A micro-led defect detection method based on multi-scale feature enhancement and hypergraph fusion

Abstract

The present application relates to the technical fields of machine vision and industrial defect detection, more particularly, it relates to a Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion, comprising the following steps: S1, obtaining a Micro-LED defect dataset; S2, constructing an end-to-end defect detection deep neural network, training the deep neural network using the defect dataset, and obtaining a trained detection model; the deep neural network comprises a backbone network, a neck network and a prediction module; the backbone network is embedded with a multi-scale edge information enhancement module and a channel-space attention module, and the neck network adopts a hypergraph modulation feature pyramid network; S3, inputting a Micro-LED image to be detected into the trained detection model, performing non-maximum suppression processing on the obtained preliminary detection result, and outputting the category and position of the Micro-LED defect. The present application has the advantages of significant micro-defect edge feature enhancement effect, precise background redundant feature suppression, and sufficient cross-scale high-order information fusion.

Description

Technical Field

[0001] This invention relates to the technical field of machine vision and industrial defect detection, and more specifically, to a Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion. Background Technology

[0002] With the rapid development of new display technologies, Micro-LED display devices, with their advantages of high brightness, low power consumption, long lifespan, and high resolution, are widely used in high-definition displays, wearable devices, and automotive displays. In the mass production process of Micro-LEDs, processes such as mass transfer, bonding, and packaging are prone to various defects, including empty wafers, black spots, misalignment, deflection, breakage, foreign matter, and crystal bridging. These defects directly affect the yield and reliability of the display devices; therefore, defect detection is a crucial quality control step in the production process.

[0003] Existing methods for detecting defects in Micro-LEDs mainly include manual visual inspection, traditional machine vision inspection, and deep learning-based target detection methods. Among them, manual visual inspection suffers from low detection efficiency, high labor costs, large subjective judgment errors, and high false negative rates. Machine vision methods such as template matching and traditional feature extraction rely on manually designed features, which have poor adaptability to lighting fluctuations, sample shifts, background interference, and weak contrast defects, resulting in insufficient robustness.

[0004] In recent years, deep learning detection models such as Faster R-CNN and YOLO series have been gradually applied to industrial defect detection scenarios. However, due to the characteristics of Micro-LED defects, such as small size, dense distribution, high class similarity, and complex background texture, existing deep learning methods are still prone to problems such as insufficient expression of small target features, low efficiency of multi-scale information fusion, and obvious interference from redundant background information. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion, which has the advantages of significant enhancement of small defect edge features, accurate suppression of background redundancy features, and sufficient fusion of cross-scale high-order information.

[0006] The above-mentioned technical objective of this invention is achieved through the following technical solution: a Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion, comprising the following steps:

[0007] S1. Obtain Micro-LED defect dataset;

[0008] S2. Construct an end-to-end deep neural network for defect detection, and train the deep neural network for defect detection using the Micro-LED defect dataset to obtain a trained detection model.

[0009] The defect detection deep neural network includes a backbone network for multi-level feature extraction, a neck network for cross-scale context fusion, and a prediction module for outputting detection results.

[0010] The shallow feature paths of the backbone network are embedded with multi-scale edge information enhancement modules, and the high-level feature paths are embedded with channel-spatial attention modules. The neck network adopts a hypergraph modulation feature pyramid network.

[0011] S3. Input the Micro-LED image to be detected into the trained detection model to obtain preliminary detection results, perform non-maximum suppression processing on the preliminary detection results, and output the category and location information of Micro-LED defects.

[0012] In one embodiment, step S1 includes the following steps:

[0013] S11. Acquire Micro-LED sample images through an automatic scanning system, manually annotate the defect targets in the sample images, generate a label file containing defect category and bounding box information, and construct an initial defect dataset;

[0014] S12. Perform data augmentation processing on the initial defect dataset and update the corresponding labels synchronously. Divide the augmented dataset into a training set, a validation set, and a test set.

