A solar panel defect detection method and system

Abstract

This invention relates to the field of defect detection technology, and provides a method and system for detecting defects in solar panels. The method includes: acquiring an image of a solar panel; and obtaining a solar panel spot detection result through a solar panel defect detection network. The solar panel defect detection network includes a feature extraction module and a multi-scale feature fusion module connected in sequence. The feature extraction module includes a convolutional layer, a C3 module, three C3 modules, a Mobilenetbone module, three C3GHOST modules, a convolutional layer, an SPP module, and a C3 module connected in sequence. This method improves the accuracy and reliability of image detection for solar panel defects.

Description

Technical Field

[0001] This invention belongs to the field of defect detection technology, and in particular relates to a method and system for detecting defects in solar panels. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Photovoltaic power generation is now a way for humans to use solar energy to solve traditional energy problems. With the technological advancement and continuous development of clean energy, the application of solar panels is becoming more and more widespread.

[0004] During long-term use, solar panels can suffer varying degrees of damage due to weather conditions, bird droppings, and plant shading. Among these, the hot spot effect has a significant impact on solar panels, as it consumes some of the photovoltaic module's energy and, in severe cases, may damage or burn the module.

[0005] Existing spot recognition methods mainly rely on manual identification and image thresholding, but these methods suffer from inaccuracies, low efficiency, and high costs. Summary of the Invention

[0006] To address the technical problems mentioned above, this invention provides a method and system for detecting defects in solar panels. By incorporating a Mobilenetbone module and a C3GHOST module into the image recognition network structure, the accuracy and reliability of image detection for solar panel defects are improved.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] The first aspect of the present invention provides a method for detecting defects in a solar panel, comprising:

[0009] Acquire images of solar panels;

[0010] Based on solar panel images, the solar panel defect detection network is used to obtain the solar panel spot detection results;

[0011] The solar panel defect detection network includes a feature extraction module and a multi-scale feature fusion module connected in sequence. The feature extraction module includes a convolutional layer, a C3 module, a convolutional layer, three C3 modules, a Mobilenetbone module, three C3GHOST modules, a convolutional layer, an SPP module, and a C3 module connected in sequence. The three C3 modules, the Mobilenetbone module, and the last C3 module output features of different sizes, and the three features of different sizes are all fed into the multi-scale feature fusion module to obtain the spot features.

[0012] Furthermore, the multi-scale feature fusion module includes an FPN structure and a PAN structure;

[0013] The FPN structure includes a convolutional layer, an upsampling layer, a tensor concatenation layer, a C3 module, a convolutional layer, an upsampling layer, and a tensor concatenation layer connected in sequence.

[0014] The PAN structure includes a C3 module, a convolutional layer, a tensor splicing layer, a C3 module, a convolutional layer, a tensor splicing layer, and a C3 module connected in sequence.

[0015] Furthermore, the features output by the three C3 modules of the feature extraction module enter the last tensor concatenation layer of the FPN structure; the features output by the Mobilenetbone module of the feature extraction module enter the first tensor concatenation layer of the FPN structure; and the features output by the last C3 module of the feature extraction module enter the first convolutional layer of the FPN structure.

[0016] Furthermore, the features output from the first convolutional layer of the FPN structure enter the last tensor stitching layer of the PAN structure; the features output from the C3 module of the FPN structure enter the first tensor stitching layer of the PAN structure; the last tensor stitching layer of the FPN structure enters the first C3 module of the PAN structure; and the three C3 modules of the PAN structure output three features of different sizes respectively.

[0017] Furthermore, in addition to the input / output modules, the Mobilenetbone module also includes an inverse residual structure of five linear bottlenecks connected in sequence.

[0018] Furthermore, the C3 module includes a convolutional layer, a Bottleneck layer, a tensor splicing layer, and a convolutional layer connected in sequence, and then fused with a convolutional residual structure from the input to the Bottleneck layer.

[0019] Furthermore, the SPP module first passes the input data through a convolutional layer, then performs max pooling operations three times in parallel, and then concatenates the output of the three max pooling operations with the output tensor of the convolutional layer, and finally inputs it into a convolutional layer.

