Mura defect intelligent detection method and system in OLED display panel

Abstract

The application provides a kind of Mura defect intelligent detection method and system in OLED display panel, it is related to defect detection field, the method flow is: based on the original image of OLED display panel obtains multiscale feature image and fusion feature image;To the multiscale feature image and fusion feature image are carried out feature enhancement and feature fusion, to obtain initial prediction image and encoding feature image;To initial prediction image and encoding feature image are carried out edge thinning and image segmentation, to obtain fine prediction image;End-to-end multi-objective optimization training is carried out using hybrid loss function, to comprehensively optimize network performance.The application adopts the way that three module fusion decoder architecture and hybrid loss function are combined to carry out feature enhancement and network optimization, so that network model can accurately detect Mura defect, prevent the problem of missed detection and wrong detection.

Description

Technical Field

[0001] The present invention relates to the field of defect detection, and in particular to a method and system for intelligently detecting mura defects in OLED display panels. Background Art

[0002] OLED display panels have been widely used in high-end consumer electronics due to their excellent display performance. However, OLED display panels are prone to mura defects. Mura defects refer to spots or marks with uneven local brightness or color on display devices (such as LCD and OLED screens). Because they are too faint, they are difficult to detect and have seriously affected product quality and production costs.

[0003] Existing technologies for detecting mura defects in OLED display panels primarily include: 1. Traditional machine vision-based defect detection, such as template matching and linear filtering; and 2. Deep learning model-based defect detection. In actual production environments, existing traditional machine vision-based defect detection suffers from high false detection and missed detection rates, making it difficult to detect mura defects with low contrast, diverse morphologies, and blurred boundaries. Existing deep learning-based defect detection methods also suffer from inadequate feature representation, insufficient edge perception, and poor adaptability. Furthermore, there is a trade-off between the real-time requirements of production lines and detection accuracy. Furthermore, existing technologies often utilize single-modality defect images and lack comprehensive feature enhancement mechanisms at multiple angles and scales. Summary of the Invention

[0004] The present invention provides a method and system for intelligent detection of mura defects in OLED display panels, which solves the problem that the existing technology cannot efficiently and accurately perform comprehensive intelligent detection of mura defects of different shapes, different scales, and different contrasts, which leads to the easy omission and false detection of mura defects.

[0005] In a first aspect, an embodiment of the present invention provides a method for intelligently detecting mura defects in an OLED display panel, the method comprising the following process:

[0006] Obtain multi-scale feature images and fused feature images based on the original image of the OLED display panel;

[0007] The channel attention enhancement module, dual-path feature enhancement module, weak target enhancement module and fusion decoder module are used to perform feature enhancement and feature fusion on the multi-scale feature image and the fused feature image to obtain the initial prediction image and the encoded feature image;

[0008] An edge refinement attention module is used to perform edge refinement and image segmentation on the initial prediction image and the encoded feature image to obtain the final refined prediction image;

[0009] A hybrid loss function is used to perform end-to-end multi-objective optimization training on the channel attention enhancement module, dual-path feature enhancement module, weak target enhancement module and edge refinement attention module to comprehensively optimize the network model performance.

[0010] In the above-mentioned embodiment, the network model of the present invention adopts a three-module fusion decoder architecture combined with a hybrid loss function. This solves the problem that existing defect detection technologies are unable to efficiently and accurately perform comprehensive intelligent detection of mura defects of different shapes, scales, and contrasts, which in turn leads to missed detection and false detection of mura defects.

[0011] As some optional embodiments of the present application, the process of obtaining multi-scale feature images and fused feature images based on the original image of the OLED display panel is as follows:

[0012] Performing multi-scale feature extraction on the original image of the OLED display panel to obtain a multi-scale feature image;

[0013] Feature fusion is performed on multi-scale feature images to obtain a fused feature image.

[0014] In the above embodiment, the present invention selects ResNeSt-34 with a split attention mechanism as the backbone network to obtain multi-scale feature images and fused feature images. The backbone network can more effectively capture the texture information and brightness changes of the defect area through feature reorganization and group representation learning.

[0015] As some optional implementations of the present application, the process of performing feature enhancement and feature fusion on the multi-scale feature image and the fused feature image using the channel attention enhancement module, the dual-path feature enhancement module, the weak target enhancement module, and the fusion decoder module is as follows:

[0016] The channel attention enhancement module is used to dynamically adjust the channel weights of the fused feature image to obtain the enhanced feature image;

[0017] A dual-path feature enhancement module is used to extract the features of the deep path and shallow path from the enhanced feature image, and the dual-path feature fusion is performed on the features of the deep path and shallow path to obtain the middle-layer feature image;

[0018] The weak target enhancement module is used to dynamically adjust the texture mapping weights of the enhanced feature image to obtain the regional feature image;

[0019] A fusion decoder module is used to upsample and fuse the enhanced feature image, the mid-level feature image and the regional feature image to obtain a three-channel fused feature image; and a fusion decoder module is used to upsample and fuse the three-channel fused feature image and the multi-scale feature image in turn to obtain an initial predicted image and a coded feature image.