[0015] In one embodiment, the processing steps of the multi-scale edge information enhancement module in step S2 are as follows:

[0016] Input feature map X, perform channel compression through 1×1 convolution to obtain initial features. ;

[0017] The initial features The input is fed into the main feature extraction branch for cascaded convolution, extracting local semantic features layer by layer to obtain intermediate features. Output features of the main feature extraction branch ;

[0018] Extract the intermediate features of the branches from the main features. The input is fed into the multi-scale edge enhancement branch, and after multi-scale pooling and edge enhancement processing, the fused multi-scale edge enhancement features are obtained. ;

[0019] Initial features The output features of the main feature extraction branch With the multi-scale edge enhancement features Perform fusion to obtain features , is represented as:

[0020]

[0021] Then, it is concatenated with the feature map X, and channel remapping is completed through 1×1 convolution to obtain the output of the multi-scale edge information enhancement module. , is represented as:

[0022] .

[0023] In one embodiment, the processing steps for the multi-scale edge enhancement branch are as follows:

[0024] Multiple pooling branches with different scales are set up to process the intermediate features of the input. Pooling, convolution mapping, and upsampling are performed at the corresponding scales to obtain reconstructed features at each scale. , is represented as:

[0025]

[0026] in, This is a pre-defined set of multiple pooling scales;

[0027] Reconstruction features at various scales Perform edge enhancement operations by reconstructing features The high-frequency edge components are extracted from the difference between the pooling features and the edge enhancement features, which are then weighted by activation functions and convolutions to obtain edge enhancement features at various scales. , is represented as:

[0028]

[0029] in, Use the Sigmoid activation function;

[0030] Edge enhancement features at various scales The multi-scale edge enhancement branch is obtained by concatenating it with local branch features and then compressing it through a 1×1 convolution channel. , is represented as:

[0031]

[0032] in, Indicates intermediate features of the input Local features after convolution.

[0033] In one embodiment, the processing steps of the channel-spatial attention module in step S2 are as follows:

[0034] Input features It is divided into two paths. One path is transmitted directly as the original feature-preserving branch, and its output is denoted as... Another input path is used for feature enhancement by an adaptive sparse self-attention enhancement branch, and the output of the adaptive sparse self-attention enhancement branch... Represented as:

[0035]

[0036] The original features are retained as the output features of the branch. With the output of the adaptive sparse self-attention enhancement branch The data is concatenated and then subjected to a 1×1 convolution to complete channel remapping, thus obtaining the output of the channel-spatial attention module. , is represented as:

[0037] .

[0038] In one embodiment, the processing steps of the adaptive sparse self-attention enhancement branch are as follows:

[0039] Input features Divide the data into non-overlapping windows and obtain the query matrix through linear mapping. Key matrix AND-value matrix , respectively represented as:

[0040]

[0041] Spatial correlation is modeled using dense and sparse self-attention branches, respectively. The dense self-attention branch uses the Softmax function to calculate global attention weights, while the sparse self-attention branch uses the squared ReLU function to suppress the attention response in weakly correlated regions. These are expressed as follows:

[0042]

[0043] Where d represents the feature dimension of the query matrix and the key matrix;

[0044] By dynamically adjusting the contribution ratio of dense and sparse self-attention using learnable weights and introducing residual connections, a fused attention output is obtained. , is represented as:

[0045]

[0046] The fusion attention output Channel remapping is performed using consecutive 1×1 convolutions to obtain the output of the adaptive sparse self-attention enhancement branch. , is represented as:

[0047] .

[0048] In one embodiment, the processing steps of the hypergraph modulation feature pyramid network in step S2 are specifically as follows:

[0049] The semantic collection module spatially aligns the multi-scale features output by the backbone network to unify the feature interaction space.

[0050] The modulation fusion module adaptively calculates the fusion weights for the aligned multi-scale features and performs weighted fusion.

[0051] A hypergraph association matrix is ​​constructed through the hypergraph computation module, and higher-order relationship modeling is performed through hypergraph convolution to complete the evolution of higher-order features;

[0052] By fusing local details and global semantics through a hybrid aggregation feature module, a multi-scale enhanced feature map is reconstructed.

[0053] In one embodiment, the processing steps of the modulation fusion module are specifically as follows:

[0054] Multi-source input features are spatially converged, and modulation weights are dynamically calculated using global average pooling, a multilayer perceptron, and a Softmax function to perform weighted fusion. The output of the modulation fusion module is... Represented as:

[0055]

[0056] GAP stands for Global Average Pooling, and MLP stands for Multilayer Perceptron.

