Photovoltaic module defect detection method based on multi-mode intermediate frequency fusion and high-frequency perception enhancement

Abstract

The invention discloses a photovoltaic module defect detection method based on multi-mode intermediate frequency fusion and high-frequency perception enhancement, and the method comprises the steps: building a multi-mode image, an EL image and a depth image of a photovoltaic module defect, and constructing a YOLOMM-Mid-HFP model; a double-path feature extraction stage: performing double-path parallel processing on the visible light image and the auxiliary modal image, and respectively generating double-path P3, P4 and P5 multi-scale feature maps; the dual-path P4 layer and the dual-path P5 layer are respectively fused and then are enhanced through high-frequency perception; defect detection and detection result output are carried out through a multi-scale feature pyramid stage, the constructed model is trained, and photovoltaic module defect detection is carried out through the trained model. Compared with the prior art, the method has the advantages that the detection capability of the model is enhanced by combining the multi-mode intermediate frequency fusion and the high-frequency sensing model, and the performance of detecting the defects of the photovoltaic module is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a method for detecting defects in photovoltaic modules based on multimodal mid-frequency fusion and high-frequency perception enhancement. Background Technology

[0002] With the rapid development of the photovoltaic industry, the quality testing of photovoltaic modules has become a key link in ensuring power generation efficiency and safe operation.

[0003] Traditional detection methods have the following problems:

[0004] 1) Single-mode detection is difficult to fully capture surface and internal defect features;

[0005] 2) High-frequency detail information is easily lost in deep networks, resulting in poor small target detection performance;

[0006] 3) Existing fusion strategies have limitations in feature alignment and detail preservation.

[0007] For multimodal photovoltaic module detection methods: Early fusion (e.g., Concatenation at Input) presents the challenge of significant intermodal differences and difficulty in feature alignment, potentially leading to severe information interference and limited detection accuracy. This model, however, reduces modal interference through independent backbone network preprocessing. Late fusion (e.g., Decision-level Fusion) suffers from poor information complementarity and significant detail loss, potentially resulting in low detection rates for microcracks and other minor defects. This model, however, performs deep fusion at the feature level, preserving crucial details. Summary of the Invention

[0008] Purpose of the invention: In view of the problems pointed out in the background art, the present invention discloses a photovoltaic module defect detection method based on multimodal mid-frequency fusion and high-frequency sensing enhancement. A mid-frequency feature fusion layer and a high-frequency sensing enhancement module (HFP) are set in both the P4 layer and the P5 layer to perform dual-path feature extraction and fusion of multimodal images, which effectively improves the performance of photovoltaic module defect detection.

[0009] Technical Solution: This invention discloses a method for detecting defects in photovoltaic modules based on multimodal mid-frequency fusion and high-frequency sensing enhancement, comprising the following steps:

[0010] Step 1: Obtain multimodal images of photovoltaic module defects, including EL images and depth images, to form a dataset;

[0011] Step 2: Construct the YOLOMM-Mid-HFP detection model, which is as follows:

[0012] 1) Dual-path feature extraction stage: EL path: processes visible light images and extracts surface defect features; Depth path: processes auxiliary modal images and extracts internal defect features. The two paths are processed in parallel to generate P3, P4, and P5 multi-scale feature maps respectively.

[0013] 2) In the mid-frequency feature fusion stage, P4 layer fusion: the EL_P4 feature map obtained by the EL path and the Depth_P4 feature map obtained by the Depth path are concatenated; P5 layer fusion: the EL_P5 feature map obtained by the EL path and the Depth_P5 feature map obtained by the Depth path are concatenated; after fusion, feature optimization and dimensionality reduction are performed.

[0014] 3) High-frequency perception enhancement stage: Feature enhancement is performed on the fused features of P4 layer and P5 layer respectively;

[0015] 4) Defect detection is achieved through a multi-scale feature pyramid stage, and the detection results are output;

[0016] Step 3: Use the acquired dataset to train the constructed YOLOMM-Mid-HFP detection model, and use the trained detection model to detect defects in photovoltaic modules in the images to be detected.

[0017] Furthermore, the EL path and Depth path in the dual-path feature extraction stage have the same structure but independent parameters, and are both implemented based on the YOLO model backbone network:

[0018] The EL path is layer 0 to layer 10, and the Depth path is layer 11 to layer 21.

