A defect detection method of a display panel

Abstract

This invention discloses a defect detection method for display panels. The method includes: acquiring global and local features of the display panel under test through a lightweight sampling module of a target defect detection network model to obtain a fused feature map of the display panel under test. The target defect detection network model includes a lightweight sampling module, a defect identification module, a frequency domain feature acquisition module, and a deconvolution module; identifying defects in the fused feature map through the defect identification module to obtain defect location information of the display panel under test; acquiring frequency domain features of the fused feature map through the frequency domain feature acquisition module; and fusing and upsampling the defect location information and frequency domain features through the deconvolution module to obtain a defect detection map of the display panel under test. This invention features a lightweight design for the display panel under test, reducing computational load and improving detection speed; and integrates defect location information and frequency domain features to more accurately determine defect location.

Description

Technical Field

[0001] This invention relates to the field of semiconductor display technology, and more specifically to a method for detecting defects in a display panel. Background Technology

[0002] Mura refers to a defect in brightness or color uniformity on a screen. The Mura phenomenon causes inconsistencies in brightness or color on the screen, affecting the user's visual experience and reducing viewing comfort. Severe Mura may interfere with the normal display function of the monitor, leading to unclear images or other quality problems. From device fabrication to mass transfer in the template, Mura defects are inevitably introduced into the chip. Mura defect detection is a crucial step in controlling the quality of LED chips.

[0003] However, in the current process of Mura testing of display panels, the large amount of brightness data of the display panel, due to the use of global or static threshold strategies for calculation, results in low testing efficiency. In addition, in dynamic environments, the brightness changes of the display panel are complex, and traditional methods are difficult to accurately capture subtle defects, affecting the testing accuracy.

[0004] Therefore, in the process of defect detection of display panels, there are problems of low detection efficiency and low defect location accuracy. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for detecting defects in display panels, aiming to solve the problems of low detection efficiency and low defect location accuracy in the existing technology for detecting defects in display panels.

[0006] To address the aforementioned technical problems, the embodiments of the present invention provide the following technical solutions:

[0007] A method for detecting defects in a display panel, comprising:

[0008] The target defect detection network model obtains global and local features of the display panel under test through the lightweight sampling module of the target defect detection network model, and obtains a fused feature map of the display panel under test with the global and local features. The target defect detection network model includes the lightweight sampling module, the defect identification module, the frequency domain feature acquisition module, and the deconvolution module.

[0009] The defect identification module identifies defects in the fused feature map to obtain the defect location information of the display panel under test.

[0010] The frequency domain features of the fused feature map are acquired through the frequency domain feature acquisition module;

[0011] The defect detection map of the display panel under test is obtained by fusing and upsampling the defect location information and the frequency domain features through the deconvolution module.

[0012] Optionally, the lightweight sampling module includes a lightweight downsampling module and a feature balancing module; the step of acquiring global and local features of the display panel under test through the lightweight sampling module, effectively extracting features of the display panel while reducing computational load, includes:

[0013] The lightweight downsampling module sequentially extracts single-channel spatial features and fuses cross-channel features of the display panel under test to obtain the lightweight downsampling features of the display panel under test.

[0014] The lightweight downsampling features are subjected to multi-scale feature fusion balancing processing by the feature balancing module to obtain the fused feature map of the display panel under test.

[0015] Optionally, the lightweight downsampling module includes a depthwise convolution submodule and a pointwise convolution submodule; the lightweight downsampling features include single-channel spatial features and cross-channel features; the step of sequentially extracting single-channel spatial features and fusing cross-channel features on the display panel under test through the lightweight downsampling module to obtain the lightweight downsampling features of the display panel under test includes:

[0016] The depth convolution submodule is used to independently convolve each input channel of the display panel under test to obtain the single-channel spatial features of the display panel under test.

[0017] The pointwise convolution submodule performs linear combination and dimension transformation operations on the channel dimensions of the display panel under test to obtain the cross-channel features of the display panel under test.

[0018] Optionally, the feature balancing module includes a local depthwise convolution submodule, a nonlinear feature extraction submodule, a lightweight self-attention submodule, a feedforward network submodule, and a secondary balancing and fusion submodule; the step of performing multi-scale feature fusion balancing processing on the lightweight downsampled features according to the feature balancing module to obtain the fused feature map of the display panel under test includes:

[0019] The global features and the local features are channel-segmented according to the gamma distribution factor to obtain gamma partial features and residual features;

[0020] The spatial features of the gamma-part features are obtained by performing depthwise convolution on the local depthwise convolution submodule.

[0021] The nonlinear feature extraction submodule performs multi-level nonlinear transformations on the remaining features and the spatial features to obtain multi-level nonlinear features.

[0022] The lightweight self-attention submodule is used to extract key features from the multi-level nonlinear features by performing self-attention feature extraction.

[0023] The key features and the multi-level nonlinear features are fused according to the feedforward network sub-module to obtain the feedforward fused features;

[0024] The feedforward fusion feature and the key feature are segmented by secondary gamma channels according to the secondary balancing and fusion submodule. The segmented features are then subjected to depthwise convolution and normalized activation to obtain the fusion feature map.

[0025] Optionally, the defect identification module includes a residual branch, a dynamic convolution branch, a first ghost convolutional network submodule, and a second ghost convolutional network submodule; the step of performing defect identification on the fused feature map through the defect identification module to obtain the defect location information of the display panel under test includes:

[0026] The fused feature map is convolved and normalized according to the residual branch to obtain basic detail features;

[0027] The feature extraction method of the fused feature map is adaptively adjusted according to the dynamic convolution branch to obtain the convolution weight matrix;

[0028] The fused feature map is expanded by increasing the number of channels according to the first ghost convolutional network submodule to obtain an expanded redundant feature map;

[0029] The second ghost convolutional network submodule performs dynamic weighting and dimensionality compression on the redundant feature map according to the convolutional weight matrix to obtain high-dimensional features;

[0030] Channel concatenation is performed between the high-dimensional features and the basic detail features to obtain a multi-scale feature space;

[0031] The defect location information is determined based on the multi-scale feature space.

[0032] Optionally, the frequency domain feature acquisition module includes a dual-tree complex wavelet transform submodule and an inverse dual-tree complex wavelet transform submodule; the acquisition of frequency domain features of the fused feature map through the frequency domain feature acquisition module includes:

[0033] The fused feature map is filtered by the dual-tree complex wavelet transform submodule to obtain low-frequency approximate components and high-frequency detail components.

[0034] The frequency domain features are obtained by jointly reconstructing the real and imaginary features of the low-frequency approximation component and the high-frequency detail component through the inverse dual-tree complex wavelet transform submodule.

[0035] Optionally, the step of performing feature fusion and upsampling on the defect location information and the frequency domain features through the deconvolution module to obtain the defect detection map of the display panel under test includes:

[0036] The defect location information and the frequency domain features are aligned to obtain defect location information and frequency domain features aligned in spatial resolution and feature dimension.

[0037] The defect location information and frequency domain features aligned in spatial resolution and feature dimension are concatenated along the channel dimension to obtain a time-frequency fusion feature map.

[0038] The time-frequency fusion feature map is deconvolved to restore it from a small size with high semantic information to a larger spatial resolution, thus obtaining the defect detection map.