[0057] In one embodiment, the processing steps of the hypergraph computation module are as follows:

[0058] The input features are mapped to node tensors, and a hypergraph association matrix is ​​constructed using the distances between nodes to model long-distance nonlocal dependencies. The hypergraph association matrix is ​​represented as follows:

[0059]

[0060] in, and This represents the i-th and j-th semantic nodes in the expanded node tensor. Indicates the learnable association threshold;

[0061] Hypergraph convolution aggregates node information to hyperedges and redistributes it to vertices, obtaining evolved feature representations. The output of the hypergraph computation module is... for:

[0062]

[0063] in, The feature tensor is evolved by hypergraph convolution, BN is the batch normalization operation, and Reshape is the tensor reshaping operation.

[0064] In one embodiment, the processing steps of the hybrid aggregation feature module are as follows:

[0065] Multiple parallel convolutional branches are set up, including 1×1 convolutional branches, depthwise separable convolutional branches, pointwise convolutional branches, and cascaded convolutional neck branches, which jointly model local details and global semantics;

[0066] The output of the cascaded convolution neck Represented as:

[0067]

[0068] The features from each parallel branch are aggregated along the channel dimension and then output uniformly through a convolutional layer to obtain the output of the hybrid aggregated feature module. , is represented as:

[0069] .

[0070] The aforementioned Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion has the following beneficial effects:

[0071] Firstly, by embedding a multi-scale edge information enhancement module in the shallow layer of the backbone network, the edge texture and fine-grained feature expression of small defects can be effectively enhanced, alleviating the problem of small target feature loss during network downsampling, significantly improving the detection capability of small-sized and weak-contrast defects, and reducing the false negative rate.

[0072] Secondly, by embedding a channel-spatial attention module at a high level of the backbone network, redundant feature responses of periodic backgrounds can be adaptively suppressed, the effective defect area can be accurately focused, the interference of background noise on defect identification can be reduced, and the false detection rate can be effectively reduced.

[0073] Third, by using a hypergraph modulation feature pyramid network as the neck structure, high-order semantic fusion of cross-scale features can be achieved, long-distance non-local dependencies can be modeled, and the full interaction between shallow detail information and deep semantic information can be promoted, thereby comprehensively improving the overall detection accuracy and robustness of multi-type and multi-scale defects. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating this embodiment;

[0075] Figure 2 This is a schematic diagram of the multi-scale edge information enhancement module in this embodiment;

[0076] Figure 3 This is a schematic diagram of the channel-space attention module in this embodiment;

[0077] Figure 4 This is a schematic diagram of the structure of the hypergraph modulation feature pyramid network in this embodiment;

[0078] Figure 5 This is a flowchart illustrating the end-to-end defect detection deep neural network in this embodiment. Detailed Implementation

[0079] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0080] A method for detecting defects in Micro-LEDs based on multi-scale feature enhancement and hypergraph fusion, such as... Figure 1 As shown, it includes the following steps:

[0081] S1. Obtain Micro-LED defect dataset;

[0082] S2. Construct an end-to-end deep neural network for defect detection, and train the deep neural network for defect detection using the Micro-LED defect dataset to obtain a trained detection model.

[0083] The defect detection deep neural network includes a backbone network for multi-level feature extraction, a neck network for cross-scale context fusion, and a prediction module for outputting detection results.

[0084] The shallow feature paths of the backbone network are embedded with multi-scale edge information enhancement modules, the high-level feature paths are embedded with channel-spatial attention modules, and the neck network adopts a hypergraph modulated feature pyramid network.

[0085] S3. Input the Micro-LED image to be detected into the trained detection model to obtain preliminary detection results, perform non-maximum suppression processing on the preliminary detection results, and output the category and location information of Micro-LED defects.

[0086] Specifically, S1 includes the following steps:

[0087] S11. Images of Micro-LED samples from different batches and processes are acquired using an industrial automated scanning microscope system, with image resolution adapted to the camera parameters of the inspection station. Professional quality inspectors annotate defects in the images pixel by pixel, generating label files containing defect category labels and bounding box coordinates to construct an initial defect dataset.

