Solar wing photovoltaic module crack segmentation method based on adaptive frequency domain noise suppression

Abstract

The application discloses a solar wing photovoltaic module crack segmentation method based on adaptive frequency domain noise suppression. The segmentation model comprises an adaptive frequency domain noise suppression module, a dual-domain feature extraction and gate fusion module, a crack feature enhancement module, a crack boundary perception module, a crack thinning module and a decoder; an input image is denoised through the adaptive frequency domain noise suppression module, the dual-domain feature extraction and gate fusion module extracts spatial domain features and frequency domain features of the denoised image and fuses the features to obtain dual-domain fusion features; the dual-domain fusion features are subjected to crack enhancement through the crack enhancement module, the crack enhancement features are subjected to boundary enhancement through the crack boundary perception module, the boundary enhancement features are subjected to crack thinning through the crack thinning module, the crack thinning features and the dual-domain fusion features are added element by element, and then the decoder is used to generate a crack segmentation map. The method effectively suppresses noise and improves the coherent segmentation capability and segmentation precision of slender cracks.

Description

Technical Field

[0001] This invention belongs to the field of defect segmentation technology, specifically a crack segmentation method for solar photovoltaic modules based on adaptive frequency domain noise suppression. Background Technology

[0002] Solar photovoltaic (PV) modules are critical components of spacecraft energy systems. Internal cracks, microcracks, and micro-defects directly impact the output efficiency and operational reliability of the solar array. Therefore, high-precision non-destructive testing of solar PV modules plays a vital role. Existing deep learning methods, such as DeepLabV3+, have been widely applied to defect segmentation tasks, achieving good segmentation results for common cracks and large-area defects. However, for crack segmentation of solar PV modules, due to the complex detection environment, the unique photoluminescence (PL) imaging link, and the small scale of slender cracks, existing methods still have the following shortcomings in segmenting slender cracks under strip noise interference: (1) Strip noise and periodic interference In the process of photoluminescence imaging, factors such as excitation light source, PL imaging link and driving power supply may introduce strip noise, periodic texture or fixed pattern noise into the imaging. These interferences manifest as a large number of false edges in the spatial domain, which can easily lead to false detection or missegmentation of cracks.

[0003] (2) Thin, long cracks are prone to breakage Cracks in solar photovoltaic modules typically exhibit characteristics such as being long and thin, small in scale, and having low contrast. During the spatial domain downsampling and feature fusion process of the model, excessive smoothing can easily cause the long and thin cracks to break or be missed, thus affecting the complete identification of the long and thin cracks.

[0004] (3) Unclear boundary positioning Under PL imaging conditions, crack boundaries are often accompanied by weak light transition regions or noise contamination, resulting in blurred or jagged boundaries, which reduces the accuracy of crack segmentation.

[0005] To address the aforementioned shortcomings, this invention proposes a crack segmentation method for solar photovoltaic modules based on adaptive frequency domain noise suppression. This method solves problems such as frequency domain noise interference and difficulty in identifying crack boundaries in the crack segmentation of solar photovoltaic modules, thereby improving the ability and accuracy of coherent segmentation of slender cracks. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a method for crack segmentation of solar photovoltaic modules based on adaptive frequency domain noise suppression.

