain the tenth feature map, and the ninth feature map is fused with the tenth feature map and processed by the third decoding block. The first feature map is processed by the second decoding block to obtain an eleventh feature map; the first feature map is processed by the first multi-scale fusion block to obtain a twelfth feature map, and the twelfth feature map is fused with the thirteenth feature map and processed by the first decoding block to obtain a thirteenth feature map; the thirteenth feature map and the scattering feature map of the image to be tested are processed by the self-attention mechanism block, and the display image to be tested is output; it should be noted that the first encoding block, the second encoding block, the third encoding block, the fourth encoding block and the fifth encoding block all include a downsampling operation, that is, a pooling operation; and the first decoding block, the second decoding block, the third decoding block and the fourth decoding block all include an upsampling operation, and the resolution of the feature map can be continuously restored by the upsampling operation until it is restored to the size of the image to be tested;

[0047] This application uses the encoding block to extract features from the image to be tested, and can learn the deep-level features corresponding to the image in the display screen. The deep-level features corresponding to the image in the display screen are then analyzed through the decoding block, and the actual image corresponding to the image to be tested is output, thereby eliminating distortion problems caused by factors such as camera settings and environmental influences, thereby improving the accuracy of subsequent display screen defect detection.

[0048] In the multi-scale fusion block, the following operations are performed:

[0049] The feature map input to the multi-scale fusion block is recorded as the input feature map. It should be noted that the feature map input to the multi-scale fusion block referred to here includes the first feature map, the second feature map, the third feature map, and the fourth feature map; the feature map output by the multi-scale fusion block is recorded as the output feature map, and the feature map output by the multi-scale fusion block referred to here includes the sixth feature map, the eighth feature map, the tenth feature map, and the twelfth feature map;

[0050] The input feature map undergoes a dilation factor of n 1The first feature map to be fused is obtained by performing a dilated convolution operation with a coefficient of n. The first fused feature map is subjected to a global average pooling operation to obtain a second feature map to be fused. The first feature map to be fused is subjected to a dilation operation with a coefficient of n. 2 The third feature map to be fused is fused with the first feature map to be fused, and a convolution operation is performed, the convolution kernel can be "1×1", and the fourth feature map to be fused is obtained; the fourth feature map to be fused is expanded once with a coefficient of n 3 The fifth feature map to be fused is obtained by performing a dilated convolution operation with a coefficient of n. The fifth feature map to be fused is fused with the first feature map to be fused, and a convolution operation is performed to obtain a sixth feature map to be fused. The sixth feature map to be fused is subjected to a dilation operation with a coefficient of n. 4 The hole convolution operation is performed to obtain the seventh feature map to be fused; the second feature map to be fused, the third feature map to be fused, the fifth feature map to be fused and the seventh feature map to be fused are spliced ​​according to the channel, and a convolution operation is performed once to obtain an output feature map; and {n 1 , n 2 , n 3 , n 4} is a set of expansion coefficients, which is obtained by simulation calculation through a swarm optimization algorithm and can adapt to the resolutions of different display screens, thereby improving the accuracy of display screen defect detection.

[0051] The multi-scale fusion block provided in the present application can capture receptive field information of different scales, and fuse the receptive field information of different scales, which can effectively capture multi-scale contextual information with high coverage and improve the degree of learning in the feature extraction process.

[0052] In the self-attention mechanism block, the following operations are performed:

[0053] The thirteenth feature map is multiplied by the key weight matrix and the value weight matrix respectively to obtain the key feature map K and the value feature map V; the scattering feature map of the image to be tested is multiplied by the query weight matrix to obtain the query feature map Q; the attention weight matrix ATT is calculated based on the key feature map K and the query feature map Q, ATT=softmax[QK T / (d) 0.5 ], where d is the dimension of the query feature map Q; finally, the attention weight matrix ATT is multiplied by the value feature map V to obtain the display image to be tested.

