As global energy demand continues to rise alongside growing environmental concerns, renewable energy technologies have become a central focus of research and development. Among these, solar energy has emerged as a key player, with solar panel installations expanding rapidly worldwide. However, solar panels face significant operational challenges, including cracks, dust accumulation, and other defects that reduce photoelectric conversion efficiency. addressing these issues is critical because even minor defects can lead to substantial energy losses over time. In particular, dust accumulation can corrode the protective surface layer of solar panels, cause short circuits in internal circuits, and create serious safety hazards. Therefore, developing efficient and accurate methods for detecting solar panel defects is of paramount practical importance.
Traditional approaches to inspecting solar panels, such as manual inspection, are time-consuming, labor-intensive, and often fail to provide consistent, objective results. Meanwhile, conventional image processing techniques, while useful in controlled settings, frequently struggle with the complexity and variability of real-world solar panel defects. In response, deep learning has emerged as a transformative approach for automated defect detection and classification. In this work, I present a comprehensive investigation into deep learning-based methods for solar panel defect detection, with a particular focus on image classification and edge feature extraction.
The structure of this thesis is organized as follows: chapter 2 introduces the construction and enhancement of solar panel datasets, detailing the defect types and the augmentation techniques used to improve the quantity and quality of data. Chapter 3 presents a novel randomly connected network model for solar panel defect classification. Chapter 4 introduces a dual-channel multi-scale attention mechanism for solar panel edge feature extraction. Finally, chapter 5 concludes the work and discusses future directions.
Dataset Construction and Augmentation
In the field of deep learning, the performance of neural networks is heavily dependent on the quantity and quality of training data. When defective and non-defective solar panel images are mixed within the same network, significant convergence issues can arise, particularly when the diversity of defect samples is limited. Since public datasets for solar panel defects are relatively small, data augmentation is essential to expand the dataset’s scope and depth, thereby enhancing model accuracy and robustness. Image enhancement techniques aim to improve model performance by removing noise and interference while making the similarities and differences between objects in the image more evident.
Defect Types in Solar Panels
The dataset used in this research was sourced from publicly available electroluminescence (EL) image datasets. It includes 300×300 grayscale EL images containing visible defects such as micro-cracks, fragments, and black spots, as well as defect-free EL images. To ensure experimental reliability, 600 EL images were randomly selected and divided into training, validation, and test sets. This division allows for scientific evaluation of model generalization.
Three main types of solar panel defect types are focused on:
Micro-cracks: As modern solar panels become increasingly thin, external forces during manufacturing and transportation can produce micro-cracks in the cells. These cracks are often difficult to detect visually, appearing only as subtle lines. However, with an EL testing instrument, dark brown lines become clearly visible, indicating the presence of micro-cracks. Over time, these cracks may worsen due to mechanical load and aging, potentially causing open circuit damage. This significantly threatens the safety and reliability of solar panels.
Fragments: Fragment defects are more severe than micro-cracks and typically arise from the continuous deterioration of micro-cracks. External impacts or improper soldering during production can also cause fragmentation. When fragment defects are present, EL testing reveals large irregular black areas with sharp edges. These defects directly reduce the power output of solar panels and, in extreme cases, render entire panels unusable.
Black spots: Black spot defects are severe phenomena caused by substandard materials or improper manufacturing processes. EL testing can identify black spot defects, which appear as one or more black circles in the center of the image. These defects not only reduce power generation efficiency but may also damage other components, thereby affecting the lifespan and stability of solar panels.
Data Augmentation Techniques
The performance of a convolutional neural network (CNN) is closely related to the amount of training data. When training data are insufficient or class distribution is imbalanced, network performance may degrade. The collected defect image datasets are relatively small, and the imbalance between normal and defective images may bias the model toward the majority class. To effectively improve CNN performance, preprocessing and augmentation must be applied before feeding images into the model. Specifically, image augmentation techniques such as random rotation, cropping, and flipping are used to generate additional defect samples.
