Photovoltaic module dust image feature extraction method based on improved semantic segmentation model
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2026-07-10
- Publication Date
- 2026-08-07
AI Technical Summary
[0005]本发明的目的在于克服现有技术中积灰边界定位不准确、多尺度积灰区域漏检、背景噪声误分割以及少数类积灰召回率低的问题，提供一种基于改进语义分割模型的光伏组件积灰图像特征提取方法
[0015]传统的光伏组件积灰视觉检测方法多采用有灰或无灰的图像分类，或者直接使用普通SegNet模型进行像素级分割，难以同时处理积灰尺度差异大、边缘模糊以及背景反光干扰等问题，而本发明通过在SegNet编码器—解码器网络中融合积灰边缘特征增强SERB模块、改进多尺度感知ASPP-R模块和语义注意力模块，实现对粗颗粒积灰、细颗粒积灰和湿性积灰的像素级识别，并提取积灰空间位置、覆盖度和分布占比。
Smart Images
Figure CN122530723A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of visual detection technology for surface contamination of photovoltaic modules, specifically involving a pixel-level semantic segmentation, dust coverage calculation, and dust type distribution ratio extraction method for coarse-particle dust accumulation, fine-particle dust accumulation, and wet dust accumulation on the surface of photovoltaic modules. Background Technology
[0002] Photovoltaic modules are exposed to the outdoor environment for extended periods, and the surface of the module's glass cover easily accumulates dust, sand, and industrial particles, which can form wet dust accumulation under the influence of moisture. This dust accumulation blocks and scatters solar radiation, reducing the cover's light transmittance and consequently causing a decrease in output power and power generation efficiency. Accurately identifying dust accumulation areas, types of dust, and their spatial distribution is fundamental to developing cleaning plans, assessing power loss, and ensuring the safe and economical operation of photovoltaic power plants.
[0003] Existing visual detection methods for dust accumulation on photovoltaic modules mostly rely on image classification based on whether the dust is present or absent, or simple classification based on color, texture, and shape. These methods struggle to obtain pixel-level quantitative information such as dust coverage, type proportion, and spatial location. While semantic segmentation models can achieve pixel-level recognition, dust accumulation on the surface of photovoltaic modules is characterized by large scale differences, irregular shapes, and blurred boundaries. Furthermore, the acquired images are easily affected by background noise such as metal borders, module stripes, shadows, localized reflections, and watermarks.
[0004] Traditional SegNet networks are prone to losing edge details during multiple downsampling processes, and their fixed receptive field makes it difficult to simultaneously capture fine-grained dust accumulation and large-area clustered dust accumulation. In the decoding stage, relying solely on upsampling and pooling indexes to recover spatial information is insufficient for suppressing reflections and background artifacts. Furthermore, the number of pixels in each type of dust accumulation is typically significantly less than the number of pixels in the background and clean components; directly applying ordinary cross-entropy loss can easily bias the model towards the majority class, leading to missed detections of dust accumulation. Therefore, a feature extraction method for photovoltaic module dust accumulation images that can enhance edge information, integrate multi-scale context, filter semantic features, and accommodate class imbalance is needed. Summary of the Invention
[0005] The purpose of this invention is to overcome the problems of inaccurate localization of dust accumulation boundaries, missed detection of multi-scale dust accumulation areas, missegmentation due to background noise, and low recall rate of minority classes of dust accumulation in the prior art, and to provide a method for extracting features of dust accumulation images of photovoltaic modules based on an improved semantic segmentation model.
[0006] To achieve the above objectives, the present invention adopts the following technical solution, comprising the following steps:
[0007] Step 1: Obtain images of photovoltaic modules under different illumination and dust accumulation conditions in public datasets and actual photovoltaic power plants. Mark the background, clean modules, coarse dust particles, fine dust particles, and wet dust in the images at the pixel level, and simultaneously enhance the images and their mask labels to construct a dust accumulation image dataset.
[0008] Further, the specific process of step 1 is as follows: collect and summarize the original images, create a semantic mask corresponding to each pixel, expand the samples using rotation, flipping, mesh perspective, lens distortion, Gaussian blur and Mosaic methods, and then divide the standardized samples into training set, validation set and test set.