[0039] Optionally, before obtaining the global and local features of the display panel under test through the lightweight sampling module of the target defect detection network model, the method further includes:

[0040] An initial defect detection network model is constructed, which includes a weighted loss function, comprising a gradient coordination mechanism focus loss function, a structural similarity loss function, and a mean squared error loss function.

[0041] Obtain a training sample set, which includes input data samples and target output label samples;

[0042] The input data sample is input into the initial defect detection network model to obtain the predicted label sample corresponding to the input data sample;

[0043] Obtain the gradient norm of the training sample set, and construct the gradient norm space based on the gradient norm;

[0044] The gradient norm space is dynamically balanced according to the gradient coordination mechanism focus loss function to obtain the assigned weights of the training sample set, and the gradient coordination mechanism focus loss value of the training sample set is determined based on the assigned weights.

[0045] The structural similarity loss value of the input data sample is calculated and determined based on the structural similarity loss function.

[0046] The mean squared error loss value of the input data sample is calculated and determined based on the mean squared error loss function.

[0047] The focus loss value of the gradient coordination mechanism, the structural similarity loss value, and the mean square error loss value are weighted to obtain the target loss function value. The initial defect detection network model is then adjusted based on the target loss function value until a fully trained target defect detection network model is obtained.

[0048] Optionally, the weighted processing of the focus loss value of the gradient coordination mechanism, the structural similarity loss value, and the mean squared error loss value to obtain the target loss function value includes:

[0049] The training sample set is divided into a first batch and a second batch;

[0050] During the training of the initial defect detection network model based on the first batch, the gradient coordination mechanism focus loss value, the structural similarity loss value, and the mean square error loss value are weighted based on the first weight combination;

[0051] During the training of the initial defect detection network model based on the second batch, the gradient coordination mechanism focus loss value, the structural similarity loss value, and the mean square error loss value are weighted based on the second weight combination;

[0052] In the first weight combination, the weight value of the focus loss value of the gradient coordination mechanism is greater than the weight value of the structural similarity loss value; in the second weight combination, the weight value of the structural similarity loss value is much greater than the weight value of the focus loss value of the gradient coordination mechanism.

[0053] A defect detection device for a display panel, comprising:

[0054] The fusion feature map acquisition module is used to acquire global and local features of the display panel under test through the lightweight sampling module of the target defect detection network model, and obtain a fusion feature map of the display panel under test with the global and local features. The target defect detection network model includes the lightweight sampling module, the defect recognition module, the frequency domain feature acquisition module, and the deconvolution module.

[0055] The defect location information acquisition module is used to identify defects in the fused feature map through the defect identification module to obtain the defect location information of the display panel under test.

[0056] A frequency domain feature acquisition module is used to acquire the frequency domain features of the fused feature map through the frequency domain feature acquisition module;

[0057] The defect detection module is used to perform feature fusion and upsampling on the defect location information and the frequency domain features through the deconvolution module to obtain the defect detection map of the display panel under test.

[0058] In this embodiment of the invention, a lightweight sampling module acquires global and local features of the display panel under test, significantly reducing the time required for feature acquisition. By performing a balanced fusion process on the global and local features, a fused feature map of the display panel under test is obtained, ensuring that both global and local features contribute fully to the fusion process. The output fused feature map contains both overall context and detailed defect information. A defect identification module identifies defects in the fused feature map, obtaining the defect location information of the display panel under test, enabling more accurate differentiation between defective and normal areas. A frequency domain feature acquisition module acquires the frequency domain features of the fused feature map, and a deconvolution module fuses and upsamples the defect location information and frequency domain features to obtain a defect detection map of the display panel under test. Since frequency domain features can capture structural features that are difficult to express in the spatial domain, integrating defect location information and frequency domain features allows for more accurate defect localization. In summary, this invention significantly improves the efficiency of defect identification and the accuracy of defect localization. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a schematic diagram of a scenario of an embodiment of the defect detection system for a display panel provided in this invention.

[0061] Figure 2 This is a schematic diagram of another embodiment of the defect detection system for a display panel provided in this invention.

[0062] Figure 3 This is a schematic flowchart of an embodiment of the defect detection method for a display panel provided by the present invention;

[0063] Figure 4 This is a schematic diagram of the structure of an embodiment of the feature balancing module provided in this invention;

[0064] Figure 5 This is a schematic diagram of the structure of an embodiment of the defect identification module provided in this invention;

[0065] Figure 6This is a schematic diagram of the structure of the first embodiment of the target defect detection network model provided in this invention.

[0066] Figure 7 This is a schematic diagram of the structure of the second embodiment of the target defect detection network model provided in this invention.

[0067] Figure 8 A schematic diagram of a defect detection device for a display panel provided in an embodiment of the present invention;

[0068] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] In the following description, specific embodiments of the invention will be illustrated with reference to steps and symbols performed by one or more computers, unless otherwise stated. Therefore, these steps and operations will be referred to several times as being performed by a computer, and computer execution as referred to herein includes operations by a computer processing unit representing electronic signals of data in a structured format. This operation transforms the data or maintains it at a location in the computer's memory system, which can be reconfigured or otherwise alter the operation of the computer in a manner well known to those skilled in the art. The data structure maintained by the data is the physical location of the memory, which has specific characteristics defined by the data format. However, the principles of the invention described above are not intended to be limiting, and those skilled in the art will understand that many of the steps and operations described below can also be implemented in hardware.

[0071] The terms "module" or "unit" as used herein can be considered as software objects executing on the computing system. The different components, modules, engines, and services described herein can be considered as implementation objects on the computing system. The apparatus and methods described herein are preferably implemented in software, but can also be implemented in hardware, both of which are within the scope of this invention.

[0072] This invention provides a method for detecting defects in a display panel.

[0073] Please see Figure 1 , Figure 1This is a schematic diagram of a scenario of an embodiment of the display panel defect detection system provided in this invention. The display panel defect detection system may include a client 100 and a server 200, which are connected via a network. The server 200 integrates a display panel defect detection device. The server 200 may be a work platform server (i.e., a server loaded with a work platform), such as... Figure 1 In this embodiment of the invention, server 200 is mainly used to obtain global and local features of the display panel under test through the lightweight sampling module of the target defect detection network model, and obtain a fused feature map of the display panel under test. The target defect detection network model includes a lightweight sampling module, a defect identification module, a frequency domain feature acquisition module, and a deconvolution module. The defect identification module performs defect identification on the fused feature map to obtain the defect location information of the display panel under test. The frequency domain feature acquisition module acquires the frequency domain features of the fused feature map. The deconvolution module performs feature fusion and upsampling on the defect location information and frequency domain features to obtain the defect detection map of the display panel under test.

[0074] In this embodiment of the invention, the server 200 can be a standalone server, a server network, or a server cluster. For example, the server 200 described in this embodiment includes, but is not limited to, a computer, a network host, a single network server, a set of multiple network servers, or a cloud server composed of multiple servers. The cloud server is composed of a large number of computers or network servers based on cloud computing. In this embodiment, communication between the server and the client can be achieved through any communication method, including but not limited to, mobile communication based on the 3rd Generation Partnership Project (3GPP), Long Term Evolution (LTE), and Worldwide Interoperability for Microwave Access (WiMAX), or computer network communication based on the TCP / IP Protocol Suite (TCP / IP) and User Datagram Protocol (UDP).