[0019] The P3 feature generation for the EL path consists of layers 0 to 4, including convolutional layers and the C3k2 module; the P4 feature generation consists of layers 5 to 6, including convolutional layers and the C3k2 module, with P3 features as input; the P5 feature generation consists of layers 7 to 10, including convolutional layers, the C3k2 module, the SPPF module, and the C2PSA module, with P4 features as input, and so on for the P3, P4, and P5 features of the Depth path.

[0020] Furthermore, the mid-frequency feature fusion stage and the high-frequency perception enhancement stage are performed sequentially at the P4 and P5 feature levels, with cross-modal feature fusion and subsequent enhancement occurring in turn.

[0021] The features of the P4 layer of the EL path and the Depth path are concatenated along the channel dimension and optimized and reduced in dimensionality by a C3k2 convolution module. The fused features are then fed into the High Frequency Perception Enhancement (HFP) module. The features of the P5 layer of the EL path and the Depth path are concatenated, first processed by a C2PSA module, and then optimized by a C3k2 module. Similarly, the fused P5 features are enhanced by the High Frequency Perception Enhancement (HFP) module.

[0022] Furthermore, the high-frequency sensing enhancement module (HFP) enhances defect features through frequency domain processing and attention mechanisms, as follows:

[0023] The input features are transformed to the frequency domain using a Fast Fourier Transform.

[0024] Low-frequency components are masked by a binary mask, while high-frequency components are preserved, thus preserving edge texture features.

[0025] The frequency domain features after masking are transformed back to the spatial domain to obtain the high-frequency features.

[0026] A spatial weight map is generated based on high-frequency features, and the weights are applied to the original input features to highlight defective areas;

[0027] Calculate channel importance and optimize feature combinations;

[0028] The spatial and channel attention results are weighted and fused.

[0029] Furthermore, a binary mask is used to block low-frequency components while retaining high-frequency components. The mask is defined as follows:

[0030] ;

[0031] Where H and W are the height and width of the frequency domain feature map, respectively, and h and w are the spatial location indices, r h and r w This is a proportional hyperparameter used to control the size of the low-frequency region.

[0032] Furthermore, the multi-scale feature pyramid stage adopts a structure combining feature pyramids and path aggregation networks, specifically:

[0033] The enhanced P5 features are upsampled and then concatenated and fused with the enhanced P4 features;

[0034] The obtained FPN_P4 features are upsampled again and then concatenated and fused with the original P3 features output by the backbone network.

[0035] After downsampling the FPN_P3 features, they are concatenated and fused with the FPN_P4 features;

[0036] After downsampling the PAN_P4 features, they are spliced ​​and fused with the original P5 enhanced features from the fusion enhancement layer.

[0037] Furthermore, loss calculation and optimization are performed. The total loss function is a weighted sum of four parts: localization loss, classification loss, target-specific loss, and fusion consistency loss. Backpropagation is then used to calculate the gradient, update the network parameters, and finally, iterative optimization is performed, repeating training until the model converges. The total loss function is:

[0038]

[0039] in, , , , These are the weights of the localization loss, classification loss, targetability loss, and fusion consistency loss, respectively. It is the total loss. It is the localization loss, which calculates the error between the defect bounding box predicted by the model and the actual labeled box; It is the classification loss, which calculates the error between the model's prediction of the defect category and the true category label; It is a targeted loss function used to determine whether a defect exists within a mesh cell; This is the fusion consistency loss, used to constrain the alignment of bimodal features before fusion. It is obtained by calculating the L2 norm between the EL path P4 feature and the Depth path P4 feature after a 1x1 convolution transformation.

[0040]

[0041] in, This represents a 1x1 convolution operation used for channel alignment.

[0042] Beneficial effects:

[0043] 1. Innovative breakthrough in multimodal mid-frequency fusion strategy

[0044] This invention performs mid-frequency fusion at the P4 / P5 feature layers, achieving an optimal balance between shallow detail information and deep semantic information. Both the P4 and P5 layers incorporate mid-frequency feature fusion layers and high-frequency perceptual enhancement modules (HFP). The innovation of this "hierarchical feature enhancement" architecture lies not in designing different enhancement modules for different levels, but in strategically deploying the same meticulously designed fusion and enhancement mechanism across different key nodes of the feature pyramid, thereby achieving targeted optimization of defects at different scales.

[0045] 1) Function of P4 layer (intermediate feature layer): The P4 layer feature map has medium spatial resolution and medium semantic abstraction. Its features focus more on the local regions of components and medium-scale structural patterns. Fusion and enhancement are performed at this layer, and its main goal is to optimize the model's ability to perceive medium-scale defects (such as broken gates and local scratches). The HFP module here focuses on enhancing the mid-to-high frequency details that define the boundaries and textures of such defects.