[0007] The present invention solves the aforementioned technical problem by adopting the following technical solution: A method for crack segmentation of solar photovoltaic modules based on adaptive frequency domain noise suppression includes the following steps: Step 1: Collect photoluminescence images of several solar photovoltaic modules under different operating conditions, including normal and cracked defects, and construct a dataset; Step 2: Construct a segmentation model, including an adaptive frequency domain noise reduction module, an encoder, and a decoder; the encoder includes a dual-domain feature extraction and gating fusion module, a crack feature enhancement module, a crack boundary perception module, and a crack refinement module; In the adaptive frequency domain noise reduction module, the photoluminescence image of the solar panel photovoltaic module is subjected to a two-dimensional Fourier transform to obtain a frequency domain complex spectrum. The amplitude spectrum and phase spectrum are extracted from the frequency domain complex spectrum. The amplitude spectrum and phase spectrum are passed through a mask prediction network to generate a frequency domain noise reduction mask. The frequency domain complex spectrum and the frequency domain noise reduction mask are multiplied by a Hadamard to obtain a filtered frequency domain complex spectrum. The filtered frequency domain complex spectrum is subjected to an inverse Fourier transform to obtain the denoised photoluminescence image of the solar panel photovoltaic module. The dual-domain feature extraction and gating fusion module extracts spatial features from the denoised solar photovoltaic module photoluminescence image and extracts frequency features from the logarithmic amplitude spectrum. The spatial and frequency features are then fused to obtain dual-domain fused features. In the crack feature enhancement module, the dual-domain fusion features are processed in parallel through 1×1 convolution branches, multi-scale dilated convolution branches, directional convolution branches, and global average pooling branches. The output features of all branches are concatenated and then processed through 1×1 convolution, batch normalization, and ReLU activation function to obtain crack enhancement features. The crack boundary perception module includes a boundary prediction branch and a boundary enhancement branch. In the boundary prediction branch, the crack enhancement features are sequentially processed by 3×3 convolution, batch normalization, ReLU activation function, and 1×1 convolution to generate a boundary probability distribution. A boundary prediction map is generated based on the boundary probability distribution. The boundary prediction map is then processed by 1×1 convolution and Sigmoid activation function to generate a spatial gating map. In the boundary enhancement branch, the spatial gating map undergoes a residual gating fusion operation to inject boundary information into the crack enhancement features. The output features of the residual gating fusion operation are then processed by 1×1 convolution, batch normalization, and ReLU activation function to obtain the boundary enhancement features. In the crack refinement module, the boundary enhancement features are processed in parallel by multi-scale dilated convolution and oriented convolution branches. The output features of all branches are concatenated and then subjected to 1×1 convolution to obtain multi-branch fusion features. The multi-branch fusion features are then processed by the SE module and k×k convolution. The output features of the SE module and k×k convolution are concatenated and then subjected to 3×3 convolution to obtain crack refinement features. Here, k is an integer greater than 1. After adding the crack refinement features and the dual-domain fusion features element by element, and then passing them through a decoder, the crack segmentation map is obtained. Step 3: Train the segmentation model using the dataset, and then use the trained segmentation model to segment cracks in the solar panel photovoltaic module.

[0008] Furthermore, the dual-domain feature extraction and gating fusion module includes a spatial encoder, a frequency encoder, and a gating fusion module. The denoised photoluminescence image of the solar photovoltaic module is processed by the spatial encoder to extract spatial features; the logarithmic amplitude spectrum is processed by the frequency encoder to extract frequency features. In the gating fusion module, firstly, the frequency features are aligned with the spatial features, then the spatial features are concatenated with the aligned frequency features along the channel dimension, and then gated response features are generated through 1×1 convolution, batch normalization, and ReLU activation. The gated response features are processed by the Sigmoid activation function to generate a gating coefficient matrix. The gating coefficient matrix is ​​multiplied element-wise with the aligned frequency features to obtain weighted frequency features. The spatial features are added element-wise with the weighted frequency features to obtain the dual-domain fused features.

[0009] Furthermore, the spatial domain encoder and the frequency domain encoder have the same structure, both including multiple coding layers, with downsampling operations embedded between adjacent coding layers; each coding layer includes two 3×3 convolutions, followed by batch normalization and ReLU activation functions after each 3×3 convolution.

[0010] Furthermore, the multi-scale dilated convolution branch of the crack feature enhancement module includes parallel 3×3 dilated convolutions with dilation rates of 6, 12, and 18, and the directional convolution branch includes parallel 1×k and k×1 convolutions.

[0011] Furthermore, the multi-scale dilated convolution branch of the crack refinement module includes parallel 3×3 dilated convolutions with dilation rates of 1, 3, and 5, and the directional convolution branch includes parallel 1×k and k×1 convolutions.

[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. The adaptive frequency domain noise reduction module is used to specifically suppress strip noise and periodic interference in the photoluminescence image of the solar photovoltaic module, avoiding false boundaries and improving segmentation accuracy.

[0013] 2. The crack feature enhancement module and crack refinement module capture contextual information and the overall morphology of cracks under different receptive fields through multi-scale dilated convolution branches, and enhance the response of fine cracks in the vertical and horizontal directions through directional convolution branches, thereby improving the model's separability for slender cracks, divergent cracks, and long strip defects, making the model more sensitive and robust to slender cracks.