[0009] Step 2: Construct an improved semantic segmentation model (SegNet) based on the SegNet encoder-decoder network. Embed a Sobel EdgeReinforcement Block (SERB) in the encoder, set an improved Atrous Spatial Pyramid Pooling-Residual Module (ASPP-R) at the end of the encoder, and set a Semantic Attention Module (SAM) in the decoder.
[0010] Further, the specific process of step 2 is as follows: the encoder extracts hierarchical features and saves the max-pooling indices; the SERB module enhances the ash texture and geometric boundaries; the ASPP-R module fuses the ash features and global illumination information of different receptive fields; the semantic attention module suppresses borders, stripes, shadows and reflective noise; and finally, the decoder restores spatial details and outputs pixel-level probability maps of five semantic categories.
[0011] Step 3: Set category weights according to the sample distribution of the five semantic categories, train the improved semantic segmentation model using the weighted negative log-likelihood loss function, and save the optimal model parameters based on the segmentation performance of the validation set to obtain the gray image segmentation model.
[0012] Furthermore, the specific process of step 3 is as follows: one-hot encoding of the mask labels, calculation of class-weighted loss, iterative update of network parameters using the Adam optimizer, and evaluation of the model using intersection-union ratio, accuracy, and recall.
[0013] Step 4: Input the image of the photovoltaic module to be detected into the dust accumulation image segmentation model, output the dust accumulation type mask, count the number of pixels of each type of dust accumulation, and calculate the total dust accumulation coverage, the coverage of each type of dust accumulation and its distribution ratio in the total dust accumulation area.
[0014] The beneficial technical effects brought about by this invention are as follows:
[0015] Traditional visual detection methods for dust accumulation in photovoltaic modules often employ image classification based on whether the dust is present or absent, or directly use the ordinary SegNet model for pixel-level segmentation. These methods struggle to simultaneously address issues such as large differences in dust accumulation scale, blurred edges, and background reflection interference. In contrast, this invention integrates a dust accumulation edge feature enhancement SERB module, an improved multi-scale perception ASPP-R module, and a semantic attention module into the SegNet encoder-decoder network. This enables pixel-level recognition of coarse-grained dust accumulation, fine-grained dust accumulation, and wet dust accumulation, and extracts the spatial location, coverage, and distribution ratio of the dust accumulation.
[0016] Secondly, the SERB module in this invention uses Sobel gradient and channel attention to enhance the texture and geometric boundaries of the accumulated dust, the ASPP-R module uses multi-diffraction dilated convolution and global contextual information to perceive the accumulated dust regions at different scales, and the semantic attention module uses long-range dependencies between pixels to filter the semantic features of the accumulated dust. The three modules work together to reduce the missed detection of small accumulated dust, the adhesion of blurred boundaries, and the missegmentation caused by metal borders, component stripes, shadows and local reflections, thereby improving the segmentation accuracy and robustness of the model in complex photovoltaic power station scenarios.
[0017] This invention employs a weighted negative log-likelihood loss function and an iterative optimization method to train an improved semantic segmentation model. The class weights are set according to the recognition difficulty of different semantic categories to reduce the misclassification cost of minority classes such as coarse-grained dust accumulation, fine-grained dust accumulation, and wet dust accumulation. The optimal model parameters are selected through a validation set, so that the model can balance the overall segmentation accuracy and dust accumulation category recall during the training process.
[0018] Experimental results from this invention on public datasets and multi-scenario datasets collected from actual photovoltaic power plants show that the accuracy of the proposed method reaches 96.57%, the cross-validation ratio reaches 85.73%, and the recall rate reaches 89.61%, all of which are higher than the comparison model. At the same time, it can output the total dust accumulation coverage, the coverage of various types of dust accumulation, and their distribution ratio in the total dust accumulation area, thereby providing reliable structured data for photovoltaic module cleaning priority judgment and power loss quantitative assessment. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the overall process of the photovoltaic module dust accumulation image feature extraction method based on an improved semantic segmentation model according to the present invention.