[0020] A second aspect of the present invention provides a solar panel defect detection system, comprising:

[0021] An image acquisition module is configured to acquire images of solar panels.

[0022] The defect detection module is configured to: obtain the solar panel spot detection results based on the solar panel image through the solar panel defect detection network;

[0023] The solar panel defect detection network includes a feature extraction module and a multi-scale feature fusion module connected in sequence. The feature extraction module includes a convolutional layer, a C3 module, a convolutional layer, three C3 modules, a Mobilenetbone module, three C3GHOST modules, a convolutional layer, an SPP module, and a C3 module connected in sequence. The three C3 modules, the Mobilenetbone module, and the last C3 module output features of different sizes, and the three features of different sizes are all fed into the multi-scale feature fusion module to obtain the spot features.

[0024] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a solar panel defect detection method as described above.

[0025] A fourth aspect of the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a solar panel defect detection method as described above.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] This invention provides a method for detecting defects in solar panels, which applies deep learning image recognition algorithms to the image detection of defects in solar panels. This method can greatly reduce the impact of hot spot effect on solar panels and has certain engineering value for solar panel fault detection.

[0028] This invention provides a method for detecting defects in solar panels. The method innovatively incorporates Mobilenetbone and C3GHOST modules into its image recognition network structure, which improves the accuracy and reliability of image detection of solar panel defects and has certain advantages compared with existing models. Attached Figure Description

[0029] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0030] Figure 1 This is a flowchart of a solar panel defect detection method according to Embodiment 1 of the present invention;

[0031] Figure 2 This is a flowchart of the image processing of the data input module in Embodiment 1 of the present invention;

[0032] Figure 3 This is a structural diagram of the solar panel defect detection network according to Embodiment 1 of the present invention;

[0033] Figure 4 This is a structural diagram of the Mobilenetbone module according to Embodiment 1 of the present invention;

[0034] Figure 5 This is a structural diagram of the C3GHOST module according to Embodiment 1 of the present invention;

[0035] Figure 6 This is a structural diagram of module C3 in Embodiment 1 of the present invention;

[0036] Figure 7 This is a structural diagram of the SPP module according to Embodiment 1 of the present invention;

[0037] Figure 8 This is a structural diagram of the multi-scale feature fusion module in Embodiment 1 of the present invention. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0039] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0040] Example 1

[0041] This embodiment provides a method for detecting defects in solar panels, aiming to accurately identify solar spot patterns on solar panels. The method includes the following steps:

[0042] Step 1: Obtain an image of the solar panel;

[0043] Step 2: Based on the solar panel image, obtain the solar panel spot detection results through the solar panel defect detection network.

[0044] The solar panel defect detection network consists of a feature extraction module, a multi-scale feature fusion module, and a classifier connected in sequence.

[0045] like Figure 1 As shown, the construction and training process of the solar panel defect detection network is as follows:

[0046] Step S1: Take an infrared thermal imaging photo of the solar panel to obtain an image showing the solar panel's light spot defects;

[0047] Step S2: Label the dataset for images with light spot defects;

[0048] Step S3: Construct a data input module to ensure that data can be successfully input into the training network;

[0049] Step S4: Construct a feature extraction module to extract the light spot features from the dataset;

[0050] Step S5: Construct a multi-scale feature fusion module to fuse the extracted multi-scale features and improve recognition accuracy;

[0051] Step S6: Construct a complete new network (solar panel defect detection network) structure, which can achieve image training and detection;

[0052] Step S7: Divide the dataset into a training set and a test set;

[0053] Step S8: Input the image data into the network structure for training, learn the spot features, and use a classifier to detect the spots on the solar panel based on the spot features. The detection result is to mark the location of the spot defects in the image.

[0054] The steps for acquiring image data in step S1 are as follows:

[0055] First, an infrared thermal imaging camera is used to photograph the solar panel to obtain a thermal image.