[0075] It is understood that the client 100 used in this embodiment of the invention can be understood as a client device. A client device includes both receiving and transmitting hardware, that is, a device with receiving and transmitting hardware capable of performing bidirectional communication on a bidirectional communication link. Such a client device may include cellular or other communication devices, having a single-line display, a multi-line display, or a cellular or other communication device without a multi-line display. Specifically, the client 100 may be a desktop terminal or a mobile terminal, specifically a mobile phone, tablet computer, laptop computer, etc.

[0076] Those skilled in the art will understand that Figure 1 The application environment shown is merely one application scenario of the solution in this application and does not constitute a limitation on the application scenario of the solution in this application. Other application environments may include those that are more specific to this application. Figure 1 The number of servers shown, or the server network connectivity relationships, for example... Figure 1 Only one server and two clients are shown in the diagram. It is understood that the defect detection system of this display panel may also include one or more other servers, and / or one or more clients connected to the server network, which is not limited here.

[0077] In some embodiments of the present invention, the working platform may be an enterprise office platform, such as WeChat for Business. Taking server 200 as an example, it may further include an enterprise office platform contact server, an enterprise office platform configuration management server, and a web management server. Enterprise users or developers can access the web management server using a web browser terminal to configure the field configuration information on the enterprise office platform configuration management server, and set and store the enterprise user information of enterprise employees of the enterprise office platform on the enterprise office platform contact server.

[0078] In addition, such as Figure 2 As shown, Figure 2 This is a schematic diagram of another embodiment of the display panel defect detection system provided in this invention. The display panel defect detection system may further include a storage terminal 300 for storing data, such as a storage object database. The object database stores object data, which may include application templates (such as approval templates, attendance templates, and other application templates), file data (such as Word files, Excel files, or PPT files, and other files in various formats), image data (such as JPG, PNG, BMP, and other images in various formats), and so on. Correspondingly, the object database may also be divided into multiple types of data, such as an application database, a file database, or an image database.

[0079] It should be noted that, Figure 1-2The schematic diagram of the display panel defect detection system shown is merely an example. The display panel defect detection system and scenario described in the embodiments of the present invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of the present invention, and do not constitute a limitation on the technical solutions provided by the embodiments of the present invention. As those skilled in the art will know, with the evolution of display panel defect detection systems and the emergence of new business scenarios, the technical solutions provided by the embodiments of the present invention are also applicable to similar technical problems.

[0080] The following detailed description is based on specific embodiments.

[0081] In this embodiment, the description will be from the perspective of a defect detection device for the display panel, which can be integrated into the server 200.

[0082] This invention provides a method for detecting defects in a display panel. Please refer to [link / reference]. Figure 3 , Figure 3 This is a flowchart illustrating an embodiment of a defect detection method for a display panel provided by the present invention, including:

[0083] S301: The global and local features of the display panel under test are obtained through the lightweight sampling module of the target defect detection network model to obtain the fused feature map of the display panel under test. The target defect detection network model includes a lightweight sampling module, a defect recognition module, a frequency domain feature acquisition module, and a deconvolution module.

[0084] In one specific embodiment, the target defect detection network model can be any of the following: deep learning network model, neural network model, generative model, unsupervised / self-supervised model, multi-scale feature fusion network model, and diffusion model, without limitation herein.

[0085] The display panel under test refers to the display panel that needs to be tested for defects (such as Mura defects, i.e., uneven brightness or color defects) during production, quality inspection and other processes.

[0086] The lightweight sampling module samples the input data of the display panel under test to efficiently extract its global and local features. It's important to note that during the feature acquisition process, the lightweight sampling module performs regional or downsampling processing on the display panel according to certain rules. This ensures that both the overall features (global features) of the display panel and the regional features of all areas or specific local features are obtained.

[0087] A fused feature map refers to a feature map of the display panel under test that includes both global and local features.

[0088] S302: The defect identification module identifies defects in the fused feature map to obtain the defect location information of the display panel under test;

[0089] In one specific embodiment, the defect identification module is a core functional module that accurately identifies and outputs the specific location information of defects in the display panel under test by performing key information extraction, feature matching, or classification regression processing on the fused feature map.

[0090] Defect location information is the spatial localization result of the extracted fused feature map. It is usually expressed as the bounding box coordinates of the defect area (such as the pixel coordinates of the upper left and lower right corners), the center point coordinates, or the pixel-level area mask, etc., and is used to accurately or preliminarily indicate the spatial location of the defect in the display panel image.

[0091] S303: The frequency domain features of the fused feature map are acquired through the frequency domain feature acquisition module;

[0092] In one specific embodiment, the frequency domain feature acquisition module is a core functional module that extracts frequency characteristic information related to display panel defects by converting the fused feature map from the spatial domain to the frequency domain (such as high-frequency components corresponding to defect edges / details, and low-frequency components corresponding to large-area brightness unevenness).

[0093] Frequency domain features are the distribution and intensity information of different frequency components in the frequency domain of the fused feature map. They reflect the speed and periodicity of the grayscale changes of pixels (or feature values) in the fused feature map.

[0094] S304: The defect location information and frequency domain features are fused and upsampled by the deconvolution module to obtain the defect detection map of the display panel under test.

[0095] In one specific embodiment, the deconvolution module is the core functional module that restores low-resolution abstract features (integrating defect location information and frequency domain features) to a high-resolution space, while preserving and enhancing the key details required for defect detection, and finally outputting an accurate defect detection map.

[0096] A defect detection image is an image that can clearly and intuitively show the specific location, shape, and detailed features of defects in the display panel under test.

[0097] In this embodiment of the invention, a lightweight sampling module acquires global and local features of the display panel under test, significantly reducing the time required for feature acquisition. By performing a balanced fusion process on the global and local features, a fused feature map of the display panel under test is obtained, ensuring that both global and local features contribute fully to the fusion process. The output fused feature map contains both overall context and detailed defect information. A defect identification module identifies defects in the fused feature map, obtaining the defect location information of the display panel under test, enabling more accurate differentiation between defective and normal areas. A frequency domain feature acquisition module acquires the frequency domain features of the fused feature map, and a deconvolution module fuses and upsamples the defect location information and frequency domain features to obtain a defect detection map of the display panel under test. Since frequency domain features can capture structural features that are difficult to express in the spatial domain, integrating defect location information and frequency domain features allows for more accurate defect localization. In summary, this invention significantly improves the efficiency of defect identification and the accuracy of defect localization.

[0098] In one specific embodiment, in S301, the lightweight sampling module includes a lightweight downsampling module and a feature balancing module. To obtain the global and local features of the display panel under test through the lightweight sampling module, the process includes: sequentially performing single-channel spatial feature extraction and cross-channel feature fusion on the display panel under test using the lightweight downsampling module to obtain the lightweight downsampling features of the display panel under test; and performing multi-scale feature fusion balancing processing on the lightweight downsampling features according to the feature balancing module to obtain the global and local features of the display panel under test.

[0099] Specifically, the lightweight downsampling module sequentially performs "single-channel spatial feature extraction" and "cross-channel feature fusion," capturing both global structural features (such as the overall display area distribution and macroscopic anomalies) and local detail features (such as microscopic defects like minor scratches and local bright / dark spots) of the display panel under test. Global features ensure a grasp of the overall panel structure, avoiding the neglect of cross-regional related defects; local features focus on detailed anomalies. The collaborative extraction of both comprehensively covers the feature dimensions of display panel defects, avoiding information loss caused by single feature extraction. The lightweight sampling module simplifies the computational process of feature extraction and fusion (e.g., "single-channel spatial feature extraction" reduces redundant computation, and the "lightweight" structure reduces the number of parameters), significantly reducing the computational complexity and resource consumption of the model while ensuring feature extraction performance, thus making the defect detection process more efficient.