[0020] In the above embodiments, the channel attention enhancement module of the present invention applies a channel attention mechanism, which can dynamically adjust the weights of each feature channel, highlight the defect-related channels, and suppress background interference; the dual-path feature enhancement module uses cascaded residual blocks to perform deep path feature extraction on the enhanced feature image, which can effectively expand the receptive field and extract rich semantic features, thereby modeling the overall shape and contextual relationship of the defect; the weak target enhancement module adopts a weighted texture mapping strategy to dynamically adjust the texture mapping weights, which can specifically enhance low-contrast areas and weak boundary areas.

[0021] As some optional embodiments of the present application, the process of using the edge refinement attention module to perform edge refinement and image segmentation on the initial prediction image and the encoded feature image is as follows:

[0022] Performing edge feature extraction on the initial prediction image and the encoded feature image to obtain a potential edge feature image;

[0023] Constructing a spatial attention image and applying the spatial attention image to the edge feature image to obtain an enhanced edge feature image;

[0024] The enhanced edge feature image and the coded feature image are fused to obtain the final refined prediction image.

[0025] In the above embodiment, the edge refinement attention module of the present invention receives the saliency map and detail features generated in the previous stages, and utilizes the spatial attention mechanism and residual connection to effectively enhance edge clarity and maintain structural integrity.

[0026] As some optional implementations of the present application, the process of end-to-end multi-objective optimization training of the channel attention enhancement module, the dual-path feature enhancement module, the weak target enhancement module, and the edge refinement attention module using a hybrid loss function is as follows:

[0027] Construct a binary cross entropy loss function and perform iterative training using the binary cross entropy loss function;

[0028] Construct a structural similarity loss function and perform iterative training using the structural similarity loss function;

[0029] Construct an intersection-over-union loss function and perform iterative training using the intersection-over-union loss function;

[0030] Construct a gradient loss function and perform iterative training using the gradient loss function;

[0031] A hybrid loss function is constructed based on the binary cross entropy loss function, structural similarity loss function, intersection-over-union loss function and gradient loss function to comprehensively optimize the performance of the network model.

[0032] In the above embodiment, the present invention introduces four loss functions to collaboratively constrain the performance of the network model at different levels. This not only ensures good segmentation accuracy but also accurately restores the true structure and edge features of the defect, making it particularly suitable for mura defect detection scenarios with extremely high requirements for details.

[0033] As some optional implementations of the present application, the formula for the three-channel fusion feature image is expressed as:

[0034] ;in, represents the fusion feature image of three channels, represents the middle-level feature image, represents the regional feature image, represents the upsampled enhanced feature image.

[0035] As some optional implementations of the present application, the formula of the hybrid loss function is expressed as:

[0036] ;

[0037] Among them, L total represents the mixed loss function, L BCE represents the binary cross entropy loss function, L SSIM represents the structural similarity loss function, L IOU represents the intersection-over-union loss function, L Gradient represents the gradient loss function, y ture Indicates the true result of the defect, y pred Represents the defect prediction result corresponding to the refined prediction image.

[0038] In a second aspect, the present invention provides an intelligent detection system for mura defects in an OLED display panel, the system comprising:

[0039] A feature extraction unit, wherein the feature extraction unit obtains a multi-scale feature image and a fused feature image based on an original image of the OLED display panel;

[0040] A feature enhancement and fusion unit, which uses a channel attention enhancement module, a dual-path feature enhancement module, a weak target enhancement module, and a fusion decoder module to perform feature enhancement and feature fusion on the multi-scale feature image and the fused feature image to obtain an initial predicted image and an encoded feature image;

[0041] a defect edge segmentation unit, wherein the defect edge segmentation unit uses an edge refinement attention module to perform edge refinement and image segmentation on the initial prediction image and the encoded feature image to obtain a final refined prediction image;

[0042] A multi-objective optimization training unit uses a hybrid loss function to perform end-to-end multi-objective optimization training on the channel attention enhancement module, the dual-path feature enhancement module, the weak target enhancement module, and the edge refinement attention module to comprehensively optimize the network model performance.

[0043] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the intelligent detection method for mura defects in an OLED display panel when executing the computer program.

[0044] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the computer program implements the intelligent detection method for mura defects in an OLED display panel.