[0056] Secondly, the images are filtered to classify them into normal images and images with light spots;

[0057] Finally, the ratio of the image with glare to the normal image is controlled at 2:1.

[0058] The steps for dataset annotation in step S2 are as follows:

[0059] First, LabelImg was selected as the software for dataset annotation;

[0060] Secondly, the information about the light spots in the image is labeled;

[0061] Finally, construct a new, complete dataset with annotations.

[0062] The steps for constructing the data input module in step S3 are as follows:

[0063] First, perform data augmentation on the image, such as... Figure 2 As shown, four photos are randomly scaled, randomly cropped, and randomly arranged.

[0064] Secondly, the fixed input image size is set to 608 pixels × 608 pixels;

[0065] Finally, the image is scaled to a fixed size to facilitate feature training.

[0066] The steps for constructing the feature extraction module in step S4 are as follows:

[0067] First, a convolution plus C3 structure is used to extract preliminary feature information;

[0068] Secondly, the Mobilenetbone module was added to improve the network's ability to extract non-redundant features;

[0069] Secondly, the C3GHOST module is added to enhance the network's ability to extract diverse linear features;

[0070] Finally, a complete feature extraction module was built.

[0071] In one specific embodiment, a Mobilenetbone module is introduced, which includes depthwise separable convolutions and linear bottleneck inverse residual structures, ensuring network lightweighting while improving the network's ability to extract non-redundant information.

[0072] In one specific embodiment, the C3GHOST module is introduced, which uses normal convolution to generate real feature maps, and then uses these real feature maps to obtain the fantasy feature layer through linear transformation. Finally, the real feature layer and the fantasy feature layer are combined to form a complete feature layer.

[0073] The steps for constructing the multi-scale feature fusion module in step S5 are as follows:

[0074] First, a top-down FPN structure is adopted to fuse high-level feature information through upsampling;

[0075] Secondly, PAN is added to the FPN to convey strong positioning information from the bottom up;

[0076] Finally, the two layers work together to aggregate parameters from different feature extraction layers (backbone layers) for different detection layers.

[0077] The steps for constructing the complete network structure in step S6 are as follows:

[0078] First, connect the data input, feature extraction, and feature fusion modules;

[0079] Secondly, the feature output is divided into three feature maps: 19×19, 38×38, and 76×76, corresponding to the detection of defect features of different sizes.

[0080] Finally, a complete defect detection network model was constructed.

[0081] The steps for partitioning the dataset in step S7 are as follows:

[0082] First, the dataset is divided into two parts;

[0083] Secondly, the training set accounts for the first 80% of the dataset, and the test set accounts for the last 20% of the dataset.

[0084] Finally, the dataset was divided: the training set was used for training the network model, and the test set was used for recognition and detection.

[0085] This invention first inputs the training set into the feature extraction module, such as... Figure 3 As shown, the feature extraction module consists of a 3×3 convolutional kernel with a stride of 2, a C3 module, a 3×3 convolutional kernel with a stride of 2, three C3 modules, a Mobilenetbone module, three C3GHOST modules, a 3×3 convolutional kernel with a stride of 2, an SPP module, and a C3 module connected in sequence. Shallow features are output after the three C3 modules, mid-level features are output after the Mobilenetbone module, and deep features are output after the last C3 module.

[0086] The multi-scale feature fusion module includes an FPN structure and a PAN structure. The FPN structure consists of a convolutional layer, an upsampling layer, a tensor concatenation layer, a C3 module, another convolutional layer, another upsampling layer, and another tensor concatenation layer, all connected in sequence. The PAN structure consists of a C3 module, a convolutional layer, a tensor concatenation layer, another C3 module, another convolutional layer, another tensor concatenation layer, and another C3 module, all connected in sequence. Features output from the three C3 modules of the feature extraction module are fed into the last tensor concatenation layer of the FPN structure; features output from the Mobilenetbone module of the feature extraction module are fed into the first tensor concatenation layer of the FPN structure; features output from the last C3 module of the feature extraction module are fed into the first convolutional layer of the FPN structure. Features output from the first convolutional layer of the FPN structure are fed into the last tensor concatenation layer of the PAN structure; features output from the C3 module of the FPN structure are fed into the first tensor concatenation layer of the PAN structure; the last tensor concatenation layer of the FPN structure is fed into the first C3 module of the PAN structure. The three C3 modules of the PAN structure output three features of different sizes, i.e., the learned spot features.