[0020] Figure 2 This is a schematic diagram illustrating the dataset construction and synchronized data enhancement of this invention;
[0021] Figure 3 This is a schematic diagram of the overall network structure of the improved SegNet semantic segmentation model of this invention;
[0022] Figure 4 This is a schematic diagram of the SERB dust accumulation edge feature enhancement module structure of the present invention;
[0023] Figure 5 This is a schematic diagram of the ASPP-R improved multi-scale sensing module structure of the present invention;
[0024] Figure 6 This is a schematic diagram of the semantic attention module structure of the present invention;
[0025] Figure 7 This is a schematic diagram of the model training and optimal parameter selection process of the present invention;
[0026] Figure 8 This is a schematic diagram of the ash accumulation segmentation, type identification, and coverage calculation process of the present invention;
[0027] Figure 9 This is a schematic diagram comparing the accuracy of the present invention with that of the comparative model;
[0028] Figure 10 This is a schematic diagram of the segmentation results of the present invention on a public dataset;
[0029] Figure 11 This is a schematic diagram of the segmentation results of the present invention on a self-made dataset;
[0030] Figure 12 This is a schematic diagram of the dust accumulation mask, thermal map, and coverage output results of the present invention. Detailed Implementation
[0031] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the present invention and not to limit the scope of protection of the present invention; equivalent substitutions for parameters, number of network layers, input size and enhancement methods without departing from the concept of the present invention are all included within the scope of the present invention.
[0032] like Figure 1 As shown, this invention proposes a method for extracting features from dust accumulation images of photovoltaic modules based on an improved semantic segmentation model, specifically including the following steps:
[0033] Step 1: Obtain images of dust accumulation on photovoltaic modules and create pixel-level semantic labels. Perform synchronous data augmentation on the images and their mask labels to construct a dust accumulation image dataset, such as... Figure 2 As shown;
[0034] Furthermore, the specific process of step 1 is as follows:
[0035] Step 1.1: Obtain dust accumulation samples from the public dataset DeepSolarEye, and collect images of photovoltaic modules under different lighting conditions, different background reflections, and different dust accumulation conditions in actual photovoltaic power plants to form an original image set; the dust accumulation conditions include uniformly distributed fine-particle dust accumulation, clustered coarse-particle dust accumulation, and wet dust accumulation formed by water vapor.
[0036] Step 1.2: Use a polygon annotation tool to annotate the original image pixel by pixel. The semantic categories include background, clean components, coarse-grained dust accumulation, fine-grained dust accumulation, and wet dust accumulation. Generate a single-channel mask of the same size for each original image so that the mask pixel values correspond one-to-one with the five semantic categories.
[0037] Step 1.3: Perform synchronous enhancement while keeping the geometric transformation parameters of the original image and the mask consistent. The synchronous enhancement includes one or more of the following: 90° random rotation, horizontal flip, vertical flip, mesh perspective, lens distortion, and Gaussian blur. Mosaic enhancement is used to stitch four randomly cropped and affine transformed images and their masks onto a 2×2 canvas, combining different backgrounds, scales, and grayscale patterns within a single training sample.
[0038] Step 1.4: Divide the enhanced images and mask labels into training, validation and test sets in a ratio of 8:1:1, and normalize the images and masks to 256×256 pixels.
[0039] This embodiment uses 213 original images, which are enhanced to obtain 1491 samples, and the training set, validation set and test set are divided in an 8:1:1 ratio.
[0040] Step 2: Construct an improved semantic segmentation model based on the SegNet encoder-decoder network, such as... Figure 3 As shown;
[0041] Furthermore, the specific process of step 2 is as follows:
[0042] Step 2.1: Input the 256×256 pixel image into the SegNet encoder, and use multi-level convolution, normalization, activation and max pooling operations to extract texture and semantic features step by step and save the max pooling index; the decoder uses the corresponding max pooling index to perform non-linear upsampling to reduce the loss of spatial location information.
[0043] Step 2.2: Set up SERB modules in the third to fifth stages of the encoder, with corresponding feature map resolutions of 32×32, 16×16, and 8×8, respectively. The SERB module includes a main branch, an edge enhancement branch, and an attention filtering branch, as shown below. Figure 4 As shown;
[0044] Furthermore, the specific process of step 2.2 is as follows:
[0045] Step 2.2.1: The main branch uses a 1×1 convolution to perform channel dimensionality reduction on the input features, with a channel compression ratio r of 16; the edge enhancement branch uses horizontal and vertical Sobel convolution kernels to extract gradient responses in two directions. The convolution kernels are:
[0046] (1)
[0047] Step 2.2.2: Pass the gradient responses in both directions through 3×3 convolution, normalization, and ReLU activation functions sequentially, then sum them element-wise. Finally, use 1×1 convolution to compress the channels to obtain the fused edge features.