[0045] The beneficial effects of the present invention are as follows:

[0046] The present invention adopts a three-module fusion decoder architecture, integrating a channel attention enhancement module, a dual-path feature enhancement module, and a weak target enhancement module to achieve multi-scale and multi-angle feature enhancement and feature fusion, and integrates an edge refinement attention module to achieve highly precise capture of defect edge details, thereby generating accurate and high-resolution saliency images.

[0047] The present invention adopts a comprehensive optimized hybrid loss function for model constraint, so that the model can effectively deal with multiple issues such as pixel-level accuracy, structural consistency, regional overlap and edge precision. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For those skilled in the art, other relevant drawings can be obtained based on these drawings without creative work.

[0049] Figure 1 is a flow chart of the intelligent detection method for mura defects according to an embodiment of the present invention;

[0050] Figure 2 is a structural diagram of a channel attention enhancement module according to an embodiment of the present invention;

[0051] Figure 3 is a schematic structural diagram of a dual-path feature enhancement module according to an embodiment of the present invention;

[0052] Figure 4 is a schematic structural diagram of a weak target enhancement module according to an embodiment of the present invention;

[0053] Figure 5 2 is a schematic diagram of the structure of the edge refinement attention module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0055] To address the problem that existing technologies cannot efficiently and accurately perform comprehensive intelligent detection of mura defects of different shapes, scales, and contrasts, which in turn leads to missed and false detection of mura defects, the present invention provides an intelligent detection method and system for mura defects in OLED display panels. Through a multi-module collaborative working mechanism, it can comprehensively detect mura defects of different shapes, scales, and contrasts, providing reliable technical support for improving product yield and reducing production costs in the OLED display panel manufacturing industry.

[0056] See also Figure 1 , Figure 1 Flowchart of the intelligent detection method for mura defects is as follows:

[0057] (1) Collect the original image of the OLED display panel.

[0058] In an embodiment of the present invention, a high-resolution industrial camera is used to capture images of an OLED display panel to obtain a raw image of the OLED display panel. This raw image covers common mura defect types, such as bright spots, dark bands, color blocks, and cloudiness. During the raw image capture process, stable lighting conditions and uniform shooting parameters are maintained to minimize the impact of environmental changes on image quality.

[0059] Specifically, each original image should be accompanied by the following field information: ① Image number: unique image identifier; ② Acquisition time: image capture time; ③ Defect label: defect inclusion / exclusion, and mura defect type; ④ Defect location information: defect mask; and all original images and corresponding field information will be uniformly formatted as standard input for use by the subsequent neural network model of the embodiment of the present invention.

[0060] In an embodiment of the present invention, the original image is input into a neural network model for model training, and intelligent detection of mura defects is achieved through the trained neural network model.

[0061] (2) Obtaining multi-scale feature images based on the original image of the OLED display panel And the fusion feature image F en .

[0062] In an embodiment of the present invention, ResNeSt-34 with a split attention mechanism is selected as the backbone network to obtain multi-scale feature images and fused feature images. The backbone network can more effectively capture the texture information and brightness changes of the defect area through feature reorganization and group representation learning.

[0063] In an embodiment of the present invention, the process of obtaining a multi-scale feature image and a fused feature image based on an original image of an OLED display panel is as follows:

[0064] (2.1) The low-level semantic features of the original image are extracted through the convolution layer. That is, the basic visual information of the mura defect, such as the edge and texture, is first obtained through a convolution layer and a maximum pooling layer, and then the low-level semantic features T1 are generated.

[0065] (2.2) Through the four residual stages in the ResNeSt-34 standard structure, the mid- and high-level semantic features of four scales are extracted respectively. ; Among them, each residual stage consists of multiple ResNeSt modules, with a segmentation attention mechanism embedded inside, allowing attention-weighted selective feature fusion to be performed within each channel group, thereby more flexibly adapting to different brightness levels and low-contrast texture changes.

[0066] (2.3) Through a convolutional layer to multi-scale feature maps Perform feature fusion to generate fusion feature F en .

[0067] (3) Using the channel attention enhancement module (CAEM), dual path feature enhancement module (DPFEM), weak target enhancement module (DPFEM) and fused decoder block to multi-scale feature images And the fusion feature image F en Perform feature enhancement and feature fusion to obtain the initial prediction image F init And the encoded feature image F de .

[0068] Specifically, for multi-scale feature images and fused feature image F en The process of feature enhancement and feature fusion is as follows:

[0069] (3.1) Channel Attention Enhancement Module (CAEM): For the fused feature image F en And the upsampled feature image F of the original image up Apply the channel attention mechanism to dynamically adjust the weight of each feature channel, highlight the defect-related channels, and suppress background interference, thereby generating an enhanced feature image. .