[0087] To improve the network's ability to extract non-redundant features, this invention utilizes the Mobilenetbone module. MobileNetv3 introduces depthwise separable convolutions and the inverted residual with a linear bottleneck, making it a lightweight network. This invention incorporates a portion of MobileNetv3 called Mobilenetbone, such as... Figure 4 As shown, the Mobilenetbone module, in addition to the input / output modules, includes five sequentially connected InvertedResiduals (inverse residual structures of the linear bottleneck). The novel feature extraction network structure proposed in this invention maintains the first five layers of the feature extraction network unchanged, and uses the five-layer InvertedResidual structure of the Mobilenetv3 network to improve the network's ability to extract non-redundant information. The main structure of the InvertedResidual is a convolutional layer, and the residual structure exists only in the first and last layers. Therefore, this structure has a strong feature extraction capability for non-redundant information. The InvertedResidual adopts a structure that is wide in the middle and narrow at both ends. Although there are fewer residual structures, it greatly reduces the probability of information loss during feature extraction. This makes it lose less information when extracting non-redundant information.

[0088] To enhance the network's ability to extract diverse linear features, this invention proposes the use of the C3GHOST module. Ghostnet uses normal convolutions to generate partially real feature maps, then uses these real feature maps through linear transformations (cheap operations) to obtain ghost feature maps. Finally, the complete feature layer is composed of the real feature layer and the ghost feature layer, making it a common feature extraction network. Figure 6 The diagram shows a traditional C3 module. Its feature extraction module consists of a 1×1 convolutional kernel with a stride of 1 (convolutional layer), a Bottleneck layer, a tensor concatenation layer, another 1×1 convolutional kernel with a stride of 1, and then a convolutional residual structure from the input to the Bottleneck layer. The Bottleneck layer significantly reduces the amount of data computation by using small convolutional kernels instead of large ones. Its structure is as follows: first a 1×1 convolutional kernel, then a 3×3 convolutional kernel, and finally a 1×1 convolutional kernel.

[0089] and Figure 5 Compared to the C3GHOST module, this invention replaces the 1×1 convolutional kernel module with a stride of 1 with a GHOST convolutional kernel module. Ghostnet uses the features extracted by convolution to map new features. Placing this feature extraction structure, which can extract non-redundant features, at the end of the feature extraction structure improves the anti-interference ability of the feature extraction network. This gives the new feature extraction network stronger generalization ability, and therefore, the new network is more sensitive to targets with poor features.

[0090] In this invention, the SPP module is as follows: Figure 7 As shown, the input data is first passed through a 1×1 convolutional kernel with a stride of 1 (one convolutional layer), then max pooling is performed three times in parallel. The tensors from the first 1×1 convolutional kernel with a stride of 1 are then concatenated together, and finally passed through another 1×1 convolutional kernel with a stride of 1.

[0091] In convolutional neural networks, deeper networks provide stronger feature information about the target, leading to better predictions. However, this also weakens the target's positional information. Furthermore, the continuous convolution process can cause information loss for small targets. Therefore, predictions at different scales are necessary. Figure 8To perform multi-scale feature fusion of images using FPN and PAN, the upper-layer feature maps, due to their deeper network layers, contain stronger semantic information about light spots, while the lower-layer features, having passed through fewer convolutional layers, suffer less loss of positional information about the light spots. The FPN structure performs top-down upsampling, ensuring that the lower-layer feature maps contain stronger semantic information about the light spots; the PAN structure performs bottom-up downsampling, ensuring that the top-layer features contain strong positional information about the light spots. The two features are then fused, so that feature maps of different sizes contain both strong semantic and feature information about the light spots, guaranteeing accurate prediction of light spot images of different sizes.