[0048] (2)
[0049] Step 2.2.3: Perform global average pooling on the fused edge features to obtain the descriptor of the c-th channel:
[0050] (3)
[0051] Step 2.2.4: Input the channel descriptors sequentially into two fully connected layers, the ReLU function, and the Sigmoid function to obtain the channel attention weights:
[0052] (4)
[0053] Step 2.2.5: Apply channel attention weights to the fused edge features channel by channel, and add them to the residuals of the main branch features to obtain the SERB module output:
[0054] (5)
[0055] The above structure not only enhances the ash accumulation texture and geometric boundaries, but also preserves shallow information and improves gradient propagation through residual connections.
[0056] Step 2.3: Set up the ASPP-R module at the end of the encoder; the input image is compressed by the backbone network to form an 8×8 feature map. The ASPP-R module uses three types of processing paths: multi-scale dilated convolution, global context, and reflection thinning, such as... Figure 5 As shown;
[0057] Furthermore, the specific process of step 2.3 is as follows:
[0058] Step 2.3.1: First, compress the input features to 64 channels using a channel compression ratio r=4. Then, input the features into four 3×3 dilated convolution branches with different dilation rates. The output of the i-th branch is:
[0059] (6)
[0060] Step 2.3.2: The global context branch compresses the 8×8 feature map into a 1×1 channel statistic through global average pooling, and then interpolates it to the original feature resolution after 1×1 convolution and Sigmoid activation to perceive the illumination distribution of the entire component.
[0061] Step 2.3.3: Concatenate the four dilated convolutional features and one global context feature along the channel dimension to obtain multi-scale concatenated features:
[0062] (7)
[0063] Step 2.3.4: Use 3×3 grouped convolutions for cross-scale information exchange, and use 1×1 convolutions to generate gating signals for residual calibration, obtaining the ASPP-R module output:
[0064] (8)
[0065] Step 2.4: Embed a semantic attention module when the decoder restores the feature map to 32×32 resolution, such as... Figure 6 As shown;
[0066] Furthermore, the specific process of step 2.4 is as follows:
[0067] Step 2.4.1: Map the input features to query features Q, key features K and value features V respectively; where the semantic dimension of the key features is 5, corresponding to background, clean components, coarse-grained dust accumulation, fine-grained dust accumulation and wet dust accumulation.
[0068] Step 2.4.2: Perform matrix multiplication on the query features and key features, and obtain the attention map by normalization using the Softmax function.
[0069] (9)
[0070] Step 2.4.3: Weight the value features using the attention map and reshape them into a spatial feature map. After dimensionality increase through a 1×1 convolution, add it to the input feature residual to obtain the output of the semantic attention module:
[0071] (10)
[0072] This attention map represents long-range dependencies between pixels, enabling self-enhancing of features within the dust accumulation area while mitigating noise responses caused by component stripes, metallic borders, and localized reflections.
[0073] Step 2.5: Use the decoder to upsample step by step, and combine it with the pooling index of the corresponding layer of the encoder to restore spatial details. After passing through the classification layer and the Softmax function, output the pixel-level probability map of five semantic classes.
[0074] Step 3: Train the improved semantic segmentation model using the gray image dataset and the weighted negative log-likelihood loss function to obtain the gray image segmentation model, such as... Figure 7 As shown;
[0075] Furthermore, the specific process of step 3 is as follows:
[0076] Step 3.1: Convert pixel-level real labels into one-hot encoding and set category weights for the five semantic categories; set higher weights for wet dust and other difficult-to-identify dust categories, and lower weights for background and clean components;
[0077] Step 3.2: Assuming a single sample contains N pixels and the number of semantic categories is 5, calculate the weighted negative log-likelihood loss based on the true label, category weights, and the predicted probability output by the model using the Softmax function:
[0078] (11)
[0079] In the formula, c represents the semantic category, w_c represents the category weight of the c-th category, y represents the true label, and the predicted probability represents the output probability of the model for the corresponding category.
[0080] Step 3.3: Train for 200 rounds using the Adam optimizer, an initial learning rate of 0.001, and a batch size of 16. Calculate the validation set loss and segmentation metrics in each round. Save the current parameters when the validation performance improves, thereby obtaining the optimal gray image segmentation model.