Image quality is an important factor influencing the efficiency and accuracy of defect recognition algorithms. However, in practice, poor image quality may prevent algorithms from correctly analyzing image data, thus affecting overall system performance. Image enhancement addresses this issue by improving the informative content of images. The methods applied in this work include image preprocessing, geometric transformation, and a generative adversarial network (GAN) approach.
Image Preprocessing: Preprocessing is used to enhance image quality by removing noise and improving contrast. Spatial domain methods, such as local averaging and median filtering, are employed to suppress noise. Gamma transformation and adaptive histogram equalization are also applied to improve contrast. Image binarization is used to convert grayscale images into binary images, enabling clearer segmentation of defect regions. Image smoothing is applied to reduce noise while preserving edge information.
Geometric Methods: To address the issue of non-uniform defect distribution, the image dataset was adjusted so that each defect type occupies a reasonable proportion. Since the collection of defective images is challenging, techniques including rotation, horizontal and vertical flipping, and symmetry transformations are used to enhance the dataset. Although the augmented images may appear similar to the original ones, a CNN trained on these different augmented images demonstrates improved robustness and accuracy. The augmented EL image samples include original images, x-axis symmetry reversal, y-axis symmetry reversal, and x-y axis symmetry reversal.
Generative Adversarial Network Method: Dataset quality is crucial for network training, as it helps the network avoid learning irrelevant features. A deep convolutional generative adversarial network (DCGAN) is used to generate realistic images that can effectively improve the discriminator’s judgment capability. The DCGAN model consists of a generator and a discriminator. The generator produces synthetic images from random noise, and the discriminator compares them with real images to determine authenticity. When the discriminator labels a real image as “1” and a generated image as “0”, the generator adjusts its parameters accordingly to produce increasingly realistic images. After iterative training, the generated images become nearly indistinguishable from real ones and are added to the real dataset for augmentation. The DCGAN architecture includes four deconvolutional layers in the generator and four convolutional layers in the discriminator.
Through the combination of geometric transformation and DCGAN methods, the dataset was expanded to 1100 images, ensuring both quantity and quality suitable for network training.
Random Connection Network for Solar Panel Defect Detection
In the past, solar cell defect inspection suffered from high costs, limited functionality, and low efficiency. With the rapid development of computer science and the availability of large-scale data, deep learning has emerged as a powerful tool. By extracting useful features from target samples, deep learning enables machines to perform autonomous analysis and self-learning. Traditional solar panel defect detection methods, such as manual inspection and remote sensing image judgment, have many limitations. To address these issues, I adopt deep learning technology for solar panel defect classification.
Fundamentals of Convolutional Neural Networks
CNN architectures are composed of several fundamental layers: input layers, convolutional layers, pooling layers, fully connected layers, and activation functions.
Input Layer: The input layer is the starting point of a CNN. Data are typically preprocessed before being fed into the network to avoid differences in data units and to shorten training time. In computer vision applications, the input is typically an image represented as a multi-dimensional tensor.
Convolutional Layer: The convolutional layer is composed of a set of filters (kernels) and bias values, and its primary function is feature extraction and dimensionality reduction. Multiple filters efficiently extract complex features by performing dot product operations on the input matrix. The convolution operation is expressed as:
$$ P(i,j) = \sum_{a=0}^{m-1}\sum_{b=0}^{n-1} Q(a,b) \cdot f(i-a, j-b) $$
where P is the convolution output, Q is the convolution kernel of size (m,n), and f is the input matrix. Different parameter settings for kernel size, stride, and padding determine the output scale of the convolution layer.
Pooling Layer: Pooling layers are inserted between successive convolution layers to reduce spatial dimensions, thus decreasing computational complexity. Pooling operations provide translation invariance, reduce overfitting, and increase the receptive field. Two common types are average pooling and max pooling.
Fully Connected Layer: The fully connected layers, typically located at the end of a CNN, perform classification or regression tasks. Each neuron in a fully connected layer is connected to every neuron in the previous layer. The operation is expressed as:
$$ z = Wx + b $$
where x is the input feature vector, W is the weight matrix, b is the bias, and z is the output.