[0070] For details, please refer to Figure 2 , Figure 2 This is a structural diagram of the channel attention enhancement module, which mainly includes upsampling operation (Upconv) and global average pooling operation (Ch), activation function (PReLU) and convolution layer (Conv).

[0071] Specifically, the channel attention enhancement module is used to obtain enhanced feature images The formula is:

[0072] ;

[0073] ;

[0074] in, represents the enhanced feature image, F up represents the upsampled feature image, F en represents the fused feature image, represents the connection operation along the channel dimension, F GAP Represents the global average pooling operation, Conv represents the convolution layer, Represents the PReLU activation function.

[0075] (3.2) Dual-Path Feature Enhancement Module (DPFEM): It mainly includes the convolution layer (Conv), batch normalization (BatchNorm), activation function (PReLU), and channel connection operation (Channel Shuffle), etc. Figure 3 , Figure 3 Schematic diagram of the structure of the dual-path feature enhancement module.

[0076] First, the cascaded residual blocks are used to enhance the feature image The deep path feature extraction effectively expands the receptive field and extracts rich semantic features, thereby modeling the overall shape and contextual relationship of the defect. It consists of multiple stacked residual blocks, each of which contains two convolutional layers, a batch normalization layer, and an activation function.

[0077] Specifically, the characteristics of the deep path are expressed as follows:

[0078] ;

[0079] Among them, Conv represents the convolution layer, C represents the total number of channels, BN1 and BN2 represent batch normalization, Represents the PReLU activation function.

[0080] Then, a single convolutional layer is used to extract features of the shallow path, and channel reordering is performed to retain detail information, reducing the number of output channels to half of the input channels.

[0081] Specifically, the characteristic representation of the shallow path is:

[0082] ;

[0083] Among them, Conv represents the convolution layer, C represents the total number of channels, BN1 and BN2 represent batch normalization, Represents the PReLU activation function.

[0084] Finally, the features of the deep path and the shallow path are combined 、 Perform dual-path feature fusion to generate richer mid-level feature images .

[0085] Specifically, the middle-level feature image is represented as:

[0086] ;

[0087] in, represents the middle-level feature image, represents the characteristics of the deep path, Characteristic representing shallow paths.

[0088] (3.3) Weak Target Enhancement Module (WTEM): includes convolutional layers (Conv), batch normalization (BatchNorm), activation function (PReLU), and channel connection operations (Channel Shuffle), etc. Figure 4 , Figure 4 Schematic diagram of the structure of the weak target enhancement module.

[0089] Specifically, a weighted texture mapping strategy is used to enhance the feature image Dynamically adjust the texture mapping weights to specifically enhance low-contrast areas and weak boundary areas to obtain regional feature images ;

[0090] Specifically, the regional feature image is represented as:

[0091] ;

[0092] in, Represents the regional feature image, Conv represents the convolution layer, BN1 represents batch normalization, represents the PReLU activation function, and C represents the total number of channels.

[0093] (3.4) Use the fused decoder block to enhance the feature image , middle-level feature image and regional feature images Perform upsampling and feature fusion to obtain a three-channel fusion feature image ; And the fusion decoder module is used to sequentially decode the multi-scale feature images Perform upsampling and feature fusion to obtain the initial prediction image F init And the encoded feature image F de .

[0094] Specifically, the fusion feature image of the three channels is expressed as:

[0095] ;

[0096] in, represents the fusion feature image of three channels, represents the middle-level feature image, represents the regional feature image, represents the upsampled enhanced feature image.

[0097] (4) Use edge refinement attention module (EFAM) to refine the initial prediction image F init And the encoded feature image F de Perform edge refinement and high-precision segmentation to obtain the final refined prediction image F final , the edge refinement attention module mainly includes convolution layer (Conv), batch normalization (BatchNorm), activation function (PReLU), channel connection operation (Channel Shuffle) and activation function (Singmoid), etc., please refer to Figure 5 , Figure 5 Schematic diagram of the structure of the edge refinement attention module.

[0098] In this embodiment of the present invention, in order to accurately capture the edge details of mura defects, an edge refinement attention module is also proposed. The edge refinement attention module receives the saliency map and detail features generated in the previous stages, and uses a spatial attention mechanism and residual connections to effectively enhance edge clarity and maintain structural integrity.

[0099] Specifically, for the initial prediction image F init And the encoded feature image F de The process of edge refinement and high-precision segmentation is as follows:

[0100] (4.1) For the initial prediction image F init And the encoded feature image F de Perform edge feature extraction to obtain potential edge feature image F conv .

[0101] (4.2) Constructing spatial attention image A sp , and the spatial attention image A sp Acting on edge feature image F conv , to obtain the enhanced edge feature image F conv .