[0081] The training and inference process can be implemented using the following pseudocode.
[0082] Algorithm 1: Pseudocode for training and gray feature extraction of the improved semantic segmentation model. Input: Training image and mask D; validation set V; image to be detected I; number of training epochs E Output: Optimal model M*; dust accumulation type mask P; coverage and distribution ratio R 1. Initialize the parameters of the SegNet encoder-decoder, SERB, ASPP-R, and semantic attention modules. 2 for epoch = 1 to E do 3 for each batch of training samples (x,y) in D do 4. Perform synchronization enhancements with consistent parameters for x and y. 5 F ← Encoder(x); F ← SERB(F); F ← ASPP-R(F) 6 p ← Decoder(SemanticAttention(F)) 7 L ← WeightedNLL(p,y); Backpropagate and update parameters using Adam. 8 end for 9. Calculate IoU, Accuracy, and Recall on V; save M* if performance is improved. 10 end for 11 P ← argmax(M*(I)) 12. Count the number of pixels with various types of gray accumulation in P, and calculate the total coverage, coverage of each type, and distribution ratio. 13 return M*, P, R
[0083] Step 4: Input the image of the photovoltaic module to be detected into the dust accumulation image segmentation model, output a dust accumulation type mask, and extract the spatial location, coverage, and distribution ratio of various types of dust accumulation, such as... Figure 8 As shown;
[0084] Furthermore, the specific process of step 4 is as follows:
[0085] Step 4.1: Perform pixel-by-pixel maximum probability decision on the five semantic probability maps output by the model to generate a dust accumulation type mask; exclude the image background through the effective area mask of the photovoltaic module;
[0086] Step 4.2: Count the number of pixels with coarse dust particles, fine dust particles, and wet dust in the mask respectively; calculate the total dust coverage by the ratio of the sum of the number of pixels with each type of dust accumulation to the total number of pixels in the effective area of the photovoltaic module.
[0087] (12)
[0088] In the formula, N_coarse, N_fine, and N_wet represent the number of pixels with coarse-grained dust accumulation, fine-grained dust accumulation, and wet dust accumulation, respectively, and N_total represents the total number of pixels in the effective area of the photovoltaic module.
[0089] Step 4.3: Calculate the coverage of each type of dust accumulation or the distribution percentage of that type of dust accumulation in the total dust accumulation area:
[0090] (13)
[0091] In the formula, i belongs to the set of coarse-grained dust accumulation, fine-grained dust accumulation, and wet dust accumulation categories; when the denominator is N_total, the result is the coverage of this type of dust accumulation; when the denominator is the sum of the number of pixels of the three types of dust accumulation, the result is the distribution ratio of this type of dust accumulation in the total dust accumulation area.
[0092] Step 4.4: Output the dust accumulation mask, category heat map, total coverage, coverage of each category and distribution ratio, and store the data as a structured record for use as input for cleaning priority judgment or photovoltaic module power loss assessment model.
[0093] To verify the effectiveness of the method of this invention, a mixture of publicly available datasets and a self-made dataset was used for training, and the method SegNet-SAS was compared with HRNet and DeepLabv3+. Evaluation metrics included accuracy, intersection-over-union (IoU), and recall. The comparison results are shown in the table below.
[0094] Model Accuracy (%) IoU (%) Recall (%) SegNet-SAS 96.57 85.73 89.61 HRNet 92.93 82.16 84.27 DeepLabv3+ 90.13 78.93 82.15
[0095] As shown in the table above, the method of this invention achieves an accuracy of 96.57%, an IoU of 85.73%, and a recall of 89.61%, all three indicators being higher than the comparative model. Figure 9 As shown, the improved SegNet-SAS method outperforms the HRNet and DeepLabv3+ comparison models in the final convergence value of the gray image feature extraction and segmentation task, and tends to plateau after about 100 rounds.
[0096] In test images of the public dataset, such as Figure 10 As shown, the method of this invention can accurately segment coarse and fine dust particles and reduce false identification of clean areas; in reflective, shadowy, and mixed dust scenes in the self-made dataset, such as... Figure 11 As shown, the method of the present invention can still distinguish between coarse particles, fine particles, and wet ash. For example... Figure 12 As shown, the total dust coverage of the example image can be output as 8.34%, while in another example, the coarse particle dust coverage is 4.88% and the fine particle dust coverage is 2.06%, indicating that this method can provide quantitative results that can be used for operation and maintenance decisions.