Activation Functions: Activation functions introduce nonlinearity into neural networks, allowing models to learn complex patterns. Common activation functions include Sigmoid, tanh, and ReLU:
$$ \sigma(x) = \frac{1}{1+e^{-x}} $$
$$ \tanh(x) = \frac{2}{1+e^{-2x}} – 1 $$
$$ ReLU(x) = \max(0, x) $$
Classic CNN Architectures
Several classic CNN architectures have played important roles in the development of deep learning:
LeNet-5: Developed in the 1990s, LeNet-5 was successfully applied to handwritten and printed character recognition. It consists of 8 layers, alternating convolutional and pooling layers, followed by fully connected layers.
AlexNet: AlexNet contains 8 layers, with 5 convolutional layers and 3 fully connected layers, using ReLU activation. It achieved a breakthrough in the ImageNet classification task in 2012, significantly outperforming traditional methods.
VGGNet: VGG networks employ small 3×3 convolutional kernels in deeper architectures, demonstrating that network depth improves performance. VGG networks are known for their simplicity and effectiveness.
Residual Networks and Dense Networks
Residual Networks (ResNet): ResNet introduces residual blocks to alleviate the degradation problem in very deep networks. The residual mapping is expressed as:
$$ F(x) = H(x) – x $$
where x is the input, H(x) is the desired output, and F(x) is the residual mapping. The network learns the residual rather than the complete transformation, making optimization easier.
Dense Networks (DenseNet): DenseNet connects each layer to every other layer in a feed-forward fashion, promoting feature reuse and alleviating vanishing gradients. This dense connectivity significantly reduces the number of parameters while improving information flow.
Proposed Random Connection Network
In this work, I propose a randomly connected network that combines the idea of Dropout with feature reuse. The random network’s feature reuse property allows different random parameter configurations to determine the network architecture and feature repetition rate. Higher random parameter settings lead to deeper networks with increased feature reuse, which reduces error without significantly increasing weight and bias parameters. The proposed solar panel defect detection method leverages multi-scale and multi-level information through random connections, improving detection efficiency while maintaining accuracy.
The proposed random connection network consists of 4 random blocks, 5 convolutional layers, 4 pooling layers, 1 activation layer, and 1 fully connected layer. Batch normalization is added between convolution and activation layers to accelerate training and reduce sensitivity to initialization. LeakyReLU activation functions are used throughout. The overall structure of the proposed network is illustrated in the architecture with random blocks that perform stochastic forward connections.
In the random block model, four convolutional layers w1, w2, w3, w4 are set with corresponding kernels. The random block R contains the number of jumps N and the jump length L. The parameters are defined as:
$$ N_{min} = \lceil \frac{n – i}{2} \rceil, \quad N_{max} = n – i, \quad L_{max} = n – i $$
where n is the number of the last layer, i is the number of the current layer, and ⌈·⌉ denotes the ceiling function. Thus, the random matrix R is mathematically represented as an upper triangular matrix. Each current layer jumps to later layers through the random block mechanism. The output of the RandomNet is expressed as:
$$ P = R \cdot Y + Q $$
where P is the output matrix, R is the random matrix generated by the random function based on the number of jumps N and the number of skip layers L, Y is the dimension transformation matrix, and Q is the convolution transformation matrix. Each layer output can be expressed recursively, where the input is x and the output is y, sequentially processed through convolutional layers w1, w2, etc. From the recursive relationships, the Random Block model connects contextual features by using upper-layer outputs as lower-layer inputs through skip connections. The convolutional layer output features are frequently reused, greatly enhancing the network’s depth and dimension.
Experimental Configuration and Evaluation Metrics
The experiments were implemented in Python 3.8 using PyCharm as the integrated development environment. TensorFlow 2.x was the primary deep learning framework, with NVIDIA GPU acceleration and CUDA for high-performance parallel computing. Hyperparameters included a batch size of 32 and an initial learning rate of 0.001. When the loss stabilized, the learning rate was reduced to 10% of its previous value. Binary cross-entropy was used as the loss function, LeakyReLU as the activation function, and the Adam optimizer for model optimization.