[0046] 2) Function of P5 layer (high-level features): The P5 layer feature map has a lower spatial resolution but the richest semantic information. Its features correspond to a more macroscopic and holistic visual pattern. Fusion and enhancement are performed at this layer, with the main goal of ensuring that the model's semantic representation of large-scale defects (such as large-area hidden cracks and hot spots) is sufficiently accurate and robust. The role of the HFP module here is to preserve and enhance the key contour and distribution texture information of these large defects that still exist after downsampling, preventing them from being over-smoothed in deep networks.

[0047] Although the modules at the P4 and P5 levels are composed of the same "fusion + HFP", the actual functions and effects produced by these modules at different levels are hierarchical and complementary because the input features they process are fundamentally different in scale and semantic abstraction.

[0048] 2. Frequency Domain Innovation of High-Frequency Sensing Enhancement Mechanism

[0049] This invention organically combines frequency domain signal processing with deep learning to achieve adaptive high-frequency feature enhancement. Compared to traditional attention mechanisms, the SE module only focuses on channel importance, ignoring spatial high-frequency features. HFP combines frequency domain filtering, spatial attention, and channel attention for triple enhancement. CBAM processes spatial and channel attention sequentially, resulting in a single computational path, while HFP processes frequency, spatial, and channel information in parallel, leading to more comprehensive feature enhancement.

[0050] 3. Architectural innovation with enhanced hierarchical features

[0051] High-frequency perception enhancement modules are hierarchically inserted at key feature fusion points to achieve progressive feature optimization. Layer P4 enhancement targets medium-scale defects (broken grids, localized scratches), while layer P5 enhancement targets large-scale defects (large-area hidden cracks, hot spots). This hierarchical optimization targets defects of different scales, avoiding feature confusion. Compared to traditional single-point enhancement, which only adds an attention module to the backbone, resulting in significant loss of high-frequency information in deep networks, this new module enhances features before they degrade, preserving more complete information. Attached Figure Description

[0052] Figure 1 This is a diagram illustrating the overall architecture of the YOLOMM-Mid-HFP detection model of this invention.

[0053] Figure 2 This is a diagram showing the architecture of the YOLOMM-Mid-HFP detection model with dimensions according to an embodiment of the present invention;

[0054] Figure 3 This is a thermodynamic comparison diagram of different models for photovoltaic defect detection in embodiments of the present invention. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0056] This invention discloses a method for defect detection of photovoltaic modules combining a multimodal intermediate frequency fusion network with high-frequency sensing enhancement. Figure 1 and Figure 2 Its core architecture is as follows Figure 2 As shown:

[0057] Module input and overall structure:

[0058] Input multimodal images, EL images (electroluminescent images): And depth images: Where: B is the batch size, the number of channels is 3, and H×W is the spatial size.

[0059] The overall architecture includes: a dual-path backbone network (EL path + Depth path), a mid-frequency feature fusion layer (P4 / P5 fusion point), a high-frequency perception enhancement module HFP (High Frequency Perception), and a multi-scale feature pyramid (FPN+PAN).

[0060] 1. Dual-path backbone network: The backbone network consists of two structurally identical but parameter-independent convolutional neural network paths.

[0061] 1.1 EL Path (Layers 0-10): This path is responsible for processing electroluminescent images. Through layer-by-layer forward propagation, feature maps at three scales, P3, P4, and P5, are output sequentially. This process can be described as follows:

[0062]

[0063]

[0064] P3 Feature Generation: Consists of layers 0 to 4 (including convolutional layers and the C3k2 module), receives the input image, and outputs a feature map. .

[0065]

[0066]

[0067] P4 feature generation: consists of layers 5 to 6 (including convolutional layers and C3k2 modules). =Taking P3 features as input, the output is a feature map. .

[0068]

[0069]

[0070] P5 feature generation: Consists of layers 7 to 10 (including convolutional layers, C3k2 modules, SPPF modules, and C2PSA modules), taking P4 features as input and outputting feature maps. .

[0071] 1.2 Depth Path (Layers 11-21): This path receives depth image input. Its module composition is exactly the same as the EL path, but its parameters are independent. Its calculation process can be described as follows:

[0072]

[0073]

[0074] P3 Feature Generation: Consists of layers 11 to 15 (including convolutional layers and the C3k2 module), receives the input image, and outputs a feature map. .