[0054] The present application performs feature enhancement on the thirteenth feature map through the self-attention mechanism and the scattering feature map of the image to be tested, and can enhance the image in the display screen according to the scattering optical characteristic information contained in the scattering feature map of the image to be tested, highlight the feature information corresponding to the image in the display screen, and thereby improve the effect of subsequent image distortion repair.

[0055] The training of the image distortion restoration model includes the following steps:

[0056] Acquire several images to be tested and their corresponding scattering feature maps of the images to be tested. It should be noted that the images to be tested and their corresponding scattering feature maps of the images to be tested referred to here can be obtained by professionals based on actual display screen displays, or can be obtained through simulation software through simulation, and the images to be tested and their corresponding scattering feature maps of the images to be tested are annotated with standard images, and all the annotated images to be tested and their corresponding scattering feature maps of the images to be tested are composed into a training set, and then the training set is sent to a parameter-initialized image distortion repair model for training, with the standard image as the target, and the loss value is calculated. If the training conditions are met, the trained image distortion repair model is output, and the training conditions can be reaching a certain number of training times, or the loss value is within a preset range; if the training conditions are not met, the image distortion repair model continues to be trained through the training set.

[0057] Example 2

[0058] A display screen defect detection system based on visual perception, such as Figure 1 As shown, including:

[0059] The test image acquisition module is used to acquire the test image presented by the display screen in response to the standard image through a high-resolution camera device;

[0060] A standard image storage module, used for storing standard images;

[0061] A scattered light analysis module is used to perform scattered light analysis on the image to be tested to construct a scattering characteristic map of the image to be tested;

[0062] An image distortion restoration module is used to send the image to be tested and the scattering characteristic map of the image to be tested into the image distortion restoration model for processing, and output the display image to be tested;

[0063] A difference analysis module is used to perform difference calculation and normalization operations on the display image to be tested and its corresponding standard image, and output a difference feature map;

[0064] The defect analysis module is used to splice all the difference feature maps according to channels, construct the display feature map to be tested, perform defect analysis on the display feature map to be tested, and output the display defect detection result; specifically, the display feature map to be tested can be defect analyzed through a neural network model, and if there is a defect, the defect type and the location of the defect are output, and the defect types include bright spots, dark spots, and color deviations.