The binary cross-entropy loss is defined as:
$$ Loss = -\frac{1}{N}\sum_{i=1}^{N}\left[ y_i \log(\hat{y}_i) + (1-y_i)\log(1-\hat{y}_i) \right] $$
The LeakyReLU activation function is expressed as:
$$ LeakyReLU(x) = \begin{cases} ax & \text{if } x < 0 \\ x & \text{if } x \geq 0 \end{cases} $$
The Adam optimization algorithm updates network weights as follows:
$$ m_t = \beta_1 m_{t-1} + (1-\beta_1)g_t $$
$$ v_t = \beta_2 v_{t-1} + (1-\beta_2)g_t^2 $$
$$ \hat{m}_t = \frac{m_t}{1-\beta_1^t} $$
$$ \hat{v}_t = \frac{v_t}{1-\beta_2^t} $$
$$ \theta_t = \theta_{t-1} – \frac{\alpha \hat{m}_t}{\sqrt{\hat{v}_t} + \epsilon} $$
where α = 0.001, β₁ = 0.09, β₂ = 0.009, and ε = 10⁻⁸.
Evaluation metrics include accuracy, precision, recall, and F1 score:
$$ Accuracy = \frac{TP + TN}{TP + FN + FP + FN} $$
$$ Precision = \frac{TP}{TP + FP} $$
$$ Recall = \frac{TP}{TP + FN} $$
$$ F1 = \frac{2 \times Precision \times Recall}{Precision + Recall} $$
where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively.
Results and Discussion
In the first experimental step, five traditional CNN models and the proposed random network model were trained and evaluated on the CIFAR-10 dataset. The comparison results are shown in the following table:
| Model | Accuracy |
|---|---|
| AlexNet | 78.81% |
| VGG11 | 92.61% |
| VGG19 | 93.65% |
| ResNet50 | 94.04% |
| DenseNet50 | 93.93% |
| RandomNet50 | 96.15% |
From these results, the AlexNet network achieved the lowest accuracy because it was originally designed for 227×227 input images, requiring adaptation when applied to the 32×32 CIFAR-10 dataset. The VGG11 and VGG19 networks achieved accuracies of 92.61% and 93.65%, respectively. ResNet50 outperformed DenseNet50 by 0.11%, reaching 94%. Importantly, the proposed RandomNet50 network achieved an accuracy of 96.15%, surpassing all other models. Furthermore, the parameter count of RandomNet50 was smaller than that of ResNet50, demonstrating that neural networks need not be strictly hierarchical — a layer can depend not only on the immediately preceding layer but also on features learned by earlier layers.
In the second experimental step, the three highest-performing models were selected and evaluated on the EL dataset. The results are presented in the table below:
| Algorithm | Accuracy | Loss | Parameters |
|---|---|---|---|
| RandomNet50 | 88.23% | 0.34 | 1.5×10⁶ |
| ResNet50 | 83.15% | 0.41 | 21.3×10⁶ |
| DenseNet50 | 87.46% | 0.35 | 7×10⁶ |
The RandomNet50 network achieved a substantially higher accuracy of 88.23% on the EL dataset, which is 5.08% higher than ResNet50 and 1.23% higher than DenseNet50. The loss of RandomNet50 was close to that of DenseNet50 and significantly lower than that of ResNet50. The EL images contain high grayscale values in the background with substantial interference and noise, which can cause high loss values when the network oversamples background noise. Despite this, the RandomNet50 network demonstrated strong performance, starting with a relatively high initial loss that decreased rapidly.