[0075]

[0076]

[0077] P4 Feature Generation: Consists of layers 16 and 17 (including convolutional layers and the C3k2 module), taking P3 features as input and outputting feature maps. .

[0078]

[0079]

[0080] P5 feature generation: Consists of layers 18 to 21 (including convolutional layers, C3k2 module, SPPF module, and C2PSA module), taking P4 features as input and outputting feature maps. .

[0081] 2. Mid-frequency fusion and high-frequency sensing enhancement

[0082] In this stage, cross-modal feature fusion and subsequent enhancement are performed sequentially at the P4 and P5 feature levels.

[0083] 2.1 P4 Layer Fusion and Enhancement:

[0084] Fusion: The features from layer P4 of the EL path and the Depth path are concatenated along their channel dimensions, and then optimized and reduced in dimensionality using a C3k2 convolutional module. The calculation is as follows:

[0085]

[0086]

[0087] The fused features are fed into the high-frequency sensing enhancement module, which outputs the enhanced P4 features:

[0088]

[0089] 2.2 P5 Layer Fusion and Enhancement:

[0090] Fusion: The features from the EL path and the P5 layer of the Depth path are concatenated. Unlike the P4 layer, this is first processed using a C2PSA module, and then optimized using a C3k2 module. The calculation is as follows:

[0091]

[0092]

[0093] Enhancement: Similarly, the fused P5 features are enhanced using the HFP module:

[0094]

[0095] 3. Technical details of the High Frequency Sensing Enhancement Module (HFP):

[0096] The High Frequency Perception (HFP) module enhances defect features through frequency domain processing and attention mechanisms.

[0097] 3.1 High-frequency extraction in the frequency domain:

[0098] Transform the input features from the spatial domain to the frequency domain:

[0099]

[0100] in, These are the original input features.

[0101] Low-frequency components are masked using a binary mask, while high-frequency components are preserved. This mask is defined as follows:

[0102] ;

[0103] The frequency domain features after masking are transformed back to the spatial domain to obtain the high-frequency features:

[0104]

[0105] 3.2 Spatial Attention Mechanism: A spatial weight map is generated based on high-frequency features to emphasize spatial regions where defects may exist. The calculation process is as follows:

[0106]

[0107] This weight is then applied to the original input features:

[0108]

[0109] 3.3 Channel attention mechanism:

[0110] Channel importance weights are calculated using global pooling and a shared multilayer perceptron. First, max pooling and average pooling features are calculated:

[0111]

[0112]

[0113] Subsequently, a channel attention map is obtained through a weight-shared fully connected layer and an activation function. And it acts on the input features:

[0114]

[0115] 3.4 Feature Fusion Output:

[0116] The outputs of the spatial and channel attention paths are summed, and then processed by convolution and normalization to obtain the final output of the module:

[0117]

[0118] 4. Multi-scale feature pyramid

[0119] The neck network employs a structure combining feature pyramids and path aggregation networks to process multi-scale features from the backbone network and fusion enhancement layers.

[0120] 4.1 Top-down path (FPN):

[0121] The enhanced P5 features are upsampled and then concatenated and fused with the enhanced P4 features:

[0122]

[0123]

[0124] The obtained FPN_P4 features are upsampled again and then concatenated and fused with the original P3 features output by the backbone network:

[0125]

[0126] 4.2 Bottom-up path (PAN):

[0127] After downsampling the FPN_P3 features, they are concatenated and fused with the FPN_P4 features:

[0128]

[0129] After downsampling the PAN_P4 features, they are concatenated and fused with the original P5 enhanced features from the fusion enhancement layer:

[0130]

[0131] 5. Loss Function Design

[0132] The total loss function is a weighted average of four parts: localization loss, classification loss, target-orientation loss, and fusion consistency loss. Its calculation is as follows:

[0133]

[0134] The fusion consistency loss is used to constrain the alignment of bimodal features before fusion, and is obtained by calculating the L2 norm between the EL path P4 feature and the Depth path P4 feature after a 1x1 convolution transformation.