[0100] The "multi-scale feature fusion balancing processing" module solves the imbalance problems that may exist in traditional fusion, such as "global features suppressing local details" or "local noise interfering with the global structure," by dynamically adjusting the fusion weights of global and local features. The fused feature map obtained after balancing processing can both preserve the integrity of the global structure to locate defect areas and highlight the salience of local details to distinguish defect types (such as scratches, stains, and dead pixels), thereby significantly improving the detection accuracy of defects at different scales and locations (especially the recognition rate of small, low-contrast defects).

[0101] In summary, this embodiment achieves a balance between "high efficiency" and "high precision" in display panel defect detection by reducing data processing volume and improving detection speed through lightweight design and balanced fusion of global and local features.

[0102] Furthermore, the lightweight downsampling module includes a depthwise convolution submodule and a pointwise convolution submodule; the lightweight downsampling features include single-channel spatial features and cross-channel features; the lightweight downsampling module sequentially extracts single-channel spatial features and fuses cross-channel features of the display panel under test to obtain the lightweight downsampling features of the display panel under test, including: independently convolving each input channel of the display panel under test according to the depthwise convolution submodule to obtain the single-channel spatial features of the display panel under test; and performing linear combination and dimensional transformation operations on the channel dimensions of the display panel under test through the pointwise convolution submodule to obtain the cross-channel features of the display panel under test.

[0103] Specifically, the deep convolution submodule focuses on independent convolutions within a single channel, avoiding cross-channel interference. This allows for the precise capture of local spatial details of the display panel under test (such as local features like texture and structure in specific areas), thus preserving local spatial features. Meanwhile, the pointwise convolution submodule achieves cross-channel linear combination and dimensionality transformation through 1x1 convolutions, integrating semantic relationships between different channels (such as the correlation between different areas of the panel) to generate global features. This allows for the fusion of global channel information. Therefore, through the synergistic effect of the deep convolution and pointwise convolution submodules, the lightweight downsampling module can simultaneously output both local spatial features and global channel features, avoiding information loss caused by relying on a single feature (only local or only global), and improving the richness and completeness of feature representation.

[0104] In one specific embodiment, the depthwise convolution submodule performs independent convolution on each input channel (each channel corresponds to one convolution kernel), which, compared to traditional convolution (input channel * output channel * kernel size), achieves higher efficiency. 2 The parameters and computational cost are significantly reduced (only input channels * kernel size). 2The pointwise convolution submodule uses 1*1 convolution for linear channel combination, further simplifying the computation. Therefore, by combining the depthwise convolution submodule and the pointwise convolution submodule, the data processing difficulty is significantly reduced while maintaining feature extraction capabilities, thus improving the efficiency of defect detection.

[0105] In summary, the lightweight downsampling module, through the collaborative work of the depthwise convolution submodule and the pointwise convolution submodule, significantly reduces the amount of data computation while retaining the sensitivity of depthwise convolution to local spatial details. It also achieves inter-channel information exchange through pointwise convolution, avoiding the degradation of feature representation capabilities caused by lightweight design. The resulting "global + local features" more comprehensively reflect the overall characteristics and local details of the display panel under test, facilitating more accurate defect detection.

[0106] In one specific embodiment, the feature balancing module includes a local depthwise convolution submodule, a nonlinear feature extraction submodule, a lightweight self-attention submodule, a feedforward network submodule, and a secondary balancing and fusion submodule. The feature balancing module performs multi-scale feature fusion balancing on global and local features to obtain a fused feature map of the display panel under test. This includes: channel segmentation of global and local features according to the gamma distribution factor to obtain gamma partial features and remaining features; depthwise convolution processing of the gamma partial features by the local depthwise convolution submodule to obtain spatial features of the gamma partial features; multi-level nonlinear transformation of the remaining features and spatial features by the nonlinear feature extraction submodule to obtain multi-level nonlinear features; self-attention feature extraction of the multi-level nonlinear features by the lightweight self-attention submodule to obtain key features; fusion processing of the key features and multi-level nonlinear features by the feedforward network submodule to obtain feedforward fused features; and secondary gamma channel segmentation of the feedforward fused features and key features by the secondary balancing and fusion submodule, followed by depthwise convolution and normalized activation processing of the segmented features to obtain the fused feature map.

[0107] Specifically, channel segmentation is performed based on the gamma distribution factor to selectively separate components with different distribution characteristics (such as key information and secondary information) in the features, and to filter out the gamma-part features (core information) and the remaining features (secondary information) to achieve targeted feature processing; through a local depthwise convolution submodule, depthwise convolution processing is performed on the gamma-part features to obtain the spatial features of the gamma-part features, while keeping the remaining features unchanged, thereby enhancing the spatial representation of key features while reducing the amount of computation.

[0108] The nonlinear feature extraction submodule breaks through the limitations of linear models by performing multi-level nonlinear transformations (such as activation functions and nonlinear mappings) on the remaining features and spatial features. It gradually abstracts complex patterns in the remaining features and spatial features (such as irregular shapes and texture details of display panel defects) to obtain multi-level nonlinear features, thus realizing the extraction of more complex, robust and semantically meaningful features from the remaining features and spatial features.

[0109] The lightweight self-attention submodule simplifies the attention mechanism (such as reducing the number of heads and low-dimensional projection), while retaining the ability to capture global associations and avoiding the high complexity of traditional self-attention. It further focuses on key regions (such as defect candidate regions and texture salient regions) in multi-level nonlinear features and suppresses irrelevant noise and redundant information interference.

[0110] The feedforward network submodule integrates key features (core information extracted by self-attention) with multi-level nonlinear features (hierarchical features that are progressively abstracted), thereby further integrating feature information at different levels of abstraction and significantly improving the nonlinear representation capability of features.

[0111] After the secondary gamma channel segmentation, the secondary balancing and fusion submodule stabilizes the feature distribution through normalized activation processing (such as BN normalization + activation function), avoids gradient vanishing / exploding, and accelerates model training convergence. At the same time, the combination of depthwise convolution and normalization further standardizes the feature scale, reduces the impact of input perturbations (such as changes in illumination and differences in panel material) on feature extraction, and improves the robustness of the fused feature map.

[0112] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a feature balancing module provided in an embodiment of the present invention. In this module, DWConv 3*3 is a local depthwise convolution submodule, BN and Swish are nonlinear feature extraction submodules, LMHSA is a lightweight self-attention submodule, FFN is a feedforward network submodule, and the portion within the dashed box is the secondary balancing and fusion submodule. Conv 1*1 indicates that no processing is performed on the output image; that is, the input and output are identical.

[0113] In summary, by using gamma distribution factor channel segmentation and quadratic gamma channel segmentation, components with different distribution characteristics (such as key information and secondary information) in the features are selectively separated. Combined with local depthwise convolution to capture local spatial details and lightweight self-attention to capture global correlations, a precise balance is achieved between global features (long-range dependencies) and local features (fine spatial structure). Multi-scale fusion avoids the one-sidedness of single-scale features, enabling the fused feature map to simultaneously represent global context and local details, thus improving feature integrity.