[0097] In summary, this invention, through the synergistic effect of SERB edge enhancement, ASPP-R multi-scale perception, semantic attention filtering, and weighted loss training, can achieve high-precision segmentation, dust accumulation type identification, and quantitative feature extraction of the dust accumulation area of photovoltaic modules under complex operating conditions.
Claims
1. A method for extracting features from images of dust accumulation in photovoltaic modules based on an improved semantic segmentation model, characterized in that, Includes the following steps: Step 1: Obtain images of dust accumulation on photovoltaic modules and create pixel-level semantic tags corresponding to the images of dust accumulation on photovoltaic modules. Perform synchronous data augmentation on the images of dust accumulation on photovoltaic modules and their semantic tags to construct a dust accumulation image dataset. Step 2: Construct an improved semantic segmentation model (SegmentationNetwork, SegNet) based on the SegNet encoder-decoder network. Embed a Sobel EdgeReinforcement Block (SERB) in the encoder, set an improved Atrous Spatial Pyramid Pooling-Residual Module (ASPP-R) at the end of the encoder, and set a Semantic Attention Module (SAM) in the decoder. Step 3: Train the improved semantic segmentation model using the gray image dataset and the weighted negative log-likelihood loss function to obtain the gray image segmentation model; Step 4: Input the image of the photovoltaic module to be detected into the dust accumulation image segmentation model, and output the pixel-level segmentation results of the background, clean module, coarse dust accumulation, fine dust accumulation and wet dust accumulation. Then, extract the coverage and distribution ratio of each type of dust accumulation based on the pixel-level segmentation results.
2. The method for extracting features from photovoltaic module dust accumulation images based on an improved semantic segmentation model according to claim 1, characterized in that, Step 1 further includes the following steps: Step 1.1: Collect images of photovoltaic modules under different lighting conditions and different dust accumulation conditions, and obtain publicly available dust accumulation images to form an original image set; Step 1.2: According to the physical causes and appearance of the ash accumulation, label the ash accumulation area as coarse-grained ash accumulation, fine-grained ash accumulation, or wet ash accumulation, and generate a mask label corresponding to each pixel of the original image. Step 1.3: While keeping the spatial correspondence between the image and the mask label unchanged, synchronous enhancement is performed using one or more of the following methods: 90° random rotation, horizontal flip, vertical flip, mesh perspective, lens distortion, and Gaussian blur. Mosaic enhancement is then used to stitch multiple images that have been randomly cropped and affine transformed onto a 2×2 canvas. Step 1.4: Divide the enhanced images and mask labels into training, validation and test sets in a ratio of 8:1:1, and standardize the image size to 256×256 pixels.
3. The method for extracting features from photovoltaic module dust accumulation images based on an improved semantic segmentation model according to claim 1, characterized in that, Step 2 further includes the following steps: Step 2.1: Use the SegNet encoder to perform hierarchical convolution and downsampling on the input image to obtain encoded features at different resolutions; Step 2.2: Embed SERB modules in the third to fifth stages of the encoder, and fuse shallow texture, edge gradient and channel semantic information through the main branch, edge enhancement branch and attention filtering branch; Step 2.3: Set up the ASPP-R module at the end of the encoder to obtain multi-scale ash accumulation features through four dilated convolution branches with different dilation rates, a global context branch, and a reflection thinning branch; Step 2.4: When the decoder restores the feature map to 32×32 resolution, a semantic attention module is set up to perform dual spatial and semantic filtering based on the query-key-value mechanism; Step 2.5: Use the decoder to upsample step by step and combine it with the encoder's max-pooling indices to recover spatial details and output pixel-level probability maps of five semantic classes.