[0135]

[0136] Here, This represents a 1x1 convolution operation for channel alignment:

[0137]

[0138] Experimental results:

[0139] The test performance results are shown in Table 1:

[0140] Table 1

[0141] model Precision Recall mAP@0.5 mAP@0.5:0.95 YOLOv11n 0.869 0.832 0.893 0.577 YOLOv11n-SE 0.879 0.847 0.899 0.574 YOLOv11n-CBAM 0.883 0.849 0.897 0.543 YOLOMM - Early 0.902 0.845 0.903 0.577 YOLOMM-Late (Evening) 0.905 0.868 0.905 0.581 YOLOMM-Mid (Chinese) 0.904 0.871 0.909 0.582 YOLOMM-Mid-HFP 0.909 0.875 0.925 0.596

[0142] In this context, YOLOMM-Early, YOLOMM-Late, and YOLOMM-Mid represent early-stage fusion, late-stage fusion, and mid-frequency fusion, respectively. The YOLOMM-Mid-HFP of this invention represents mid-frequency fusion with the integration of a high-frequency sensing enhancement module (HFP). Table 1 lists YOLOv11n-SE and YOLOv11n-CBAM as enhanced variants of the single-modal baseline model, used to illustrate the necessity of multimodal methods. Specifically:

[0143] YOLOv11n-SE: Based on the baseline unimodal model YOLOv11n (using only EL images), a channel attention module is embedded at the end of its backbone network, aiming to improve model performance by recalibrating the channel feature responses.

[0144] YOLOv11n-CBAM: Also based on the single-modal model YOLOv11n, it embeds a convolutional block attention module, which combines channel and spatial attention in sequence to verify the optimization effect of more complex attention mechanisms on single-modal data.

[0145] Experimental results show that although SE and CBAM attention mechanisms improve the performance of single-modal benchmark models to some extent (e.g., YOLOv11n-SE improves accuracy by 1.0% compared to YOLOv11n), all their key metrics are significantly lower than any strategy employing multimodal fusion (including early fusion, late fusion, and this invention). This strongly demonstrates that for photovoltaic module defect detection tasks, information fusion from single-modal to multimodal is key to achieving a performance leap, and the resulting performance gain far exceeds the effect of feature optimization within the single-modal model. Based on this, this invention further leverages the advantages of multimodal data through innovative mid-frequency fusion and high-frequency sensing enhancement, thereby achieving optimal detection performance.

[0146] The proposed YOLOMM-Mid-HFP detection model achieves optimal performance across all key metrics: Precision: 0.909, a 4.0% improvement over the benchmark YOLOv11n and a 0.4-0.7% improvement over other multimodal fusion strategies. Recall: 0.875, a significant 4.3% improvement over the benchmark YOLOv11n, indicating the model can detect more real defects. Mean Accuracy (mAP@0.5): an excellent level of 0.925, a 3.2% improvement over the benchmark and a 1.6% improvement over ordinary fusion. Overall Mean Accuracy (mAP@0.5:0.95): 0.596, a 1.9% improvement over the benchmark, maintaining its leading position even under more stringent evaluation criteria.

[0147] This invention was tested on specific photovoltaic modules. Thermodynamic comparison diagrams of different models for photovoltaic defect detection are shown below. Figure 3 Experimental results show that the YOLOMM-Mid-HFP network model proposed in this invention achieves a comprehensive performance breakthrough in photovoltaic module defect detection tasks through the organic combination of multimodal mid-frequency fusion and high-frequency sensing enhancement. Compared with the existing best solution, it achieves significant improvements in key indicators such as precision, recall, and average accuracy, providing an effective technical solution for quality inspection in the photovoltaic industry.

[0148] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting defects in photovoltaic modules based on multimodal mid-frequency fusion and high-frequency sensing enhancement, characterized in that, Includes the following steps: Step 1: Obtain multimodal images of photovoltaic module defects, including EL images and depth images, to form a dataset; Step 2: Construct the YOLOMM-Mid-HFP detection model, which is as follows: 1) Dual-path feature extraction stage: EL path: processes visible light images and extracts surface defect features; Depth path: processes auxiliary modal images and extracts internal defect features. The two paths are processed in parallel to generate P3, P4, and P5 multi-scale feature maps respectively. 2) In the mid-frequency feature fusion stage, P4 layer fusion: the EL_P4 feature map obtained by the EL path and the Depth_P4 feature map obtained by the Depth path are concatenated; P5 layer fusion: the EL_P5 feature map obtained by the EL path and the Depth_P5 feature map obtained by the Depth path are concatenated; after fusion, feature optimization and dimensionality reduction are performed. 3) High-frequency perception enhancement stage: Feature enhancement is performed on the fusion features of P4 layer and P5 layer respectively; 4) Defect detection is achieved through a multi-scale feature pyramid stage, and the detection results are output; Step 3: Use the acquired dataset to train the constructed YOLOMM-Mid-HFP detection model, and use the trained detection model to detect defects in photovoltaic modules in the images to be detected.