[0102] (4.3) The enhanced edge feature image F conv And the encoded feature image F de Perform image fusion to obtain the final refined prediction image F final .

[0103] Specifically, the formula for the refined prediction image is expressed as:

[0104] ;

[0105] ;

[0106] ;

[0107] Among them, Conv represents the convolution layer, BN1 and BN2 represent batch normalization, represents the PReLU activation function, Represents the Sigmoid function.

[0108] (5) Using the hybrid loss function L totalEnd-to-end multi-objective optimization training is performed on the channel attention enhancement module, dual-path feature enhancement module, weak target enhancement module and edge refinement attention module to comprehensively optimize the network model performance.

[0109] Specifically, the process of end-to-end multi-objective optimization training using a hybrid loss function is as follows:

[0110] (5.1) Constructing the binary cross entropy loss function L BCE , and through the binary cross entropy loss function L BCE Iterative training is performed to achieve high-precision segmentation accuracy.

[0111] Specifically, the binary cross entropy loss function is expressed as:

[0112] ;

[0113] Among them, y i Indicates the real label 1 or the virtual label 0, Represents the predicted probability value, and N represents the total pixel value of the image.

[0114] (5.2) Constructing the structural similarity loss function L SSIM , and through the structural similarity loss function L SSIM Iterative training is performed to achieve consistency in visual structure between the segmentation results and the true defect area.

[0115] Specifically, the structural similarity loss function is expressed as:

[0116] ;

[0117] in, and represents the local pixel mean of the image, and represents the local pixel variance of the image, represents the local pixel covariance of the image, and C1 and C2 represent constants.

[0118] (5.3) Construct the intersection-over-union loss function L IOU , and through the intersection-over-union loss function L IOU Iterative training to optimize the overall coverage of the defect area;

[0119] Specifically, the intersection-over-union loss function is expressed as:

[0120] ;

[0121] Among them, y represents the real label 1 or virtual label 0 of the image, Represents the predicted probability value of the image.

[0122] (5.4) Construct the gradient loss function L Gradient , and through the gradient loss function L Gradient Iterative training to constrain the clarity and continuity of defect edge refinement;

[0123] Specifically, the gradient loss function is:

[0124] ;

[0125] in, and Represents gradients in horizontal and vertical directions.

[0126] (5.5) Constructing the hybrid loss function L total , to comprehensively optimize the network model performance.

[0127] Specifically, the hybrid loss function is:

[0128] ;

[0129] Among them, y ture represents the actual result of the defect, that is, the result of manual annotation, y pred Represents the defect prediction result corresponding to the refined prediction image.

[0130] Specifically, the following describes a specific implementation of an embodiment of the present invention for mura defects that occur during the production of OLED display panels. It should be noted that this implementation is only one implementation of an embodiment of the present invention, and not the only implementation of the embodiments of the present invention. Furthermore, the intelligent detection method for mura defects in OLED display panels proposed in an embodiment of the present invention is not limited to the field of OLED display panel manufacturing.

[0131] S10: Collect original images.

[0132] Specifically, a high-definition industrial camera (e.g., 5 megapixels and above) is used to capture images of the OLED display panel to be tested at the inspection station to capture mura defects that may occur under different working conditions.

[0133] Specifically, use annotation software to manually annotate the defective areas of the original image and generate a mask file. All mask images must be exactly the same size as the original image.

[0134] S20: Multi-scale feature extraction.

[0135] In the embodiment of the present invention, in order to capture the different scale information of the defect, ResNeSt-34 multi-scale convolution is used to extract detail features and overall features respectively:

[0136] The original image passes through the following structural modules in sequence:

[0137] ① Initial convolutional layer: The convolution kernel size is 7×7, the stride is 2, the number of output channels is 64, and it is connected to a 3×3 maximum pooling layer; then the feature image T1 is obtained, with a size of 1 / 4H×1 / 4W.

[0138] ② Backbone feature extraction layer: The ResNeSt-34 backbone network contains 4 stages, each stage contains multiple residual blocks, and introduces a split attention mechanism to obtain feature images .

[0139] ③Image fusion layer: multi-scale feature images through convolution layer Perform feature fusion to generate a fusion feature image F en .

[0140] S30: Feature enhancement and fusion.

[0141] After completing multimodal image extraction and fusion, this embodiment further enhances and fuses these features. This is achieved primarily through the Channel Attention Enhancement Module (CAEM), the Dual Path Feature Enhancement Module (DPFEM), the Weak Target Enhancement Module (WTEM), and the Fusion Decoder module, which together form a fusion decoder architecture. This architecture enhances defect-related features, preserves edge information, and compensates for low-contrast areas of weak targets.