[0114] In one specific embodiment, in S302, please refer to Figure 5 , Figure 5 This is a schematic diagram of the structure of a defect identification module provided in an embodiment of the present invention. The defect identification module includes a residual branch (identify), a dynamic convolutional branch (DFC), a first ghost convolutional network submodule (Ghost1), and a second ghost convolutional network submodule (Ghost2). The defect identification module performs defect identification on the fused feature map to obtain the defect location information of the display panel under test. This includes: performing convolution and normalization processing on the fused feature map according to the residual branch to obtain basic detail features; adaptively adjusting the feature extraction method of the fused feature map according to the dynamic convolutional branch to obtain a convolution weight matrix; expanding the number of channels of the fused feature map according to the first ghost convolutional network submodule to obtain an expanded redundant feature map; performing dynamic weighting and dimensionality compression processing on the redundant feature map according to the convolution weight matrix of the second ghost convolutional network submodule to obtain high-dimensional features; concatenating the high-dimensional features and basic detail features by channels to obtain a multi-scale feature space; and determining the defect location information based on the multi-scale feature space.

[0115] Specifically, the residual branch includes a 3x3 convolutional layer, two batch normalization (BN) layers, and a 1x1 convolutional layer. The 3x3 convolutional layer extracts spatial details from the fused feature map, followed by normalization through the first BN layer to stabilize the feature distribution. Then, the 1x1 convolutional layer compresses the channel dimension to further extract basic detail features. Finally, the second BN layer performs normalization to improve the stability of the residual branch's output features.

[0116] The Dynamic Convolutional Branch (DFC) comprises an average pooling layer, a 3x3 convolutional layer, three batch normalization (BN) layers, a 1x5 convolutional layer, a 5x1 convolutional layer, and an interpolation layer. The average pooling layer performs global feature compression on the fused feature map to obtain a global feature vector. This feature vector undergoes feature transformation via a 3x3 convolutional layer and is normalized by the first BN layer to further extract global feature representations. Subsequently, multi-scale channel feature mapping is performed using the 1x5 and 5x1 convolutional layers to output dynamic convolutional weights. Finally, the interpolation layer maps the convolutional weights back to the original feature map size, resulting in a dynamic weight matrix that matches the input feature map. Through this structure, the Dynamic Convolutional Branch can adaptively determine the convolutional weight matrix.

[0117] The first ghost convolutional network submodule consists of a 3x3 convolutional layer, two BN+ReLU layers, a 1x1 convolutional layer, and a Contact concatenation layer. Local features are extracted from the fused feature map using the 3x3 convolutional layer and activated by the first BN+ReLU layer to enhance feature representation. Subsequently, the 1x1 convolutional layer expands the feature channels to generate extended features. These extended features are then processed by the second BN+ReLU layer and input to the Contact concatenation layer to form a redundant feature map.

[0118] The second ghost convolutional network submodule has the same structure as the first ghost convolutional network submodule, but its input is a weighted result of redundant feature maps and dynamic convolutional weight matrices. Local features are extracted from the weighted redundant feature maps through 3*3 convolutional layers and activated by the first BN+ReLU layer to enhance feature representation. Subsequently, the feature channels are compressed using 1*1 convolutional layers to generate compressed features. After being processed by the second BN+ReLU layer, the compressed features are input to the Contact concatenation layer to form high-dimensional features. This structure can dynamically adjust the feature distribution of redundant feature maps, further enhancing the accuracy and adaptability of feature extraction.

[0119] After obtaining the high-dimensional features, in order to determine the defect location information of the display panel under test, it is also necessary to perform channel splicing of the high-dimensional features and the basic detail features to obtain a multi-scale feature space, and then accurately determine the defect location information based on the multi-scale feature space.

[0120] In this embodiment, through the collaborative mechanism of "detail preservation - dynamic adaptation - efficient expansion - multi-scale fusion", the accuracy of defect location information (especially for small / complex defects) is improved, while taking into account computational efficiency and scene adaptability, providing high-precision, high-efficiency and high-robust technical support for display panel defect detection.

[0121] In one specific embodiment, in S303, the frequency domain feature acquisition module includes a dual-tree complex wavelet transform submodule and an inverse dual-tree complex wavelet transform submodule; the frequency domain features of the fused feature map are acquired through the frequency domain feature acquisition module, including: filtering the fused feature map through the dual-tree complex wavelet transform submodule to obtain low-frequency approximate components and high-frequency detail components; and jointly reconstructing the real and imaginary features of the low-frequency approximate components and high-frequency detail components through the inverse dual-tree complex wavelet transform submodule to obtain the frequency domain features.

[0122] It should be noted that the Dual-Tree Complex Wavelet Transform (DTCWT) can accurately decompose the fused feature map into low-frequency approximate components (corresponding to the overall structure and smooth areas of the display panel) and high-frequency detail components (corresponding to local details such as edges, textures, and minor defects). By separating the high-frequency detail components, the masking of minor defects (such as microcracks, bright spots, dark spots, scratches, etc.) by low-frequency information can be effectively avoided, thus improving the sensitivity of extracting subtle defect features.

[0123] The complex nature of dual-tree complex wavelet transform allows its output to include both real and imaginary features. Phase information is crucial for describing the structural contours, edge continuity, and local deformations (such as minute protrusions / depressions caused by defects) of an image. By jointly reconstructing the real and imaginary features, phase characteristics can be fully preserved, avoiding structural distortion caused by the loss of phase information in traditional real wavelet transforms. This results in a more accurate representation of the geometric morphology and spatial distribution of defects, enhancing the distinction between defects and background noise.

[0124] Dual-tree complex wavelet transform achieves approximate translation invariance through its dual-tree structure. When the fused feature map undergoes a slight translation due to acquisition deviation or panel positional variations, the low-frequency / high-frequency components obtained from the decomposition change minimally, avoiding feature fluctuations caused by "translation sensitivity" in traditional wavelet transform. This ensures that consistent frequency domain features can be stably extracted from display panel images under different positions and orientations, reducing the risk of missed / false detections of defects caused by minor perturbations in the input image.

[0125] The inverse dual-tree complex wavelet transform jointly reconstructs the separated low-frequency approximation components (overall structural information) and high-frequency detail components (local defect information), achieving complementary fusion of features at different frequency scales. The resulting frequency domain features contain both the overall structural context of the panel (avoiding misjudgments caused by isolated analysis of details) and retain the local details of defects (avoiding the overall information from obscuring minor defects), giving the features a more comprehensive representational capability for better defect identification.

[0126] In summary, by leveraging the high-frequency detail preservation, phase information integrity, and translation stability of dual-tree complex wavelet transform, combined with multi-frequency feature fusion achieved through joint reconstruction of real and imaginary parts, the detail capture accuracy, structural characterization accuracy, and extraction stability of frequency domain features of display panel defects are significantly improved, thereby enhancing the sensitivity, robustness, and recognition accuracy of defect detection.

[0127] In one specific embodiment, in S304, the defect location information and frequency domain features are fused and upsampled by the deconvolution module to obtain a defect detection map of the display panel under test. This includes: aligning the defect location information and frequency domain features to obtain defect location information and frequency domain features aligned in spatial resolution and feature dimension; stitching the defect location information and frequency domain features aligned in spatial resolution and feature dimension along the channel dimension to obtain a time-frequency fusion feature map; and performing a deconvolution operation on the time-frequency fusion feature map to restore the time-frequency fusion feature map from a small size and high semantic information to a larger size spatial resolution to obtain the defect detection map.