2. The photovoltaic module defect detection method based on multimodal mid-frequency fusion and high-frequency sensing enhancement according to claim 1, characterized in that, The EL path and Depth path in the dual-path feature extraction stage have the same structure but independent parameters, and are both implemented based on the YOLO model backbone network: The EL path is layer 0 to layer 10, and the Depth path is layer 11 to layer 21. P3 feature generation for the EL path: consists of layers 0 to 4, including convolutional layers and C3k2 modules; P4 feature generation: consists of layers 5 to 6, including convolutional layers and C3k2 modules, with P3 features as input; P5 feature generation: It consists of layers 7 to 10, including convolutional layers, C3k2 module, SPPF module and C2PSA module. It takes P4 features as input, and P3 features, P4 features and P5 features of the Depth path are generated in sequence.

3. The photovoltaic module defect detection method based on multimodal mid-frequency fusion and high-frequency sensing enhancement according to claim 1, characterized in that, The mid-frequency feature fusion stage and the high-frequency perception enhancement stage are performed sequentially at the P4 and P5 feature levels, with cross-modal feature fusion and subsequent enhancement occurring in turn. The features of the P4 layer of the EL path and the Depth path are concatenated along the channel dimension and optimized and reduced in dimensionality by a C3k2 convolution module. The fused features are then fed into the High Frequency Perception Enhancement (HFP) module. The features of the P5 layer of the EL path and the Depth path are concatenated, first processed by a C2PSA module, and then optimized by a C3k2 module. Similarly, the fused P5 features are enhanced by the High Frequency Perception Enhancement (HFP) module.

4. The photovoltaic module defect detection method based on multimodal mid-frequency fusion and high-frequency sensing enhancement according to claim 3, characterized in that, The High Frequency Perception Enhancement (HFP) module enhances defect features through frequency domain processing and attention mechanisms, as detailed below: The input features are transformed to the frequency domain using a Fast Fourier Transform. Low-frequency components are masked by a binary mask, while high-frequency components are preserved, thus preserving edge texture features. The frequency domain features after masking are transformed back to the spatial domain to obtain the high-frequency features. A spatial weight map is generated based on high-frequency features, and the weights are applied to the original input features to highlight defective areas; Calculate channel importance and optimize feature combinations; The spatial and channel attention results are weighted and fused.

5. The photovoltaic module defect detection method based on multimodal mid-frequency fusion and high-frequency sensing enhancement according to claim 4, characterized in that, Low-frequency components are masked using a binary mask, while high-frequency components are preserved. The mask is defined as follows: ； Where H and W are the height and width of the frequency domain feature map, respectively, and h and w are the spatial location indices, r h and r w This is a proportional hyperparameter used to control the size of the low-frequency region.

6. The photovoltaic module defect detection method based on multimodal mid-frequency fusion and high-frequency sensing enhancement according to claim 1, characterized in that, The multi-scale feature pyramid stage adopts a structure combining feature pyramids and path aggregation networks, specifically: The enhanced P5 features are upsampled and then concatenated and fused with the enhanced P4 features; The obtained FPN_P4 features are upsampled again and then concatenated and fused with the original P3 features output by the backbone network. After downsampling the FPN_P3 features, they are concatenated and fused with the FPN_P4 features; After downsampling the PAN_P4 features, they are spliced ​​and fused with the original P5 enhanced features from the fusion enhancement layer.

7. The photovoltaic module defect detection method based on multimodal mid-frequency fusion and high-frequency sensing enhancement according to claim 1, characterized in that, Finally, loss calculation and optimization are performed. The total loss function is a weighted sum of four parts: localization loss, classification loss, target loss, and fusion consistency loss. Backpropagation is then used to calculate the gradient, update the network parameters, and finally, iterative optimization is performed, repeating training until the model converges. The total loss function is: ； in, , , , These are the weights of the localization loss, classification loss, targetability loss, and fusion consistency loss, respectively. It is the total loss. It is the localization loss, which calculates the error between the defect bounding box predicted by the model and the actual labeled box; It is the classification loss, which calculates the error between the model's prediction of the defect category and the true category label; It is a targeted loss function used to determine whether a defect exists within a mesh cell; This is the fusion consistency loss, used to constrain the alignment of bimodal features before fusion. It is obtained by calculating the L2 norm between the EL path P4 feature and the Depth path P4 feature after a 1x1 convolution transformation. ； in, This represents a 1x1 convolution operation used for channel alignment.

Images