[0014] 3. The crack boundary perception module predicts and enhances crack boundaries. Supervised training using the boundary prediction map makes the model's crack boundary localization clearer, effectively avoiding crack boundary ambiguity and enhancing the clarity and consistency of crack boundaries. The crack boundary perception module and the crack refinement module achieve cascaded enhancement through explicit boundary learning and detail-oriented compensation. Attached Figure Description

[0015] Figure 1 This is an overall structural diagram of the segmentation model of the present invention; Figure 2 This is a structural diagram of the adaptive frequency domain noise suppression module of the present invention; Figure 3 This is a structural diagram of the dual-domain feature extraction and gating fusion module of the present invention; Figure 4 This is a structural diagram of the crack boundary sensing module of the present invention; Figure 5 This is a structural diagram of the crack refinement module of the present invention; Figure 6 Visualization of crack segmentation results for different models. Detailed Implementation

[0016] Specific embodiments are given below with reference to the accompanying drawings. These specific embodiments are only used to describe the technical solution of the present invention in detail, and are not intended to limit the scope of protection of this application.

[0017] This invention provides a method for crack segmentation of solar photovoltaic modules based on adaptive frequency domain noise suppression, comprising the following steps: Step 1: Collect photoluminescence images of several solar photovoltaic modules under different operating conditions, including normal images and crack defect images. Perform pixel-level annotation on the crack defect images to construct a dataset. During the training phase, data augmentation techniques such as random scaling, flipping, rotation, and brightness perturbation can be used to improve the model's generalization ability.

[0018] Step 2: Construct the segmentation model; such as... Figure 1 As shown, the model includes an adaptive frequency domain noise reduction module, an encoder, and a decoder. The encoder includes a dual-domain feature extraction and gating fusion module, a crack feature enhancement module, a crack boundary perception module, and a crack refinement module. The photoluminescence image of the solar photovoltaic module is denoised by an adaptive frequency domain denoising module, resulting in a denoised photoluminescence image and its logarithmic amplitude spectrum. A dual-domain feature extraction and gating fusion module extracts spatial and frequency domain features from the denoised photoluminescence image and its logarithmic amplitude spectrum, respectively, and fuses these features to obtain dual-domain fused features. These dual-domain fused features are then enhanced by a crack feature enhancement module to obtain crack-enhanced features. The crack-enhanced features are further enhanced by a crack boundary perception module to extract boundary enhancement features, which are then refined by a crack refinement module to extract crack refinement features. Finally, the crack refinement features and the dual-domain fused features are added element-wise, and then decoded by a decoder. The decoder uses layer-by-layer upsampling to restore the original resolution, resulting in the segmentation result.

[0019] like Figure 2 As shown, in the adaptive frequency domain noise reduction module, the photoluminescence image of the solar panel photovoltaic module is transformed by a two-dimensional Fourier transform to obtain the frequency domain complex spectrum. Calculate the amplitude spectrum based on the complex spectrum in the frequency domain. Phase spectrum According to the amplitude spectrum Obtain the logarithmic amplitude spectrum ;in, This represents the pixel's representation in the frequency domain; the amplitude and phase spectra are passed through a mask prediction network (this implementation uses the MaskNet model) to generate a frequency domain noise reduction mask. According to equation (1), the frequency domain complex spectrum and the frequency domain noise suppression mask are multiplied by a Hadamard product to suppress stripe noise and periodic interference, thus obtaining the filtered frequency domain complex spectrum. The filtered frequency domain complex spectrum is subjected to inverse Fourier transform to obtain the denoised photoluminescence image of the solar panel photovoltaic module. (1) in, This represents the Hadama product.

[0020] like Figure 3 As shown, the dual-domain feature extraction and gating fusion module includes a spatial domain encoder, a frequency domain encoder, and a gating fusion module. The denoised photoluminescence image of the solar panel photovoltaic module undergoes layer-by-layer downsampling and feature extraction via the spatial domain encoder to obtain spatial features containing crack details and local edge information. The logarithmic amplitude spectrum of the photoluminescence image of the solar panel photovoltaic module... Frequency domain features representing strip noise and periodic interference are extracted by a frequency domain encoder; the spatial domain features and frequency domain features are then fused by a gated fusion module to obtain dual-domain fused features.

[0021] The spatial encoder and the frequency encoder have the same structure, both including multiple coding layers. The output features of the previous coding layer are downsampled and used as the input features of the next coding layer. Each coding layer includes two 3×3 convolutions, followed by batch normalization and ReLU activation function.