[0128] In this embodiment, by aligning the spatial resolution with the feature dimension, the defect location information (spatial domain positioning features) and frequency domain features (texture and frequency detail features) are effectively fused at the same scale, avoiding information loss or interference caused by feature misalignment. The time-frequency fusion feature map formed by channel splicing covers both spatial positioning and frequency domain detail information, enhancing the characterization ability of different types of defects (such as fine texture defects, blurred edge defects, etc.) and improving detection accuracy.

[0129] The time-frequency fusion feature map integrates spatial location constraints and frequency domain detail features, which can adapt to the feature differences of different defects (such as brightness defects, structural defects, etc.) and reduce the limitations of a single feature mode. Alignment processing optimizes feature utilization efficiency, avoids redundant information interference, and makes the model more robust to the detection of complex backgrounds, noise interference, or weak defects.

[0130] The deconvolution operation recovers high-resolution feature maps through upsampling. Combined with the semantic space mapping of the convolution kernel, the fused high-dimensional features are accurately mapped to the target semantic space of defect detection (such as defect region mask or probability map), which effectively improves the spatial resolution of the output detection map, making the defect boundary clearer and the localization more precise (such as sub-pixel level localization).

[0131] In summary, through the integration of the deconvolution module and upsampling processing, the final output defect detection map achieves both location accuracy and detail integrity through multimodal feature fusion and high-resolution recovery, ensuring the reliability of defect detection.

[0132] In summary, for a clear demonstration of the global structure of the target defect detection network model, please refer to [link / reference needed]. Figure 6 , Figure 6 This is a schematic diagram of the structure of the first embodiment of the target defect detection network model provided in this invention. DSConv+LUB is a lightweight sampling module, DGN is a frequency domain feature acquisition module (within the dashed box), and transConv is a deconvolution module. This structure clearly reflects the direction and changes of image data from the display panel within the target defect detection network model, enabling defect detection of the display panel.

[0133] It should be noted that, in order to improve the reliability of defect detection results, in Figure 6 Based on this, the structure of the target defect detection network model can be further restricted. Please refer to [link / reference]. Figure 7 , Figure 7 This is a schematic diagram of the second embodiment of the target defect detection network model provided in this invention. In the process of performing dual-tree complex wavelet transform, the defect identification module only uses the last lightweight downsampling result as input. Because the lightweight downsampling module downsamples the image of the display panel under test multiple times, it greatly reduces the amount of data while preserving image features. Therefore, the amount of data processing is greatly reduced during the dual-tree complex wavelet transform. In the inverse dual-tree complex wavelet transform, to improve the reliability of the data transformation result, the output of the feature balancing module LUB is used as reference data. Furthermore, the deconvolution module is divided into multiple steps. Specifically, in the first deconvolution process, the results of the defect identification module and the dual-tree complex wavelet transform are used as the basis. In the second deconvolution process, the result of the inverse dual-tree complex wavelet transform is also added. In the final deconvolution process, the result of the first convolution is used as the basis to adaptively adjust the final deconvolution result, so that the final deconvolution result can better meet the requirements of the display panel.

[0134] Furthermore, to improve the reliability of the target defect detection network model, before obtaining the global and local features of the display panel under test through the lightweight sampling module of the target defect detection network model, it is also necessary to: construct an initial defect detection network model, which includes a weighted loss function, comprising a gradient coordination mechanism focus loss function, a structural similarity loss function, and a mean squared error loss function; obtain a training sample set, which includes input data samples and target output label samples; input the input data samples into the initial defect detection network model to obtain the predicted label samples corresponding to the input data samples; obtain the gradient norm of the training sample set and the predicted label samples, and construct a gradient model based on the gradient norm. The gradient norm space is dynamically balanced according to the gradient coordination mechanism focus loss function to obtain the weights assigned to the training sample set, and the gradient coordination mechanism focus loss value of the training sample set is determined based on the assigned weights. The structural similarity loss value of the input data sample and the predicted label sample is calculated and determined according to the structural similarity loss function. The mean squared error loss value of the input data sample and the predicted label sample is calculated and determined according to the mean squared error loss function. The gradient coordination mechanism focus loss value, structural similarity loss value and mean squared error loss value are weighted to obtain the target loss function value, and the initial defect detection network model is adjusted based on the target loss function value until a fully trained target defect detection network model is obtained.

[0135] Specifically, this embodiment employs weighted loss functions to differentiate and optimize different types of defect detection tasks. Specifically: the gradient coordination mechanism focus loss function alleviates class imbalance and improves sensitivity to minute defect regions; the structural similarity loss function measures the structural consistency between input data samples and predicted label samples, enhancing the model's ability to represent defect edges and texture details; and the mean squared error loss function focuses on pixel-level error control, improving the overall detection accuracy. The synergistic effect of these three functions enables the model to efficiently utilize both global and local features during training, thereby further improving the accuracy and robustness of defect detection.

[0136] In practical applications, during the weighted processing of the gradient coordination mechanism focus loss, structural similarity loss, and mean squared error loss to obtain the target loss function value, due to the bias in the training objective, the following measures are taken to improve the reliability of model training: The training sample set is divided into a first batch and a second batch. During the training of the initial defect detection network model based on the first batch, the gradient coordination mechanism focus loss, structural similarity loss, and mean squared error loss are weighted based on a first weight combination. During the training of the initial defect detection network model based on the second batch, the gradient coordination mechanism focus loss, structural similarity loss, and mean squared error loss are weighted based on a second weight combination. Specifically, in the first weight combination, the weight of the gradient coordination mechanism focus loss is greater than the weight of the structural similarity loss; in the second weight combination, the weight of the structural similarity loss is significantly greater than the weight of the gradient coordination mechanism focus loss.

[0137] It should be noted that both the first and second batches of samples include the complete training dataset, and the sum of the first and second batches is the total number of training iterations of the initial defect detection network model.

[0138] In one specific embodiment, the weighted loss function The calculation formula is:

[0139]

[0140] in, It is a constant. The focus loss function of the gradient coordination mechanism. Let S be the structural similarity loss function. This is the mean squared error loss function.

[0141] The GHM-Focal loss function is as follows:

[0142] Let the training dataset be ,in, It is the input image. It is the target segmentation mask (for each pixel, the value is 0 (background) or 1 (defect)); the model parameters are denoted as The model predicts that ,in It is the softmax function (multi-class classification). It is the output of the model. logits .

[0143] GHM (Gradient Harmonizing Mechanism) divides the gradient norm space into M intervals. The gradient norm space is the range or set of values ​​formed by the gradient norms of all samples, and each interval has a width of M. Order No. j The intervals are ,in The loss function is as follows:

[0144]

[0145]

[0146]

[0147]

[0148]

[0149]

[0150] in, This is the modulation factor, typically set to 0.25 to balance positive and negative samples; The gradient norm (absolute value); It is the gradient density function; The predicted probability of the target category; The cross-entropy loss function; For gradient; when exist When within a neighborhood, =1, otherwise =0; d The interval length; The interval length of this neighborhood where the sample size is calculated is used as a standardization factor; The number of samples; A represents the exponential function of A.