[0088] In this embodiment, the automatic scanning system includes a microscopic imaging unit, a motion control unit, and a detection unit. The microscopic imaging unit, composed of a telecentric lens, a white light source, and a CMOS camera, is used to acquire high-resolution microscopic images of the Micro-LED sample. The motion control unit, composed of a moving platform and a motion controller, is used to control the precise movement of the sample in the X and Y directions, enabling region-by-region scanning and acquisition of the entire sample. The detection unit, composed of a computer, is used to complete image reception, parameter communication, model inference, and detection result output. This system is used to implement the automatic acquisition and detection process of Micro-LED images, and is not limited to any specific equipment model.

[0089] Meanwhile, the defect categories include normal, empty wafer, black spot, offset, deflection, breakage, foreign matter and crystal bridging, covering common defect types in Micro-LED production.

[0090] S12. Perform data augmentation on the initial defect dataset, synchronously updating the coordinates and category information of the corresponding labels to expand the data volume and improve the model's generalization ability. The data augmentation methods used in this embodiment include random scaling, random cropping, random rotation, brightness perturbation, and stitching enhancement.

[0091] The augmented dataset is divided into training, validation, and test sets, which are used for network parameter training, hyperparameter tuning, and final performance evaluation, respectively.

[0092] Specifically, such as Figure 5 As shown, in step S2, an end-to-end deep neural network for defect detection is constructed, which consists of three interconnected parts: a backbone network, a neck network, and a prediction module.

[0093] The backbone network employs a multi-level convolutional downsampling structure to extract multi-scale initial features from the input image; a multi-scale edge information enhancement module is embedded in the shallow feature path of the backbone network to enhance the edge texture and fine-grained features of minor defects; a channel-spatial attention module is embedded in the high-level feature path of the backbone network to suppress periodic background redundancy features and enhance the feature response of defective regions.

[0094] The neck network employs a hypergraph modulation feature pyramid network, which receives multi-scale features output from the backbone network, completes cross-scale semantic alignment, dynamic weighted fusion, and high-order feature evolution, and outputs an enhanced multi-scale feature map.

[0095] The prediction module uses a multi-scale detection head structure to classify and predict the location of the enhanced feature map, and outputs the class probability, confidence score and bounding box coordinates of each candidate box.

[0096] Specifically, such as Figure 2 As shown, the multi-scale edge information enhancement module is embedded in the shallow feature layer of the backbone network. The input is the feature map X of the corresponding layer of the backbone network, and the output is the feature map after edge enhancement. It adopts a dual-branch parallel architecture internally. The specific processing steps are as follows:

[0097] Input feature map X, perform channel compression through 1×1 convolution to obtain initial features. ;

[0098] Initial features The input is fed into the main feature extraction branch, and local semantic features are extracted layer by layer through two levels of 3×3 cascaded convolutions. The first-level convolution outputs intermediate features. The second-level convolution outputs the main feature extraction branch's output features. ;

[0099] intermediate features The input is fed into the multi-scale edge enhancement branch, and after multi-scale pooling and edge enhancement processing, the fused multi-scale edge enhancement features are obtained. ;

[0100] Initial features Output features of the main feature extraction branch Multi-scale edge enhancement features Element-wise summation is performed, followed by 3×3 convolution for feature fusion to obtain the features. , is represented as:

[0101]

[0102] Features The image is concatenated with the feature map X, and then subjected to a 1×1 convolution to complete channel remapping, thus obtaining the output of the multi-scale edge information enhancement module. , is represented as:

[0103] .

[0104] The processing steps for multi-scale edge enhancement branches are as follows:

[0105] Four groups of average pooling branches with different scales are set up, corresponding to the pooling scale set scale. ={3,6,9,12}. Intermediate features of the input. Pooling, 1×1 convolution mapping, and 3×3 convolution feature extraction and upsampling operations are performed at the corresponding scales to obtain reconstructed features at each scale with the same input size. , is represented as:

[0106]

[0107] Reconstruction features at various scales Perform edge enhancement operations by reconstructing features The high-frequency edge components are extracted from the difference between the pooling features and the high-frequency edge components, and then activated by the Sigmoid function. After being weighted by a 1×1 convolution, the reconstructed features are superimposed to obtain edge enhancement features at each scale. , is represented as:

[0108]

[0109] Finally, edge enhancement features at each scale are applied. - Local branching features (Intermediate features) After being concatenated along the channel dimension by 1×1 convolution, and then compressed by 1×1 convolution channel, the output of the multi-scale edge enhancement branch is obtained. , is represented as:

[0110] .