In the third experiment, the impact of different network structures on the random network was investigated. Three different configurations of convolutional blocks, with varying distribution of layers per block, were compared. The results are summarized below:
| Block Name | Activations-1 | Activations-2 | Activations-3 |
|---|---|---|---|
| Block 1 | 3×3, 64; 1×1, 64; 3×3, 64; 1×1, 256 | 1×1, 64; 3×3, 64; 1×1, 256 | 1×1, 64; 3×3, 64; 3×3, 64; 1×1, 256 |
| Block 2 | 1×1, 128; 3×3, 128; 3×3, 128; 1×1, 512 | 1×1, 128; 3×3, 128; 3×3, 128; 1×1, 512 | 1×1, 128; 3×3, 128; 3×3, 128; 1×1, 512 |
| Block 3 | 1×1, 256; 3×3, 256; 3×3, 256; 1×1, 1024 | 1×1, 256; 3×3, 256; 3×3, 256; 1×1, 1024 | 1×1, 256; 3×3, 256; 3×3, 256; 1×1, 1024 |
| Block 4 | 1×1, 512; 3×3, 512; 3×3, 512; 1×1, 2048 | 1×1, 512; 3×3, 512; 1×1, 2048 | 1×1, 512; 3×3, 512; 3×3, 512; 1×1, 2048 |
The three network structures yielded almost identical accuracy, indicating that the proposed network architecture is stable and robust. This result also confirms that random blocks can be inserted in the middle portion of the network architecture to improve reliability and accuracy.
Edge Feature Extraction Based on Dual-Channel Attention Mechanism
Edge detection is a fundamental step in image processing. In practice, obtaining sufficiently accurate edges from complex images remains challenging, particularly when edge curves are disconnected, edge segments are missing, or false edges appear. Traditional edge detection methods, such as Sobel, Prewitt, and Canny operators, rely on first- or second-order gradient information. While these methods offer good real-time performance, they are easily affected by noise and struggle to meet modern application requirements. Other approaches based on wavelet transforms, mathematical morphology, genetic algorithms, fuzzy mathematics, and combinatorial optimization have also been explored.
With advances in deep learning, CNNs have become powerful tools for improving edge detection performance. HED (Holistically-Nested Edge Detection) and RCF (Rich Convolutional Features) are two notable deep learning edge detection algorithms.
HED Network
The HED algorithm treats image edge detection as a pixel-level classification problem. It is based on the VGG16 network architecture with five convolutional groups. The HED algorithm modifies VGG16 in two ways: first, side output layers are connected to the last convolutional layer of each group to obtain edge detection results; second, the fully connected layers are removed to reduce network parameters and computational complexity. HED enables image-to-image training by computing loss functions on all pixels. The fused layer combines side outputs from different scales, allowing multi-scale learning. However, deep features gradually lose resolution, which can degrade edge detail quality.
RCF Network
The RCF network improves upon HED by collecting and integrating features from all convolutional layers, yielding richer multiscale feature representations. RCF removes one pooling layer and the fully connected layers, similar to HED, and adds 1×1 convolutional layers after each VGG16 layer to accumulate features. Deconvolution layers are added for upsampling, and loss/sigmoid layers compute losses at each stage. By connecting all stages, RCF achieves higher-quality edge detection compared with HED, though it demands more computational resources and may still struggle with complex scenes involving occlusion and noise.
Proposed Dual-Channel Multi-Scale Attention Network
I propose a novel dual-channel multi-scale attention mechanism for solar panel crack detection. The method leverages multi-scale and multi-level information to comprehensively predict low-level edges, boundaries, and object contours. The network backbone is divided into two branches: a semantic branch and a spatial detail branch. The spatial detail branch consists of three convolutional blocks, capturing low-level features and shallow spatial details. The semantic branch consists of five deep convolutional blocks, using a fast downsampling strategy to expand the receptive field and extract high-level semantic information. A multi-scale attention mechanism module is embedded in both branches to redistribute feature weights by transforming channel, height, and width dimensions for complementary fusion and multi-level utilization. Additionally, a multi-channel hole fusion module is incorporated into the semantic branch, applying hole fusion to the side outputs of the last two convolutional blocks.