[0142] ① Channel Attention Enhancement Module (CAEM):

[0143] The fused feature image F en When input to this module, a global average pooling operation is first performed to obtain the average response intensity of each channel in the entire spatial range; then, a two-layer lightweight feedforward neural network is used to model these channel averages and output a set of channel weights between 0 and 1; this set of weights is then broadcast back to the feature image, and channel-level saliency adjustment is achieved through channel-by-channel multiplication operations.

[0144] The function of the channel attention enhancement module is similar to the focusing mechanism of the human eye. It can automatically highlight the channels with strong defect responses and suppress redundant or useless channel responses in the background area, thereby improving the efficiency and accuracy of subsequent network processing. After processing by the channel attention enhancement module, the enhanced feature image is recorded as .

[0145] ②Dual Path Feature Enhancement Module (DPFEM):

[0146] The dual-path feature enhancement module is used to extract structural information with multi-scale semantics from the fused features; specifically, the input features are guided to the deep path and the shallow path respectively. The deep path consists of three residual blocks connected in series, each of which contains two 3×3 convolutional layers, two batch normalization layers and an activation function. Residual connections can help deep networks avoid gradient vanishing, so that semantic features can still be efficiently propagated in deeper networks; and convolution stacking is used to expand the receptive field and improve the model's ability to understand the contextual information of the defect area. In this path, the network can extract higher-level semantic features such as the structural shape and direction pattern of mura defects. Relatively speaking, the shallow path is more focused on preserving details and boundaries. Its structure only contains one 3×3 standard convolution layer, and a channel reordering operation is performed after the convolution to compress the number of original input channels to half to retain the most critical edge and texture information; finally, the output of the deep path and the output of the shallow path are spliced ​​along the channel dimension and fused into a mid-level enhanced feature map through 1×1 convolution. .

[0147] ③Weak Target Enhancement Module (WTEM):

[0148] In order to further improve the network model's ability to respond to weak defects, the weak target enhancement module is used to process weak target areas in the image with extremely low brightness contrast, blurred edges and small size. First, through texture analysis, the texture intensity distribution map of the input image is calculated, and areas with drastic and gentle texture changes are identified. On this basis, a set of weighted coefficient maps are constructed to give more weight to areas with weaker texture intensity. Subsequently, these weighted coefficients are applied to the feature image to highlight potential weak defect areas. In this process, the network can selectively increase the feature amplitude of the low-response area, making the small stripes or light spots hidden in the background more prominent. Finally, the enhanced feature map output by this module is recorded as .

[0149] ④Fused Decoder Block:

[0150] The fusion decoder module is used to enhance the feature image , middle-level feature image and regional feature images Perform upsampling and feature fusion to obtain a three-channel fusion feature image ; And the fusion decoder module is used to sequentially decode the multi-scale feature images Perform upsampling and feature fusion to obtain the initial prediction image F init And the encoded feature image F de .

[0151] S40: Edge refinement and high-precision segmentation.

[0152] In order to further improve the clarity and positioning accuracy of the edge of the mura defect area, the edge refinement attention module is specifically used to perform edge-level enhancement on the preliminary prediction results. First, it receives two inputs: the initial prediction image F init , and the encoder output F from the last block de First, two 3×3 standard convolutional layers are used to extract features from the two inputs to extract potential edge-related information. Then, a spatial attention map A is constructed. sp , the activation of the edge area is enhanced, so that the model focuses on the area with weak edge information and performs specific response improvement. The generation of the attention map adopts the fusion method of channel-by-channel maximum pooling and average pooling to ensure that the spatial attention map has a good perception of regional differences. Finally, the spatial attention map is applied to the edge feature map, and the original structural information is retained through the residual connection. The enhanced edge feature image is combined with the encoder output F de Fusion is performed to output a fine prediction image F with refined edges final .

[0153] S50: Multi-objective optimization training.

[0154] The embodiment of the present invention introduces the performance of four loss collaborative constraint models at different levels, which not only enables it to have good segmentation accuracy, but also accurately restore the real structure and edge features of defects. It is particularly suitable for industrial defect detection scenarios with extremely high requirements on details.

[0155] After the above process, the contrast and edge clarity of the defect area in this embodiment are significantly improved, and the defect detection accuracy of the neural network model is significantly improved.

[0156] Specifically, the accuracy of mura defect detection using the improved neural network model of the present invention can reach over 97.5%, the defect missed detection and false detection rates are less than 2.5%, and the defect edge clarity is extremely high.