[0092] This invention uses Alpha-IoU as the loss function during network training, balancing the accuracy and recall of the symplectic model and further improving the precision of the network structure. Alpha-IoU is a novel type of IoU that uses exponentiation in the IoU and penalty terms expressions. In clean or noisy environments, it does not introduce additional parameters or increase training and inference time, and its effect is more pronounced on small datasets. Its formula is as follows:

[0093] α -IOU=1-IOUα

[0094] The parameter α can be used as a hyperparameter to adjust the α-IOU loss to meet different levels of bounding box regression accuracy. α > 1 achieves high regression accuracy by focusing more on the High IOU objective (i.e., the High IOU threshold). Empirically, α is not overly sensitive to different models or datasets; in most cases, α = 3 consistently performs well. The α-IOU loss family can be easily used to improve detector performance without introducing additional parameters or increasing training / inference time in clean or noisy environments.

[0095] In a specific embodiment, the operating system supporting the experiment was Windows 10, the CPU was an Intel(R) Core(R) i7-8700 CPU 3.70GHz, the GPU was a GeForce GTX 2080ti with 11GB of video memory, and the framework used was PyTorch. The dataset was divided into training and test sets in an 8:2 ratio, and the division was repeated six times for each experiment. The initial learning rate was 0.01, the termination learning rate was 0.2, 8 images were input into the model each time, the number of data loaders was 4, and the number of training rounds was 200. The experiment showed that the average accuracy of detecting solar panel light spots was about 90%.

[0096] This embodiment provides a method for detecting defects in solar panels. First, images of solar panels are collected using a camera. Simultaneously, a novel image recognition network structure is built to effectively identify and locate light spots on the solar panels. Second, the image recognition network structure innovatively incorporates Mobilenetbone and C3GHOST modules, which improves the accuracy and reliability of the invention for detecting defects in solar panels, and has certain advantages compared with existing models.

[0097] The solar panel defect detection method provided in this embodiment has superior performance, can more accurately inspect and identify the light spots of solar panels, and reduce the maintenance cost of solar panels.

[0098] This embodiment provides a method for detecting defects in solar panels. Taking into account the hot spot effect of solar panels, a new deep learning image recognition algorithm is applied to the image detection of solar panel defects. This method can greatly reduce the impact of the hot spot effect on solar panels and has certain engineering value for solar panel fault detection.

[0099] Example 2

[0100] This embodiment provides a solar panel defect detection system, which specifically includes:

[0101] An image acquisition module is configured to acquire images of solar panels.

[0102] The defect detection module is configured to: obtain the solar panel spot detection results based on the solar panel image through the solar panel defect detection network;

[0103] The solar panel defect detection network includes a feature extraction module and a multi-scale feature fusion module connected in sequence. The feature extraction module includes a convolutional layer, a C3 module, a convolutional layer, three C3 modules, a Mobilenetbone module, three C3GHOST modules, a convolutional layer, an SPP module, and a C3 module connected in sequence. The three C3 modules, the Mobilenetbone module, and the last C3 module output features of different sizes, and the three features of different sizes are all fed into the multi-scale feature fusion module to obtain the spot features.

[0104] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0105] Example 3

[0106] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a solar panel defect detection method as described in Embodiment 1 above.

[0107] Example 4

[0108] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the solar panel defect detection method described in Embodiment 1 above.

[0109] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0110] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0111] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0112] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0113] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0114] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting defects in solar panels, characterized in that, include: Acquire images of solar panels; Based on solar panel images, the solar panel defect detection network is used to obtain the solar panel spot detection results; The solar panel defect detection network includes a feature extraction module and a multi-scale feature fusion module connected in sequence. The feature extraction module includes a convolutional layer, a C3 module, a convolutional layer, three C3 modules, a Mobilenetbone module, three C3GHOST modules, a convolutional layer, an SPP module, and a C3 module connected in sequence. The three C3 modules, the Mobilenetbone module, and the last C3 module output features of different sizes, and the three features of different sizes are all fed into the multi-scale feature fusion module to obtain the spot features. The Mobilenetbone is a part of Mobilenetv3; in addition to the input and output modules, the Mobilenetbone module also includes an inverse residual structure of five linear bottlenecks connected in sequence. The C3 module includes a convolutional layer, a Bottleneck layer, a tensor splicing layer, and a convolutional layer connected in sequence, and then fused with a convolutional residual structure from the input to the Bottleneck layer.