[0111] Specifically, such as Figure 3 As shown, the channel-spatial attention module is embedded in the high-level feature layer of the backbone network, and its input is the high-level features. The output is the attention-enhanced feature map. A dual-path parallel architecture of original feature preservation and attention enhancement is used. The specific processing steps are as follows:

[0112] Input features It is divided into two paths. One path serves as the original feature preservation branch, directly transmitting the features. Its output is denoted as... The other input branch, an adaptive sparse self-attention enhancement branch, adaptively enhances the features in terms of spatial and channel dimensions, and its output is denoted as... ;

[0113] The adaptive sparse self-attention enhancement branch consists of n cascaded ASSA units. The input features are first adjusted for the number of channels by a 1×1 convolution, and then passed through each ASSA unit sequentially. The overall output is:

[0114]

[0115] In this embodiment, the number of cascaded layers is 2.

[0116] Preserve the original features from the branch's output features. Output of the adaptive sparse self-attention enhancement branch The concatenation is performed, and channel remapping is completed through 1×1 convolution to obtain the output of the channel-spatial attention module. , is represented as:

[0117] .

[0118] The processing steps for a single-level ASSA unit are as follows:

[0119] Input features Divide the data into non-overlapping windows and obtain the query matrix through linear mapping. Key matrix AND-value matrix , respectively represented as:

[0120]

[0121] In the formula, , , It is a learnable linear mapping weight matrix.

[0122] Spatial correlation is modeled using both dense and sparse self-attention branches. The dense self-attention branch uses the Softmax function to calculate global attention weights and models the spatial correlation across the entire window. The sparse self-attention branch uses the squared ReLU function to suppress the attention response in weakly correlated regions, allowing information from the effective defect region to propagate forward. These are expressed as follows:

[0123]

[0124] Where d represents the feature dimension of the query matrix and the key matrix;

[0125] Learnable weight coefficients , Dynamically fuse the two attention results and introduce a residual connection to obtain the fused attention output. , is represented as:

[0126]

[0127] Fusion attention output The output of a single-level ASSA unit is obtained by performing channel remapping after two layers of 1×1 convolution. , is represented as:

[0128] .

[0129] The channel-space attention module can enhance the effective feature response of defective target areas and suppress interference information caused by large areas of normal background.

[0130] Specifically, such as Figure 4 As shown, the hypergraph modulation feature pyramid network, acting as the neck network, receives multi-scale features output from the backbone network. Feature enhancement is then achieved sequentially through a semantic collection module, a modulation fusion module, a hypergraph computation module, and a hybrid aggregation feature module. The specific processing steps are as follows:

[0131] The semantic collection module performs spatial size alignment on the multi-scale features P1, P2, P3, and P4 output by the backbone network, enabling feature maps of different resolutions to interact in a unified semantic space, thus preparing for subsequent fusion.

[0132] The modulation fusion module adaptively calculates fusion weights for the aligned multi-scale features and performs weighted fusion.

[0133] This involves spatially pooling multi-source input features, dynamically calculating modulation weights using Global Average Pooling (GAP), Multilayer Perceptron (MLP), and Softmax functions, and then performing weighted fusion. The output of the modulation fusion module... Represented as:

[0134]

[0135] In the formula, n represents the total number of multi-scale features input. is the input feature at the j-th scale.

[0136] The hypergraph computation module constructs a hypergraph association matrix using the distance between nodes and performs high-order relationship modeling through hypergraph convolution to complete the evolution of high-order features.