The spatial and semantic branch structures are summarized below:
| Layer | Spatial Branch | Semantic Branch |
|---|---|---|
| S1 | Conv Block 3×3, 64, stride 2; Conv Block 3×3, 64 | Deep Block 3×3, 16, stride 2; Deep Block 3×3, 16 |
| S2 | Conv Block 3×3, 128, stride 2; Conv Block 3×3, 128 × 2 | Deep Block 3×3, 32, stride 2; Deep Block 3×3, 32 |
| S3 | Conv Block 3×3, 256, stride 2; Conv Block 3×3, 256 × 2 | Deep Block 3×3, 64, stride 2; Deep Block 3×3, 64 |
| S4 | — | Deep Block 3×3, 128, stride 2; Deep Block 3×3, 64 × 2 |
| S5 | — | Deep Block 3×3, 256, stride 2; Deep Block 3×3, 256 × 2 |
The backbone uses residual convolutional blocks, which were updated into deep convolutional blocks by adding extra convolutional layers on the residual path, thereby increasing convolution depth and improving gradient flow.
Multi-Scale Attention Mechanism
To further extract deep features and improve recognition accuracy, I introduce a multi-scale attention mechanism (MCA). This module is built upon the coordinate attention idea and captures cross-channel information while also sensing directional and positional information. The MCA module transforms the input feature map dimensions and applies multi-scale learning. The attention weight computation is expressed as:
$$ M_S = \sigma(\text{conv}([F_{1max}; F_{1avg}])) \oplus \sigma(\text{conv}([F_{2max}; F_{2avg}]))^T \oplus \sigma(\text{conv}([F_{3max}; F_{3avg}]))^T $$
where σ denotes the sigmoid activation function, ⊕ denotes the concatenation operation, conv denotes convolution, F represents the max-pooled and average-pooled features at each dimension, and the superscript T denotes transpose. By employing parallel branches, each with different receptive fields through different kernel sizes, the resulting attention maps are multi-scale. This approach constructs correlations between channels and spatial pixels, capturing subtle differences in the network and significantly increasing the depth and accuracy of the attention mechanism.
Multi-Channel Hole Fusion Module
In neural networks, continuous downsampling reduces the spatial resolution of feature maps, and excessive compression makes it difficult for subsequent upsampling or deconvolution to restore the original image size. Dilated convolution offers an effective solution by expanding the receptive field without pooling and with the same computational cost. However, the sparse sampling points in dilated convolution tend to ignore small-scale details. To address this, I propose a multi-channel hole fusion module that combines semantic features from different convolutional layers. By using dilated convolutions with different dilation rates and global pooling, multi-scale feature extraction is achieved on two layers of the backbone network, generating four different scales of detailed information while increasing feature dimensions and improving output feature map quality.
Experiments on Edge Detection
To fully evaluate the proposed dual-channel multi-scale attention network model, comparative experiments were conducted on the same solar panel dataset, independently training three networks (HED, FCN, RCF) and comparing their performance. Ablation experiments were also performed to assess the contribution of the multi-scale attention mechanism and the hole fusion module.
The comparative results on the EL dataset are as follows:
| Method | Precision | Recall | F1 Score |
|---|---|---|---|
| HED | 66.51% | 76.96% | 73.35% |
| RCF | 67.80% | 87.73% | 75.81% |
| FCN | 68.22% | 81.68% | 74.64% |
| Our Method | 69.15% | 84.92% | 76.18% |
All network models are evaluated using precision, recall, and F1 score. The precision rate is calculated as the ratio of true positives to the total predicted positives, while the recall rate is the ratio of true positives to all actual positives. The F1 score is the harmonic mean of precision and recall, providing a balanced assessment of classification performance.
From the experimental results, the HED algorithm achieves lower precision and recall due to the reduction in feature resolution with deeper network layers, which leads to the loss of edge details. RCF improves upon this by fusing features from different layers, yielding better performance than HED. However, FCN ignores the spatial relationships between pixels, lacking spatial consistency. The proposed multi-scale attention network considers coordinate spatial characteristics, capturing cross-channel and positional information. The F1 score reached 76.18%, which represents an improvement of 2.83%, 0.37%, and 1.54% over HED, RCF, and FCN, respectively, demonstrating superior solar panel crack feature extraction capability and robustness.