[0151] Structural similarity loss function The calculation formula is:

[0152]

[0153] in, , The input images are respectively x and output image y The standard deviation within the local window. and It is a constant. , Images x and y The mean within a local window, For image x and images y Covariance within a local window.

[0154] Mean squared error loss function The calculation formula is:

[0155]

[0156] in, For weights, defect areas Background area =0; This represents the actual grayscale deviation. This represents the model prediction error.

[0157] In the actual training process of the initial defect detection network model, assuming a batch size of 32, and using the SGD optimizer to update the network parameters, then during the initial training phase (0-50 epochs) after 100 epochs of network training: > > Prioritize defect classification and structural fidelity to avoid the model being dominated by noise.

[0158] In the later stages of training (after 50 epochs): This enhances pixel-level positioning accuracy and optimizes defect boundaries and grayscale deviations.

[0159] To facilitate better implementation of the display panel defect detection method provided in the embodiments of the present invention, the present invention also provides an apparatus based on the above-described display panel defect detection method. The meanings of the terms used are the same as in the above-described display panel defect detection method, and specific implementation details can be found in the descriptions in the method embodiments.

[0160] Please see Figure 8 , Figure 8This is a schematic diagram of an embodiment of a defect detection device for a display panel provided by the present invention. The defect detection device 800 for the display panel may include:

[0161] The fusion feature map acquisition module 801 is used to acquire the global and local features of the display panel under test through the lightweight sampling module of the target defect detection network model, and to perform balanced fusion processing on the global and local features to obtain the fusion feature map of the display panel under test. The target defect detection network model includes a lightweight sampling module, a defect recognition module, a frequency domain feature acquisition module, and a deconvolution module.

[0162] The defect location information acquisition module 802 is used to identify defects in the fused feature map through the defect recognition module to obtain the defect location information of the display panel under test.

[0163] The frequency domain feature acquisition module 803 is used to acquire the frequency domain features of the fused feature map through the frequency domain feature acquisition module;

[0164] The defect detection module 804 is used to perform feature fusion and upsampling on defect location information and frequency domain features through the deconvolution module to obtain the defect detection map of the display panel under test.

[0165] This invention also provides an electronic device, such as... Figure 9 As shown, Figure 9 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention, specifically:

[0166] The electronic device may include components such as a processor 901 with one or more processing cores, a memory 902 with one or more computer-readable storage media, a power supply 903, and an input unit 904. Those skilled in the art will understand that... Figure 9 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0167] The processor 901 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 902, and by calling data stored in the memory 902, it performs various functions and processes data, thereby providing overall monitoring of the electronic device. Optionally, the processor 901 may include one or more processing cores; preferably, the processor 901 may integrate an application processor and a modem processor. The application processor mainly handles the operation of the storage medium, user interface, and application programs, while the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 901.

[0168] The memory 902 can be used to store software programs and modules. The processor 901 executes various functional applications and data processing by running the software programs and modules stored in the memory 902. The memory 902 may mainly include a program storage area and a data storage area. The program storage area may store application programs required for operating the storage medium and at least one function (such as sound playback function, image playback function, etc.); the data storage area may store data created according to the use of the electronic device. In addition, the memory 902 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 902 may also include a memory controller to provide the processor 901 with access to the memory 902.

[0169] The electronic device also includes a power supply 903 that supplies power to various components. Preferably, the power supply 903 can be logically connected to the processor 901 via a power management storage medium, thereby enabling functions such as charging, discharging, and power consumption management through the power management storage medium. The power supply 903 may also include one or more DC or AC power supplies, recharge storage media, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0170] The electronic device may also include an input unit 904, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0171] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 901 in the electronic device loads the executable files corresponding to the processes of one or more applications into the memory 902 according to the following instructions, and the processor 901 runs the applications stored in the memory 902 to realize various functions, as follows:

[0172] The target defect detection network model acquires global and local features of the display panel under test through a lightweight sampling module, resulting in a fused feature map of the display panel. The target defect detection network model includes a lightweight sampling module, a defect identification module, a frequency domain feature acquisition module, and a deconvolution module. The defect identification module identifies defects in the fused feature map to obtain the defect location information of the display panel under test. The frequency domain feature acquisition module acquires the frequency domain features of the fused feature map. The deconvolution module fuses and upsamples the defect location information and frequency domain features to obtain the defect detection map of the display panel under test.

[0173] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0174] Therefore, embodiments of the present invention provide a computer-readable storage medium storing a computer program thereon, the computer program being loaded by a processor to execute the steps in any of the defect detection methods for display panels provided in the embodiments of the present invention. For example, the computer program, when loaded by a processor, can execute the following steps:

[0175] The target defect detection network model acquires global and local features of the display panel under test through a lightweight sampling module, resulting in a fused feature map of the display panel. The target defect detection network model includes a lightweight sampling module, a defect identification module, a frequency domain feature acquisition module, and a deconvolution module. The defect identification module identifies defects in the fused feature map to obtain the defect location information of the display panel under test. The frequency domain feature acquisition module acquires the frequency domain features of the fused feature map. The deconvolution module fuses and upsamples the defect location information and frequency domain features to obtain the defect detection map of the display panel under test.

[0176] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0177] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0178] Since the computer program stored in the computer-readable storage medium can execute the steps in any of the defect detection methods for display panels provided in the embodiments of the present invention, the beneficial effects that any of the defect detection methods for display panels provided in the embodiments of the present invention can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0179] The above provides a detailed description of a defect detection method for a display panel provided by an embodiment of the present invention. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting defects of a display panel, characterized in that, include: The target defect detection network model acquires global and local features of the display panel under test through a lightweight sampling module, resulting in a fused feature map of the display panel. The target defect detection network model includes the lightweight sampling module, a defect identification module, a frequency domain feature acquisition module, and a deconvolution module. The lightweight sampling module includes a lightweight downsampling module and a feature balancing module. Acquiring global and local features of the display panel under test through the lightweight sampling module involves: sequentially performing single-channel spatial feature extraction and cross-channel feature fusion on the display panel under test using the lightweight downsampling module to obtain lightweight downsampled features; and performing multi-scale feature fusion balancing processing on the lightweight downsampled features according to the feature balancing module to obtain the global and local features of the display panel under test. The feature balancing module includes a local depthwise convolution submodule, a nonlinear feature extraction submodule, a lightweight self-attention submodule, a feedforward network submodule, and a secondary balancing and fusion submodule. The feature balancing module performs multi-scale feature fusion balancing on the lightweight downsampled features to obtain the fused feature map of the display panel under test. This includes: segmenting the global and local features according to the gamma distribution factor to obtain gamma partial features and remaining features; performing depthwise convolution on the gamma partial features through the local depthwise convolution submodule to obtain the spatial features of the gamma partial features; performing multi-level nonlinear transformation on the remaining features and the spatial features according to the nonlinear feature extraction submodule to obtain multi-level nonlinear features; performing self-attention feature extraction on the multi-level nonlinear features according to the lightweight self-attention submodule to obtain key features; fusing the key features and multi-level nonlinear features according to the feedforward network submodule to obtain feedforward fused features; and performing secondary gamma channel segmentation on the feedforward fused features and key features according to the secondary balancing and fusion submodule, and then performing depthwise convolution and normalized activation processing on the segmented features to obtain the fused feature map. The defect identification module identifies defects in the fused feature map to obtain the defect location information of the display panel under test. The frequency domain features of the fused feature map are acquired through the frequency domain feature acquisition module; The defect location information and the frequency domain features are fused and upsampled by the deconvolution module to obtain the defect detection map of the display panel under test.