[0137] The merged feature maps are represented as node tensors. A hypergraph association matrix is ​​constructed using the distances between nodes to model long-range nonlocal dependencies. The hypergraph association matrix is ​​expressed as:

[0138]

[0139] In the formula, and This represents the i-th and j-th semantic nodes in the expanded node tensor. Indicates the learnable association threshold. This is the function for calculating feature distance;

[0140] Hypergraph convolution is performed based on the hypergraph association matrix, aggregating node information to hyperedges and redistributing it to vertices to obtain the evolved feature tensor. After reshaping the evolutionary features to their original size, a residual connection is made with the input features. This connection is then processed by batch normalization (BN) and the SiLU activation function to obtain the output of the hypergraph computation module. , is represented as:

[0141]

[0142] In the formula, Reshape is the tensor reshaping operation.

[0143] By using a hybrid aggregation feature module, local details and global semantics are fused to reconstruct a multi-scale enhanced feature map;

[0144] This involves setting up multiple parallel convolutional branches, including 1×1 convolutional branches, depthwise separable convolutional branches, pointwise convolutional branches, and cascaded convolutional neck branches, to jointly model local details and global semantics.

[0145] The cascaded residual convolution neck branch employs a multi-layered cascaded residual convolution structure, with the output of the i-th layer being:

[0146]

[0147] The features from each parallel branch are aggregated along the channel dimension, and the channel information is integrated through a 3×3 convolution to obtain the output of the hybrid aggregated feature module. , is represented as:

[0148] .

[0149] Through the above embodiments, the present invention can accurately identify and locate tiny, dense, low-discrimination, and multi-scale defects in Micro-LED images, and realize automated and real-time Micro-LED defect detection.

[0150] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for detecting defects in Micro-LEDs based on multi-scale feature enhancement and hypergraph fusion, characterized in that, Includes the following steps: S1. Obtain Micro-LED defect dataset; S2. Construct an end-to-end deep neural network for defect detection, and train the deep neural network for defect detection using the Micro-LED defect dataset to obtain a trained detection model. The defect detection deep neural network includes a backbone network for multi-level feature extraction, a neck network for cross-scale context fusion, and a prediction module for outputting detection results. The shallow feature paths of the backbone network are embedded with multi-scale edge information enhancement modules, and the high-level feature paths are embedded with channel-spatial attention modules. The neck network adopts a hypergraph modulation feature pyramid network. S3. Input the Micro-LED image to be detected into the trained detection model to obtain preliminary detection results, perform non-maximum suppression processing on the preliminary detection results, and output the category and location information of Micro-LED defects.

2. The Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion according to claim 1, characterized in that, Step S1 includes the following steps: S11. Acquire Micro-LED sample images through an automatic scanning system, manually annotate the defect targets in the sample images, generate a label file containing defect category and bounding box information, and construct an initial defect dataset; S12. Perform data augmentation processing on the initial defect dataset and update the corresponding labels synchronously. Divide the augmented dataset into a training set, a validation set, and a test set.

3. The Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion according to claim 1, characterized in that, In step S2, the processing steps of the multi-scale edge information enhancement module are as follows: Input feature map X, perform channel compression through 1×1 convolution to obtain initial features. ; The initial features The input is fed into the main feature extraction branch for cascaded convolution, extracting local semantic features layer by layer to obtain intermediate features. Output features of the main feature extraction branch ; Extract the intermediate features of the branches from the main features. The input is fed into the multi-scale edge enhancement branch, and after multi-scale pooling and edge enhancement processing, the fused multi-scale edge enhancement features are obtained. ; Initial features The output features of the main feature extraction branch With the multi-scale edge enhancement features Perform fusion to obtain features , is represented as: Then, it is concatenated with the feature map X, and channel remapping is completed through 1×1 convolution to obtain the output of the multi-scale edge information enhancement module. , is represented as: 。 4. The Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion according to claim 3, characterized in that, The processing steps for the multi-scale edge enhancement branches are as follows: Multiple pooling branches with different scales are set up to process the intermediate features of the input. Pooling, convolution mapping, and upsampling are performed at the corresponding scales to obtain reconstructed features at each scale. , is represented as: in, This is a pre-defined set of multiple pooling scales; Reconstruction features at various scales Perform edge enhancement operations by reconstructing features The high-frequency edge components are extracted from the difference between the pooling features and the edge enhancement features, which are then weighted by activation functions and convolutions to obtain edge enhancement features at various scales. , is represented as: in, Use the Sigmoid activation function; Edge enhancement features at various scales The multi-scale edge enhancement branch is obtained by concatenating it with local branch features and then compressing it through a 1×1 convolution channel. , is represented as: in, Indicates intermediate features of the input Local features after convolution.