[0022] In the gated fusion module, firstly, the frequency domain features are size aligned (using upsampling or downsampling) and mapped to channels to ensure that the frequency domain features and spatial domain features are consistent in spatial resolution and channel dimension, resulting in aligned frequency domain features. The spatial domain features and the aligned frequency domain features are then concatenated along the channel dimension, and subsequently mapped to gated response features through 1×1 convolution, batch normalization, and ReLU activation. The gated response features are then processed by a Sigmoid activation function to generate a gate coefficient matrix. The gate coefficient matrix is ​​then multiplied element-wise with the aligned frequency domain features to obtain weighted frequency domain features. Finally, the spatial domain features are added element-wise with the weighted frequency domain features to obtain dual-domain fusion features, which are used for subsequent segmentation and boundary refinement.

[0023] In the crack feature enhancement module, the dual-domain fusion features are processed in parallel through 1×1 convolution branches, multi-scale dilated convolution branches, directional convolution branches, and global average pooling branches. The output features of all branches are concatenated, and then channel compression and fusion are performed using 1×1 convolution, batch normalization, and the ReLU activation function. This simultaneously enhances the crack representation in both scale and direction dimensions, resulting in enhanced crack features. The multi-scale dilated convolution branches include 3×3 dilated convolutions (DConv) with dilation rates of 6, 12, and 18, used to capture crack context information and overall morphology under different receptive fields. The directional convolution branches use 1×k and k×1 convolutions to extract crack features from the vertical and horizontal directions, respectively, to enhance the response of fine cracks in the vertical and horizontal directions and improve the model's ability to distinguish slender cracks, divergent cracks, and elongated defects. Here, k is an integer greater than 1 and can be adaptively set according to the crack length.

[0024] like Figure 4 As shown, the crack boundary perception module includes a boundary prediction branch and a boundary enhancement branch. In the boundary prediction branch, the crack enhancement features are sequentially processed by 3×3 convolution, batch normalization, ReLU activation function, and 1×1 convolution to generate a boundary probability distribution. This boundary probability distribution is then converted into a boundary prediction map, where pixels with high probability are darker and pixels with low probability are lighter. Figure 1On the one hand, boundary loss is introduced during the training phase to supervise model training, enabling the model to explicitly learn boundary positions and contour directions, thereby reducing the impact of boundary blurring, contour breakage, and pseudo-edge adhesion on the segmentation results. On the other hand, the boundary prediction map is sequentially processed by 1×1 convolution and a Sigmoid activation function to generate a spatial gating map for boundary enhancement. In the boundary enhancement branch, the spatial gating map undergoes a residual gating fusion operation to inject boundary information into the crack enhancement features. The fused features are then sequentially processed by 1×1 convolution, batch normalization, and a ReLU activation function for channel shaping to obtain the boundary enhancement features, which are the output features of the boundary enhancement branch. The residual gating fusion operation uses Hadamard product and residuals to maintain the stability of the crack enhancement features while enhancing the boundary response. Specifically, it is expressed as follows: (2) in, This represents the output characteristics of the residual gated fusion operation, i.e., the fusion characteristics; Indicates crack reinforcement characteristics. Represents a spatial gating diagram.

[0025] like Figure 5 As shown, in the crack refinement module, the boundary enhancement features are refined by multi-scale dilated convolution and directional convolution branches. The output features of all branches are concatenated and then shaped by 1×1 convolution to obtain multi-branch fusion features. The multi-branch fusion features are then processed by the SE module and k×k convolution, using attention and directional refinement mechanisms to compensate and correct local details of the crack, thereby improving the coherence and positioning accuracy of the slender crack. The output features of the SE module and k×k convolution are concatenated and then processed by 3×3 convolution to obtain the crack refinement features. The multi-scale dilated convolution branches include 3×3 dilated convolutions with dilation rates of 1, 3, and 5, used to compensate for the detail information of the crack under different widths and contrasts. The directional convolution branches use 1×k and k×1 convolutions to further enhance the slender structure and bifurcation trend of the crack from the vertical and horizontal directions, respectively. The SE module uses channel attention to weight and highlight "suspected fine crack areas" in order to suppress excessive smoothing and loss of detail during the decoding stage, and improve the continuity of crack boundaries, fine crack recall, and positioning accuracy.