[0157] In summary, the neural network model of the present invention adopts a three-module fusion decoder architecture, integrating a channel attention enhancement module, a dual-path feature enhancement module, and a weak target enhancement module to achieve multi-scale and multi-angle feature enhancement and feature fusion. It also integrates an edge refinement attention module to achieve highly accurate capture of defect edge details, thereby generating accurate and high-resolution saliency images. Furthermore, it uses a comprehensively optimized hybrid loss function to effectively address multiple issues such as pixel-level accuracy, structural consistency, regional overlap, and edge precision. This solves the problem that existing defect detection technologies cannot efficiently and accurately perform comprehensive intelligent detection of mura defects of different shapes, scales, and contrasts, which in turn leads to missed and false detections of mura defects.

[0158] Furthermore, in one embodiment, based on the same inventive concept as the aforementioned embodiment, an embodiment of the present invention provides an intelligent detection system for mura defects in an OLED display panel. The system corresponds one-to-one with the method, and the system includes:

[0159] A feature extraction unit, wherein the feature extraction unit obtains a multi-scale feature image and a fused feature image based on an original image of the OLED display panel;

[0160] A feature enhancement and fusion unit, which uses a channel attention enhancement module, a dual-path feature enhancement module, a weak target enhancement module, and a fusion decoder module to perform feature enhancement and feature fusion on the multi-scale feature image and the fused feature image to obtain an initial predicted image and an encoded feature image;

[0161] a defect edge segmentation unit, wherein the defect edge segmentation unit uses an edge refinement attention module to perform edge refinement and image segmentation on the initial prediction image and the encoded feature image to obtain a final refined prediction image;

[0162] A multi-objective optimization training unit uses a hybrid loss function to perform end-to-end multi-objective optimization training on the channel attention enhancement module, the dual-path feature enhancement module, the weak target enhancement module, and the edge refinement attention module to comprehensively optimize the network model performance.

[0163] It should be noted that the various units in the intelligent mura defect detection system of this embodiment correspond one-to-one to the various steps in the intelligent mura defect detection method of the aforementioned embodiment. Therefore, the specific implementation and technical effects achieved by this embodiment can refer to the implementation of the aforementioned intelligent mura defect detection method and will not be repeated here.

[0164] In addition, in one embodiment, the present application also provides a computer device, which includes a processor, a memory, and a computer program stored in the memory, and the computer program implements the method in the aforementioned embodiment when executed by the processor.

[0165] In addition, in one embodiment, the present application further provides a computer storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method in the aforementioned embodiment is implemented.

[0166] In some embodiments, the computer-readable storage medium may be a memory device such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface mount memory, optical disk, or CD-ROM; or various devices including any one or any combination of the above memories. The computer may be various computing devices including smart terminals and servers.

[0167] In some embodiments, executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0168] By way of example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file storing other programs or data, such as, for example, in one or more scripts within a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0169] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0170] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0171] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0172] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a multimedia terminal device (which can be a mobile phone, a computer, a television receiver, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0173] The above are only preferred embodiments of the present application and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An intelligent detection method for mura defects in OLED display panels, characterized in that: The method includes the following steps: Obtain multi-scale feature images and fused feature images based on the original image of the OLED display panel; The channel attention enhancement module, dual-path feature enhancement module, weak target enhancement module and fusion decoder module are used to perform feature enhancement and feature fusion on the multi-scale feature image and the fused feature image to obtain the initial prediction image and the encoded feature image; The process of feature enhancement and feature fusion of multi-scale feature images and fused feature images using the channel attention enhancement module, dual-path feature enhancement module, weak target enhancement module and fusion decoder module is as follows: The channel attention enhancement module is used to dynamically adjust the channel weights of the fused feature image to obtain the enhanced feature image; A dual-path feature enhancement module is used to extract the features of the deep path and shallow path from the enhanced feature image, and the dual-path feature fusion is performed on the features of the deep path and shallow path to obtain the middle-layer feature image; The weak target enhancement module is used to dynamically adjust the texture mapping weights of the enhanced feature image to obtain the regional feature image; A fusion decoder module is used to upsample and fuse the enhanced feature image, the middle-level feature image, and the regional feature image to obtain a three-channel fused feature image; and a fusion decoder module is used to sequentially upsample and fuse the three-channel fused feature image and the multi-scale feature image to obtain an initial predicted image and a coded feature image; An edge refinement attention module is used to perform edge refinement and image segmentation on the initial prediction image and the encoded feature image to obtain the final refined prediction image; A hybrid loss function is used to perform end-to-end multi-objective optimization training on the channel attention enhancement module, dual-path feature enhancement module, weak target enhancement module, and edge refinement attention module to comprehensively optimize the network model performance. The process of end-to-end multi-objective optimization training of the channel attention enhancement module, dual-path feature enhancement module, weak target enhancement module and edge refinement attention module using a hybrid loss function is as follows: Construct a binary cross entropy loss function and perform iterative training using the binary cross entropy loss function; Construct a structural similarity loss function and perform iterative training using the structural similarity loss function; Construct an intersection-over-union loss function and perform iterative training using the intersection-over-union loss function; Construct a gradient loss function and perform iterative training using the gradient loss function; A hybrid loss function is constructed based on the binary cross entropy loss function, structural similarity loss function, intersection-over-union loss function and gradient loss function to comprehensively optimize the performance of the network model.