5. The Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion according to claim 1, characterized in that, In step S2, the processing steps of the channel-spatial attention module are as follows: Input features It is divided into two paths. One path is transmitted directly as the original feature-preserving branch, and its output is denoted as... Another input path is used for feature enhancement by an adaptive sparse self-attention enhancement branch, and the output of the adaptive sparse self-attention enhancement branch... Represented as: The original features are retained as the output features of the branch. With the output of the adaptive sparse self-attention enhancement branch The data is concatenated and then subjected to a 1×1 convolution to complete channel remapping, thus obtaining the output of the channel-spatial attention module. , is represented as: 。 6. The Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion according to claim 5, characterized in that, The processing steps for the adaptive sparse self-attention enhancement branch are as follows: Input features Divide the data into non-overlapping windows and obtain the query matrix through linear mapping. Key matrix AND-value matrix , respectively represented as: Spatial correlation is modeled using dense and sparse self-attention branches, respectively. The dense self-attention branch uses the Softmax function to calculate global attention weights, while the sparse self-attention branch uses the squared ReLU function to suppress the attention response in weakly correlated regions. These are expressed as follows: Where d represents the feature dimension of the query matrix and the key matrix; By dynamically adjusting the contribution ratio of dense and sparse self-attention using learnable weights and introducing residual connections, a fused attention output is obtained. , is represented as: The fusion attention output Channel remapping is performed using consecutive 1×1 convolutions to obtain the output of the adaptive sparse self-attention enhancement branch. , is represented as: 。 7. The Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion according to claim 1, characterized in that: In step S2, the processing steps of the hypergraph modulation feature pyramid network are as follows: The semantic collection module spatially aligns the multi-scale features output by the backbone network to unify the feature interaction space. The modulation fusion module adaptively calculates the fusion weights for the aligned multi-scale features and performs weighted fusion. A hypergraph association matrix is ​​constructed through the hypergraph computation module, and higher-order relationship modeling is performed through hypergraph convolution to complete the evolution of higher-order features; By fusing local details and global semantics through a hybrid aggregation feature module, a multi-scale enhanced feature map is reconstructed.

8. The Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion according to claim 7, characterized in that, The specific processing steps of the modulation fusion module are as follows: Multi-source input features are spatially converged, and modulation weights are dynamically calculated using global average pooling, a multilayer perceptron, and a Softmax function to perform weighted fusion. The output of the modulation fusion module is... Represented as: GAP stands for Global Average Pooling, and MLP stands for Multilayer Perceptron.

9. The Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion according to claim 7, characterized in that, The specific processing steps of the hypergraph calculation module are as follows: The input features are mapped to node tensors, and a hypergraph association matrix is ​​constructed using the distances between nodes to model long-range nonlocal dependencies. The hypergraph association matrix is ​​represented as follows: in, and This represents the i-th and j-th semantic nodes in the expanded node tensor. Indicates the learnable association threshold; Hypergraph convolution aggregates node information to hyperedges and redistributes it to vertices, obtaining evolved feature representations. The output of the hypergraph computation module is... for: in, The feature tensor is evolved by hypergraph convolution, BN is the batch normalization operation, and Reshape is the tensor reshaping operation.

10. The Micro-LED defect detection method based on multi-scale feature enhancement and hypergraph fusion according to claim 7, characterized in that, The specific processing steps of the hybrid aggregation feature module are as follows: Multiple parallel convolutional branches are set up, including 1×1 convolutional branches, depthwise separable convolutional branches, pointwise convolutional branches, and cascaded convolutional neck branches, which jointly model local details and global semantics; The output of the cascaded convolution neck Represented as: The features from each parallel branch are aggregated along the channel dimension and then output uniformly through a convolutional layer to obtain the output of the hybrid aggregated feature module. , is represented as: 。

Images