2. The method for detecting defects in a solar panel as described in claim 1, characterized in that, The multi-scale feature fusion module includes an FPN structure and a PAN structure; The FPN structure includes a convolutional layer, an upsampling layer, a tensor concatenation layer, a C3 module, a convolutional layer, an upsampling layer, and a tensor concatenation layer connected in sequence. The PAN structure includes a C3 module, a convolutional layer, a tensor splicing layer, a C3 module, a convolutional layer, a tensor splicing layer, and a C3 module connected in sequence.

3. The method for detecting defects in solar panels as described in claim 2, characterized in that, The features output by the three C3 modules of the feature extraction module enter the last tensor concatenation layer of the FPN structure; the features output by the Mobilenetbone module of the feature extraction module enter the first tensor concatenation layer of the FPN structure; and the features output by the last C3 module of the feature extraction module enter the first convolutional layer of the FPN structure.

4. The method for detecting defects in a solar panel as described in claim 2, characterized in that, The features output from the first convolutional layer of the FPN structure enter the last tensor splicing layer of the PAN structure; the features output from the C3 module of the FPN structure enter the first tensor splicing layer of the PAN structure; the last tensor splicing layer of the FPN structure enters the first C3 module of the PAN structure; the three C3 modules of the PAN structure output three features of different sizes respectively.

5. The method for detecting defects in a solar panel as described in claim 1, characterized in that, The SPP module first passes the input data through a convolutional layer, then performs max pooling operations three times in parallel, and then concatenates the output of the three max pooling operations with the output tensor of the convolutional layer, and finally inputs it into a convolutional layer.

6. A solar panel defect detection system, characterized in that, include: An image acquisition module is configured to acquire images of solar panels. The defect detection module is configured to: obtain the solar panel spot detection results based on the solar panel image through the solar panel defect detection network; The solar panel defect detection network includes a feature extraction module and a multi-scale feature fusion module connected in sequence. The feature extraction module includes a convolutional layer, a C3 module, a convolutional layer, three C3 modules, a Mobilenetbone module, three C3GHOST modules, a convolutional layer, an SPP module, and a C3 module connected in sequence. The three C3 modules, the Mobilenetbone module, and the last C3 module output features of different sizes, and the three features of different sizes are all fed into the multi-scale feature fusion module to obtain the spot features. The Mobilenetbone is a part of Mobilenetv3; in addition to the input and output modules, the Mobilenetbone module also includes an inverse residual structure of five linear bottlenecks connected in sequence. The C3 module includes a convolutional layer, a Bottleneck layer, a tensor splicing layer, and a convolutional layer connected in sequence, and then fused with a convolutional residual structure from the input to the Bottleneck layer.

7. A solar panel defect detection system as described in claim 6, characterized in that, The multi-scale feature fusion module includes an FPN structure and a PAN structure; The FPN structure includes a convolutional layer, an upsampling layer, a tensor concatenation layer, a C3 module, a convolutional layer, an upsampling layer, and a tensor concatenation layer connected in sequence. The PAN structure includes a C3 module, a convolutional layer, a tensor splicing layer, a C3 module, a convolutional layer, a tensor splicing layer, and a C3 module connected in sequence.

8. A solar panel defect detection system as described in claim 6, characterized in that, The SPP module first passes the input data through a convolutional layer, then performs max pooling operations three times in parallel, and then concatenates the output of the three max pooling operations with the output tensor of the convolutional layer, and finally inputs it into a convolutional layer.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the solar panel defect detection method as described in any one of claims 1-5.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the solar panel defect detection method as described in any one of claims 1-5.

Images