2. The intelligent detection method for mura defects in an OLED display panel according to claim 1, characterized in that: The process of obtaining multi-scale feature images and fused feature images based on the original image of the OLED display panel is as follows: Performing multi-scale feature extraction on the original image of the OLED display panel to obtain a multi-scale feature image; Feature fusion is performed on multi-scale feature images to obtain a fused feature image.

3. The intelligent detection method for mura defects in an OLED display panel according to claim 1, characterized in that: The process of edge refinement and image segmentation using the edge refinement attention module for the initial prediction image and the encoded feature image is as follows: Performing edge feature extraction on the initial prediction image and the encoded feature image to obtain a potential edge feature image; Constructing a spatial attention image and applying the spatial attention image to the edge feature image to obtain an enhanced edge feature image; The enhanced edge feature image and the coded feature image are fused to obtain the final refined prediction image.

4. The intelligent detection method for mura defects in an OLED display panel according to claim 1, characterized in that: The formula of the three-channel fusion feature image is expressed as: ； in, represents the fusion feature image of three channels, represents the middle-level feature image, represents the regional feature image, represents the upsampled enhanced feature image.

5. The intelligent detection method for mura defects in an OLED display panel according to claim 1, characterized in that: The formula of the hybrid loss function is expressed as: ； Among them, L total represents the mixed loss function, L BCE represents the binary cross entropy loss function, L SSIM represents the structural similarity loss function, L IOU represents the intersection-over-union loss function, L Gradient represents the gradient loss function, y ture Indicates the true result of the defect, y pred Represents the defect prediction result corresponding to the refined prediction image.

6. An intelligent detection system for mura defects in OLED display panels, characterized in that: The system comprises: A feature extraction unit, wherein the feature extraction unit obtains a multi-scale feature image and a fused feature image based on an original image of the OLED display panel; A feature enhancement and fusion unit, which uses a channel attention enhancement module, a dual-path feature enhancement module, a weak target enhancement module, and a fusion decoder module to perform feature enhancement and feature fusion on the multi-scale feature image and the fused feature image to obtain an initial predicted image and an encoded feature image; The process of feature enhancement and feature fusion of multi-scale feature images and fused feature images using the channel attention enhancement module, dual-path feature enhancement module, weak target enhancement module and fusion decoder module is as follows: The channel attention enhancement module is used to dynamically adjust the channel weights of the fused feature image to obtain the enhanced feature image; A dual-path feature enhancement module is used to extract the features of the deep path and shallow path from the enhanced feature image, and the dual-path feature fusion is performed on the features of the deep path and shallow path to obtain the middle-layer feature image; The weak target enhancement module is used to dynamically adjust the texture mapping weights of the enhanced feature image to obtain the regional feature image; A fusion decoder module is used to upsample and fuse the enhanced feature image, the middle-level feature image, and the regional feature image to obtain a three-channel fused feature image; and a fusion decoder module is used to sequentially upsample and fuse the three-channel fused feature image and the multi-scale feature image to obtain an initial predicted image and a coded feature image; a defect edge segmentation unit, wherein the defect edge segmentation unit uses an edge refinement attention module to perform edge refinement and image segmentation on the initial prediction image and the encoded feature image to obtain a final refined prediction image; A multi-objective optimization training unit, which uses a hybrid loss function to perform end-to-end multi-objective optimization training on the channel attention enhancement module, the dual-path feature enhancement module, the weak target enhancement module, and the edge refinement attention module to comprehensively optimize the network model performance; The process of end-to-end multi-objective optimization training of the channel attention enhancement module, dual-path feature enhancement module, weak target enhancement module and edge refinement attention module using a hybrid loss function is as follows: Construct a binary cross entropy loss function and perform iterative training using the binary cross entropy loss function; Construct a structural similarity loss function and perform iterative training using the structural similarity loss function; Construct an intersection-over-union loss function and perform iterative training using the intersection-over-union loss function; Construct a gradient loss function and perform iterative training using the gradient loss function; A hybrid loss function is constructed based on the binary cross entropy loss function, structural similarity loss function, intersection-over-union loss function and gradient loss function to comprehensively optimize the performance of the network model.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for intelligently detecting mura defects in an OLED display panel according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the method for intelligently detecting mura defects in an OLED display panel according to any one of claims 1 to 5 is implemented.

Images