[0026] Step 3: Use the dataset built in Step 1 to train the segmentation model, and use the trained segmentation model for crack segmentation of solar panel photovoltaic modules; During the training phase, the joint loss function of Equation (3) is used to calculate the total loss of the model and optimize the model. The total loss of the model includes the segmentation main branch loss and the boundary prediction branch loss. The segmentation main branch loss is calculated based on the crack segmentation map output by the model and the real crack segmentation map, including cross-entropy loss and Dice loss, which is used to improve the detection accuracy of slender cracks under class imbalance conditions. The boundary prediction branch loss is calculated based on the boundary prediction map and the real crack boundary map, including cross-entropy loss and Dice loss, in order to enhance the continuity of crack boundaries and the positioning accuracy. (3) in, This represents the total loss of the model. , This represents the cross-entropy loss and Dice loss for splitting the main branch. , The cross-entropy loss and Dice loss represent the boundary prediction branches; , and This is the balance coefficient.

[0027] Example To verify the effectiveness of the method of the present invention, a comparative experiment was conducted on the crack segmentation task of solar photovoltaic modules. The comparison models included UNet, TransUNet, and DeepLabv3+. The test results are shown in Table 1.

[0028] Table 1. Comparison of test results for different models (%)

[0029] As shown in Table 1, all indicators of the method of this invention are significantly improved. The mIoU of the method of this invention reaches 84.3%, which is 6.5% higher than DeeplabV3+, 11.8% higher than UNet, and 8.9% higher than TransUNet, indicating the best overall segmentation performance. The Dice coefficient of the method of this invention reaches 87.5%, which is 5.4% higher than DeeplabV3+, indicating a significant improvement in the detection and segmentation continuity of small cracks. The boundary F1 score of the method of this invention reaches 82.6%, which is 8.4% higher than DeeplabV3+, indicating a significant improvement in the boundary ambiguity problem, making the boundary localization clearer. The accuracy of the method of this invention reaches 86.0%, which is 5.6% higher than DeeplabV3+, indicating advantages in reducing false detections and improving the reliability of practical applications.

[0030] Figure 6The image shows a visualization of the crack segmentation results of the method of this invention and DeeplabV3+. As can be seen from the image, the method of this invention can completely segment slender cracks, while DeeplabV3+ has defects such as crack fracture and unclear boundaries.

[0031] In summary, the method of this invention addresses the problems of strip noise and periodic noise in solar photovoltaic modules, which easily lead to false boundaries and false detections, as well as the low contrast of fine cracks, the fragility of slender structures, and the ease with which they are missed. It introduces an adaptive frequency domain noise suppression module to specifically suppress noise frequency bands in the input image, and achieves complementary enhancement of spatial and frequency features through a dual-domain feature extraction and gating fusion module. Furthermore, it explicitly strengthens boundary localization and compensates for the detailed information of slender cracks through a crack feature enhancement module, a crack boundary perception module, and a crack refinement module, thereby improving the clarity of the segmented boundaries, crack coherence, and stability. This method can simultaneously output crack segmentation maps and boundary prediction maps, and can output frequency domain noise suppression masks and denoised images as interpretable intermediate results, facilitating engineering adjustments and quality traceability. It has good practical value and promising prospects for widespread application. It is not only applicable to crack segmentation of solar photovoltaic modules, but can also be extended to industrial visual inspection tasks involving periodic noise or strip interference, such as cracks on metal surfaces, concrete cracks, and crack defects in electronic devices.