Ablation Experiments
Ablation experiments were designed on the EL dataset with consistent parameter settings, learning rates, and loss functions. The experiments were divided into three groups: the first group used the dual-channel network without the multi-scale attention mechanism or multi-channel hole fusion module; the second group added the multi-channel hole fusion module; the third group further incorporated the multi-scale attention mechanism. The specific results are presented below:
| Experiment | MCA | Hole Fusion | Precision | Recall | F1 Score |
|---|---|---|---|---|---|
| 1 | No | No | 64.43% | 78.05% | 70.81% |
| 2 | Yes | No | 66.56% | 80.14% | 72.54% |
| 3 | Yes | Yes | 69.15% | 84.92% | 76.18% |
Comparing experiments 1 and 2, applying hole fusion improved precision by 2.13%, recall by 2.09%, and the F1 score by 1.73%. Comparing experiments 2 and 3, adding the multi-scale attention mechanism improved precision by 2.59%, recall by 4.78%, and the balanced F1 score by 3.64%. Experiment 3 achieved the best performance across all metrics. In the detection images, experiments 1 and 2 produced discontinuous crack detections with missing segments and some background regions incorrectly classified as cracks, with experiment 1 showing thicker overlapping crack segments. In experiment 3, the detected cracks were continuous and smoother, with better background-crack separation. These results demonstrate that both the multi-scale attention mechanism and the hole fusion module positively contribute to the model’s performance, and their combination achieves optimal detection results.
Conclusion and Future Prospects
Image segmentation technology is widely applied across various industries, and the quality of segmentation results directly affects subsequent processing outcomes. Due to environmental factors such as illumination and clarity, image segmentation becomes more challenging in different conditions. In this work, I have addressed solar panel defect detection through two main approaches: classification recognition of electroluminescent images and edge detection of defective regions. These methods effectively reduce manual labor costs, save operational time, and improve overall efficiency.
The main contributions of this research are summarized in three aspects:
First, following an extensive review of the background and significance of solar panel defect detection, I identified remaining challenges in the field. I constructed and enhanced a solar panel dataset by detailing common defect types, applying image preprocessing to improve image quality, and using geometric transformation and generative adversarial network methods to increase both the quantity and quality of the dataset.
Second, I proposed a random connection network for solar panel defect classification. This network combines the idea of dropout with feature reuse, allowing different random parameter settings to determine network architecture and feature repetition rate. The proposed method achieves a classification accuracy of 96.15% on CIFAR-10 and 88.23% on the EL dataset, outperforming residual and dense networks while requiring fewer parameters. Experimental results confirm that the random connection network offers high classification accuracy for solar panel defect recognition.
Third, I presented a dual-channel multi-scale attention mechanism for solar panel crack detection. The dual-channel backbone network includes separate semantic and spatial detail branches. A multi-scale attention mechanism is fused into the backbone to redistribute feature weights through transformations of channel, height, and width dimensions, enabling complementary fusion and multi-level utilization. A hole fusion module is also integrated, which significantly improves information transfer and feature extraction capability while reducing network noise interference. The proposed algorithm outperforms classical edge detection algorithms, yielding richer crack details and improved segmentation metrics.
Looking forward, several promising research directions remain. First, the proposed methods could be extended to address multi-class classification of different solar panel defect types with finer granularity. Second, the algorithms could be deployed in actual engineering scenarios to enhance the intelligent management level of solar power station inspection systems, thereby improving operational efficiency. Third, as neural network technology continues to advance with improved accuracy, future defect detection approaches will likely transition from image-based to video-based detection using even more intelligent algorithms. I intend to focus on improving network performance in terms of speed and parameter reduction while maintaining accurate defect recognition across various application scenarios.
In conclusion, the research presented in this thesis provides effective deep learning-based solutions for solar panel defect detection, contributing to the safe and efficient operation of solar power generation systems. The findings have both theoretical and practical significance for improving power generation efficiency, reducing equipment defects, and lowering operation and maintenance costs in the photovoltaic industry.