[0065] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A display screen defect detection method based on visual perception, characterized in that: include: Acquire the image to be tested presented by the display screen in response to the standard image; Performing scattered light analysis on the image to be tested to construct a scattering characteristic map of the image to be tested; The image to be tested and the scattering characteristic map of the image to be tested are sent to the image distortion restoration model for processing, and the display image to be tested is output; Perform difference calculation and normalization operation on the display image to be tested and its corresponding standard image, and output a difference feature map; All difference feature maps are spliced ​​according to channels to construct the display feature map to be tested, the display feature map to be tested is subjected to defect analysis, and the display defect detection result is output; The image distortion restoration model is established based on the U-net network, including five encoding blocks, which are respectively recorded as the first encoding block, the second encoding block, the third encoding block, the fourth encoding block and the fifth encoding block, four decoding blocks, which are respectively recorded as the first decoding block, the second decoding block, the third decoding block and the fourth decoding block, four multi-scale fusion blocks, which are respectively recorded as the first multi-scale fusion block, the second multi-scale fusion block, the third multi-scale fusion block and the fourth multi-scale fusion block, and a self-attention mechanism block, and the encoding block is used to extract features of the image to be tested, and the decoding block is used to restore the image resolution; the multi-scale fusion block is used to perform feature fusion; the self-attention mechanism block is used to perform self-attention mechanism operation through the scattering feature map of the image to be tested; The image to be tested is sent to the image distortion restoration model for processing, and the display image to be tested is output, which specifically includes the following steps: The image to be tested is processed by the first encoding block, the second encoding block, the third encoding block, the fourth encoding block and the fifth encoding block in sequence, and the first feature map, the second feature map, the third feature map, the fourth feature map and the fifth feature map are output; the fourth feature map is processed by the fourth multi-scale fusion block to obtain the sixth feature map, and the fifth feature map is fused with the sixth feature map and processed by the fourth decoding block to obtain the seventh feature map; the third feature map is processed by the third multi-scale fusion block to obtain the eighth feature map, and the seventh feature map is fused with the eighth feature map and processed by the third decoding block to obtain the ninth feature map; the second feature map is processed by the second multi-scale fusion block to obtain the tenth feature map, and the ninth feature map is fused with the tenth feature map and processed by the second decoding block to obtain the eleventh feature map; the first feature map is processed by the first multi-scale fusion block to obtain the twelfth feature map, and the twelfth feature map is fused with the thirteenth feature map and processed by the first decoding block to obtain the thirteenth feature map; the thirteenth feature map and the scattering feature map of the image to be tested are processed by the self-attention mechanism block, and the display image to be tested is output; In the multi-scale fusion block, the following operations are performed: The feature map input to the multi-scale fusion block is recorded as the input feature map; the feature map output by the multi-scale fusion block is recorded as the output feature map; The input feature map is subjected to a dilation coefficient of n1 and a dilated convolution operation to obtain a first feature map to be fused; the first fused feature map is subjected to a global average pooling operation to obtain a second feature map to be fused; the first feature map to be fused is subjected to a dilation coefficient of n2 and a dilated convolution operation to obtain a third feature map to be fused; the third feature map to be fused is fused with the first feature map to be fused, and a convolution operation is performed, and the convolution kernel can be "1×1" to obtain a fourth feature map to be fused; the fourth feature map to be fused is subjected to a dilation coefficient of n3 and a dilated convolution operation to obtain a fifth feature map to be fused; the fifth feature map to be fused is fused with the first feature map to be fused, and a convolution operation is performed to obtain a sixth feature map to be fused; The sixth feature map to be fused is subjected to a dilated convolution operation with a dilation coefficient of n4 to obtain a seventh feature map to be fused; The second feature map to be fused, the third feature map to be fused, the fifth feature map to be fused, and the seventh feature map to be fused are spliced ​​according to channels, and a convolution operation is performed to obtain an output feature map; and {n1, n2, n3, n4} is a set of expansion coefficients, which are obtained by simulation calculation through a swarm optimization algorithm; In the self-attention mechanism block, the following operations are performed: The thirteenth feature map is multiplied by the key weight matrix and the value weight matrix respectively to obtain the key feature map K and the value feature map V; the scattering feature map of the image to be tested is multiplied by the query weight matrix to obtain the query feature map Q; the attention weight matrix ATT is calculated based on the key feature map K and the query feature map Q, ATT=softmax[QK T / (d) 0.5 ], where d is the dimension of the query feature map Q; finally, the attention weight matrix ATT is multiplied by the value feature map V to obtain the display image to be tested; The training of the image distortion restoration model includes the following steps: Obtain several images to be tested and their corresponding scattering feature maps of the images to be tested, and annotate the images to be tested and their corresponding scattering feature maps of the images to be tested through standard images. All the annotated images to be tested and their corresponding scattering feature maps of the images to be tested form a training set, and then send the training set to the image distortion restoration model with parameter initialization for training. Take the standard image as the target and calculate the loss value. If the training conditions are met, output the trained image distortion restoration model; if the training conditions are not met, continue to train the image distortion restoration model through the training set.

2. A display screen defect detection system based on visual perception, characterized in that: The system applies the display screen defect detection method based on visual perception described in claim 1, comprising: The test image acquisition module is used to acquire the test image presented by the display screen in response to the standard image through a high-resolution camera device; A standard image storage module, used for storing standard images; A scattered light analysis module is used to perform scattered light analysis on the image to be tested to construct a scattering characteristic map of the image to be tested; An image distortion restoration module is used to send the image to be tested and the scattering characteristic map of the image to be tested into the image distortion restoration model for processing, and output the display image to be tested; A difference analysis module is used to perform difference calculation and normalization operations on the display image to be tested and its corresponding standard image, and output a difference feature map; The defect analysis module is used to splice all the difference feature maps according to the channels, construct the display feature map to be tested, perform defect analysis on the display feature map to be tested, and output the display defect detection results.

Images