2. The defect detection method for a display panel according to claim 1, characterized in that, The lightweight downsampling module includes a depthwise convolution submodule and a pointwise convolution submodule; the lightweight downsampling features include single-channel spatial features and cross-channel features; The lightweight downsampling module sequentially extracts single-channel spatial features and fuses cross-channel features of the display panel under test to obtain lightweight downsampling features of the display panel under test, including: The depth convolution submodule is used to independently convolve each input channel of the display panel under test to obtain the single-channel spatial features of the display panel under test. The pointwise convolution submodule performs linear combination and dimension transformation operations on the channel dimensions of the display panel under test to obtain the cross-channel features of the display panel under test.

3. The defect detection method for a display panel according to claim 1, characterized in that, The defect identification module includes a residual branch, a dynamic convolution branch, a first ghost convolutional network submodule, and a second ghost convolutional network submodule; the step of identifying defects in the fused feature map through the defect identification module to obtain the defect location information of the display panel under test includes: The fused feature map is convolved and normalized according to the residual branch to obtain basic detail features; The feature extraction method of the fused feature map is adaptively adjusted according to the dynamic convolution branch to obtain the convolution weight matrix; The fused feature map is expanded by increasing the number of channels according to the first ghost convolutional network submodule to obtain an expanded redundant feature map; The second ghost convolutional network submodule performs dynamic weighting and dimensionality compression on the redundant feature map according to the convolutional weight matrix to obtain high-dimensional features; Channel concatenation is performed between the high-dimensional features and the basic detail features to obtain a multi-scale feature space; The defect location information is determined based on the multi-scale feature space.

4. The defect detection method for a display panel according to claim 1, characterized in that, The frequency domain feature acquisition module includes a dual-tree complex wavelet transform submodule and an inverse dual-tree complex wavelet transform submodule; the frequency domain features of the fused feature map acquired by the frequency domain feature acquisition module include: The fused feature map is filtered by the dual-tree complex wavelet transform submodule to obtain low-frequency approximate components and high-frequency detail components. The frequency domain features are obtained by jointly reconstructing the real and imaginary features of the low-frequency approximation component and the high-frequency detail component through the inverse dual-tree complex wavelet transform submodule.

5. The defect detection method for a display panel according to claim 1, characterized in that, The step of fusing and upsampling the defect location information and the frequency domain features using the deconvolution module to obtain the defect detection map of the display panel under test includes: The defect location information and the frequency domain features are aligned to obtain defect location information and frequency domain features aligned in spatial resolution and feature dimension. The defect location information and frequency domain features aligned in spatial resolution and feature dimension are concatenated along the channel dimension to obtain a time-frequency fusion feature map. The time-frequency fusion feature map is deconvolved to restore it from a small size with high semantic information to a larger spatial resolution, thus obtaining the defect detection map.

6. The defect detection method for a display panel according to claim 1, characterized in that, Before acquiring the global and local features of the display panel under test through the lightweight sampling module of the target defect detection network model, the following steps are also included: An initial defect detection network model is constructed, which includes a weighted loss function, comprising a gradient coordination mechanism focus loss function, a structural similarity loss function, and a mean squared error loss function. Obtain a training sample set, which includes input data samples and target output label samples; The input data sample is input into the initial defect detection network model to obtain the predicted label sample corresponding to the input data sample; Obtain the gradient norm of the training sample set, and construct the gradient norm space based on the gradient norm; The gradient norm space is dynamically balanced according to the gradient coordination mechanism focus loss function to obtain the assigned weights of the training sample set, and the gradient coordination mechanism focus loss value of the training sample set is determined based on the assigned weights. The structural similarity loss value of the input data sample is calculated and determined based on the structural similarity loss function. The mean squared error loss value of the input data sample is calculated and determined based on the mean squared error loss function. The focus loss value of the gradient coordination mechanism, the structural similarity loss value, and the mean square error loss value are weighted to obtain the target loss function value. The initial defect detection network model is then adjusted based on the target loss function value until a fully trained target defect detection network model is obtained.

7. The defect detection method for a display panel according to claim 6, characterized in that, The weighted processing of the focus loss value of the gradient coordination mechanism, the structural similarity loss value, and the mean squared error loss value to obtain the target loss function value includes: The training sample set is divided into a first batch and a second batch; During the training of the initial defect detection network model based on the first batch, the gradient coordination mechanism focus loss value, the structural similarity loss value, and the mean square error loss value are weighted based on the first weight combination; During the training of the initial defect detection network model based on the second batch, the gradient coordination mechanism focus loss value, the structural similarity loss value, and the mean square error loss value are weighted based on the second weight combination; In the first weight combination, the weight value of the focus loss value of the gradient coordination mechanism is greater than the weight value of the structural similarity loss value; in the second weight combination, the weight value of the structural similarity loss value is much greater than the weight value of the focus loss value of the gradient coordination mechanism.

8. A defect detection device for a display panel, characterized in that, include: A fusion feature map acquisition module is used to acquire global and local features of the display panel under test through the lightweight sampling module of the target defect detection network model, thereby obtaining a fusion feature map of the display panel under test. The target defect detection network model includes the lightweight sampling module, a defect recognition module, a frequency domain feature acquisition module, and a deconvolution module. The lightweight sampling module includes a lightweight downsampling module and a feature balancing module. Acquiring global and local features of the display panel under test through the lightweight sampling module includes: sequentially performing single-channel spatial feature extraction and cross-channel feature fusion on the display panel under test through the lightweight downsampling module to obtain lightweight downsampled features of the display panel under test; performing multi-scale feature fusion balancing processing on the lightweight downsampled features according to the feature balancing module to obtain the global and local features of the display panel under test. The feature balancing module includes a local depthwise convolution submodule, a nonlinear feature extraction submodule, a lightweight self-attention submodule, a feedforward network submodule, and a secondary balancing and fusion submodule. The method describes performing multi-scale feature fusion balancing processing on the lightweight downsampled features according to the feature balancing module to obtain the fused feature map of the display panel under test. This includes: segmenting the global and local features according to the gamma distribution factor to obtain gamma partial features and remaining features; performing depthwise convolution processing on the gamma partial features through the local depthwise convolution submodule to obtain the spatial features of the gamma partial features; performing multi-level nonlinear transformation on the remaining features and the spatial features according to the nonlinear feature extraction submodule to obtain multi-level nonlinear features; performing self-attention feature extraction on the multi-level nonlinear features according to the lightweight self-attention submodule to obtain key features; fusing the key features and multi-level nonlinear features according to the feedforward network submodule to obtain feedforward fused features; and performing secondary gamma channel segmentation on the feedforward fused features and key features according to the secondary balancing and fusion submodule, and then performing depthwise convolution processing and normalized activation processing on the segmented features to obtain the fused feature map. The defect location information acquisition module is used to identify defects in the fused feature map through the defect identification module to obtain the defect location information of the display panel under test. A frequency domain feature acquisition module is used to acquire the frequency domain features of the fused feature map through the frequency domain feature acquisition module; The defect detection module is used to perform feature fusion and upsampling on the defect location information and the frequency domain features through the deconvolution module to obtain the defect detection map of the display panel under test.

Images