[0032] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for crack segmentation of solar photovoltaic modules based on adaptive frequency domain noise suppression, characterized in that, Includes the following steps: Step 1: Collect photoluminescence images of several solar photovoltaic modules under different operating conditions, including normal and cracked defects, and construct a dataset; Step 2: Construct a segmentation model, including an adaptive frequency domain noise reduction module, an encoder, and a decoder; the encoder includes a dual-domain feature extraction and gating fusion module, a crack feature enhancement module, a crack boundary perception module, and a crack refinement module; In the adaptive frequency domain noise reduction module, the photoluminescence image of the solar panel photovoltaic module is transformed by two-dimensional Fourier transform to obtain the frequency domain complex spectrum; the amplitude spectrum and phase spectrum are extracted from the frequency domain complex spectrum, and the amplitude spectrum and phase spectrum are passed through a mask prediction network to generate a frequency domain noise reduction mask; the frequency domain complex spectrum and the frequency domain noise reduction mask are multiplied by Hadamard to obtain the filtered frequency domain complex spectrum. The filtered frequency domain complex spectrum is subjected to inverse Fourier transform to obtain the denoised photoluminescence image of the solar panel photovoltaic module; The dual-domain feature extraction and gating fusion module extracts spatial features from the denoised solar photovoltaic module photoluminescence image and extracts frequency features from the logarithmic amplitude spectrum. The spatial and frequency features are then fused to obtain dual-domain fused features. In the crack feature enhancement module, the dual-domain fusion features are processed in parallel through 1×1 convolution branches, multi-scale dilated convolution branches, directional convolution branches, and global average pooling branches. The output features of all branches are concatenated and then processed through 1×1 convolution, batch normalization, and ReLU activation function to obtain crack enhancement features. The crack boundary perception module includes a boundary prediction branch and a boundary enhancement branch. In the boundary prediction branch, the crack enhancement features are sequentially processed by 3×3 convolution, batch normalization, ReLU activation function, and 1×1 convolution to generate a boundary probability distribution. A boundary prediction map is generated based on the boundary probability distribution. The boundary prediction map is then processed by 1×1 convolution and Sigmoid activation function to generate a spatial gating map. In the boundary enhancement branch, the spatial gating map undergoes a residual gating fusion operation to inject boundary information into the crack enhancement features. The output features of the residual gating fusion operation are then processed by 1×1 convolution, batch normalization, and ReLU activation function to obtain the boundary enhancement features. In the crack refinement module, the boundary enhancement features are processed in parallel by multi-scale dilated convolution and oriented convolution branches. The output features of all branches are concatenated and then subjected to 1×1 convolution to obtain multi-branch fusion features. The multi-branch fusion features are then processed by the SE module and k×k convolution. The output features of the SE module and k×k convolution are concatenated and then subjected to 3×3 convolution to obtain crack refinement features. Here, k is an integer greater than 1. After adding the crack refinement features and the dual-domain fusion features element by element, and then passing them through a decoder, the crack segmentation map is obtained. Step 3: Train the segmentation model using the dataset, and then use the trained segmentation model to segment cracks in the solar panel photovoltaic module.

2. The method for crack segmentation of solar photovoltaic modules based on adaptive frequency domain noise suppression according to claim 1, characterized in that, The dual-domain feature extraction and gating fusion module includes a spatial encoder, a frequency encoder, and a gating fusion module; the denoised solar photovoltaic module photoluminescence image is processed by the spatial encoder to extract spatial features; The logarithmic amplitude spectrum is used to extract frequency domain features through a frequency domain encoder. In the gated fusion module, the frequency domain features are first aligned with the spatial features, and then the spatial features are concatenated with the aligned frequency domain features along the channel dimension. Then, the gated response features are generated by 1×1 convolution, batch normalization, and ReLU activation function. The gated response features are used to generate a gate coefficient matrix through a Sigmoid activation function. The gate coefficient matrix is ​​multiplied element-wise with the aligned frequency domain features to obtain the weighted frequency domain features. The spatial features are added element-wise with the weighted frequency domain features to obtain the dual-domain fused features.

3. The method for crack segmentation of solar photovoltaic modules based on adaptive frequency domain noise suppression according to claim 2, characterized in that, The spatial encoder and the frequency encoder have the same structure, both including multiple coding layers, with downsampling operations embedded between adjacent coding layers; each coding layer includes two 3×3 convolutions, followed by batch normalization and ReLU activation function.

4. The method for crack segmentation of solar photovoltaic modules based on adaptive frequency domain noise suppression according to claim 1, characterized in that, The multi-scale dilated convolution branch of the crack feature enhancement module includes parallel 3×3 dilated convolutions with dilation rates of 6, 12, and 18, and the directional convolution branch includes parallel 1×k and k×1 convolutions.

5. The method for crack segmentation of solar photovoltaic modules based on adaptive frequency domain noise suppression according to any one of claims 1 to 4, characterized in that, The multi-scale dilated convolution branch of the crack refinement module includes parallel 3×3 dilated convolutions with dilation rates of 1, 3, and 5, and the directional convolution branch includes parallel 1×k and k×1 convolutions.

Images