4. The method for extracting features from photovoltaic module dust accumulation images based on an improved semantic segmentation model according to claim 3, characterized in that, Step 2.2 further includes the following steps: Step 2.2.1: The main branch uses a 1×1 convolution to perform channel dimensionality reduction on the input features, with a channel compression ratio r of 16; the edge enhancement branch uses horizontal and vertical Sobel convolution kernels to extract gradient responses in two directions. The convolution kernels are: （1） Step 2.2.2: After sequentially applying 3×3 convolution, normalization, and ReLU activation to the gradient responses in both directions, sum them element-wise, and then perform channel compression using 1×1 convolution to obtain the fused edge features. （2） Step 2.2.3: Perform global average pooling on the fused edge features to obtain the descriptor of the c-th channel: （3） Step 2.2.4: Input the channel descriptors sequentially into two fully connected layers, the ReLU function, and the Sigmoid function to obtain the channel attention weights: （4） Step 2.2.5: Utilize the channel attention weights to perform channel-by-channel weighting on the fused edge features, and add the weighted features to the main branch feature residuals to obtain the SERB module output: （5） Where ReLU represents the linear rectified function and Sigmoid represents the sigmoid activation function. This represents the convolution operation.
5. The method for extracting features from photovoltaic module dust accumulation images based on an improved semantic segmentation model according to claim 3, characterized in that, Step 2.3 further includes the following steps: Step 2.3.1: Compress the 8×8 feature map output by the encoder to 64 channels, and input them into four 3×3 dilated convolution branches with different dilation rates. The output of the i-th dilated convolution branch is: （6） Step 2.3.2: Compress the 8×8 feature map into a 1×1 channel statistic using global average pooling, and obtain global context features through 1×1 convolution, Sigmoid activation and interpolation upsampling; Step 2.3.3: Concatenate the four-channel dilated convolutional features and the global context features along the channel dimension to obtain multi-scale concatenated features: （7） Step 2.3.4: Perform cross-scale information interaction on the multi-scale stitched features using 3×3 grouped convolution, and perform residual calibration using the gated signal generated by 1×1 convolution to obtain the ASPP-R module output: （8） Among them, the dilation rate combination of the four-way dilated convolution is used to cover ash accumulation areas of different sizes, and the channel compression ratio r is 4.
6. The method for extracting features from photovoltaic module dust accumulation images based on an improved semantic segmentation model according to claim 3, characterized in that, Step 2.4 further includes the following steps: Step 2.4.1: Map the input features to query features, key features and value features respectively. The semantic dimension of the key features is 5, corresponding to background, clean components, coarse-grained dust accumulation, fine-grained dust accumulation and wet dust accumulation respectively. Step 2.4.2: Perform matrix multiplication on the query features and key features, and then normalize using Softmax to obtain an attention map representing the long-range dependencies between pixels. （9） Step 2.4.3: Use the attention map to weight the value features and reshape them into a spatial feature map. Then, after increasing the dimensionality through a 1×1 convolution, add it to the input feature residual to obtain the output of the semantic attention module. （10）。 7. The method for extracting features from photovoltaic module dust accumulation images based on an improved semantic segmentation model according to claim 1, characterized in that, Step 3 further includes the following steps: Step 3.1: Perform one-hot encoding on the pixel-level ground truth labels of the training samples, and set the class weights corresponding to background, clean components, coarse-grained dust accumulation, fine-grained dust accumulation, and wet dust accumulation; Step 3.2: Calculate the single-sample weighted negative log-likelihood loss based on the category weights of the five semantic classes and the pixel category probabilities output by the model through the Softmax function. （11） Step 3.3: The improved semantic segmentation model is iteratively trained using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 16, and the model parameters are determined using the validation set.
8. The method for extracting features from photovoltaic module dust accumulation images based on an improved semantic segmentation model according to claim 1, characterized in that, Step 4 further includes the following steps: Step 4.1: Perform pixel-by-pixel maximum probability decision on the pixel-level probability maps of the five semantic categories to obtain the dust accumulation type mask; Step 4.2: Count the number of pixels with coarse-grained dust, fine-grained dust, and wet dust in the dust accumulation type mask respectively: （12） Step 4.3: Calculate the dust coverage of each type of dust accumulation based on the number of pixels with each type of dust accumulation and the total number of pixels in the effective area of the photovoltaic module. Then, calculate the distribution percentage of each type of dust accumulation in the total dust accumulation area based on the number of pixels with each type of dust accumulation and the total number of pixels with dust accumulation. （13） Step 4.4: Output the spatial location, coverage, and distribution ratio of various types of dust accumulation as structured input for photovoltaic module cleaning decisions or power loss assessments.