In this thesis, I investigate the image recognition techniques applied to the inspection of solar panels in photovoltaic power stations. I focus on two major problems: dust accumulation and hot spots on the surface of solar panels. Outdoor solar panels inevitably suffer from dust deposition, which reduces the power generation efficiency and, if not cleaned in time, leads to further failures and shorter service life. Meanwhile, hot spots caused by partial shading or dust can generate enough heat to damage the encapsulation materials of solar panels, accelerate aging, and even create safety hazards. Conventional image processing algorithms used in intelligent operation and maintenance inspection systems often suffer from low accuracy, poor illumination robustness, and slow recognition speed. To address these challenges, I propose several algorithms based on the specific characteristics of solar panel images. In this work, I combine traditional image processing with deep learning methods to improve the accuracy and robustness of dust recognition and hot spot detection. Extensive experiments demonstrate the effectiveness of the proposed methods.
1. Introduction
Energy shortage has become one of the major obstacles to human technological progress. Traditional power generation heavily relies on fossil fuels such as coal and oil, which are non-renewable and cause severe environmental pollution. In recent years, clean and renewable energy sources, including solar energy, have gained increasing attention. Solar power, in particular, has experienced rapid growth due to the declining cost of solar panels and their improving photoelectric conversion efficiency. According to the national statistics at the end of 2021, the installed capacity of photovoltaic power in China reached 306,560 MW, with a year-on-year growth of 20.9%, which is the highest among all power generation types. The growth of thermal power was only 4.1%, mainly because of its heavy pollution and high resource consumption. The following table summarizes the installed capacity and growth rates of different power generation types.
| Power Generation Type | Installed Capacity / MW | Year-on-Year Growth / % |
|---|---|---|
| Thermal power | 1,296,780 | 4.1 |
| Hydropower | 390,920 | 5.6 |
| Nuclear power | 53,260 | 6.8 |
| Wind power | 328,480 | 16.6 |
| Photovoltaic power | 306,560 | 20.9 |
The development of solar power is of strategic importance in China. Many photovoltaic power stations have been constructed in the western regions, where abundant sunlight is available. However, these areas often suffer from harsh weather conditions such as sandstorms, heavy snow, and strong winds. As a result, solar panels installed outdoors are easily covered by dust, which significantly reduces their power output. Dust also induces hot spots, which are local high-temperature regions that can permanently damage the solar panel. Therefore, efficient and accurate inspection methods are essential for the safe and stable operation of photovoltaic power stations.
2. Photovoltaic Power Station and Intelligent Inspection
A photovoltaic power station mainly consists of solar panels, mounting structures, inverters, combiner boxes, and monitoring systems. The solar panel is the core part that converts sunlight into electricity. The photoelectric effect in semiconductor materials such as monocrystalline or polycrystalline silicon is responsible for this conversion. When sunlight irradiates the semiconductor surface, photons are absorbed, creating electron-hole pairs in the P-N junction. The movement of these carriers generates a voltage across the junction. The principle can be illustrated by the following equation:
$$ I = I_{\text{ph}} – I_0 \left[ \exp\left(\frac{V + I R_s}{n V_t}\right) – 1 \right] – \frac{V + I R_s}{R_{sh}} $$
where \(I_{\text{ph}}\) is the photogenerated current, \(I_0\) is the reverse saturation current, \(R_s\) is the series resistance, \(R_{sh}\) is the shunt resistance, \(n\) is the ideality factor, and \(V_t\) is the thermal voltage.
Photovoltaic arrays are composed of many solar panels connected in series and parallel. Depending on the scale, solar power stations can be divided into centralized and distributed types. Centralized stations usually have a capacity above 6 MW, with long strings of solar panels. Distributed stations have smaller capacities and are often installed on rooftops. The maintenance of these stations traditionally relies on manual inspection, which is labor-intensive, inefficient, and dangerous. In modern intelligent inspection, unmanned aerial vehicles (UAVs) equipped with high-resolution RGB and infrared cameras are used to collect images of solar panels. These images are then transmitted to a ground station for analysis. Intelligent algorithms process the images to identify faults such as dust accumulation, hot spots, cracks, and other anomalies. The UAV inspection platform integrates GPS positioning, path planning, image acquisition, and real-time data transmission.
One of the main challenges in solar panel inspection is the variation of lighting conditions. Images captured outdoors are affected by the angle and intensity of sunlight, which may cause uneven illumination or glare on the glass surface of solar panels. These effects can significantly degrade the performance of image recognition algorithms. Therefore, robust algorithms that are insensitive to illumination changes are highly desirable.
3. Dust Recognition on Solar Panels Based on Image Segmentation
3.1 Characteristics of Dust
Dust is a mixture of fine solid particles with different sizes and chemical compositions. Inorganic dust mainly contains silicon dioxide, aluminum oxide, and calcium oxide. The deposition of dust on solar panels follows Stokes’ law. The settling velocity of a spherical particle can be expressed as:
$$ v = \frac{2}{9} \frac{(\rho_s – \rho) g r^2}{\mu} $$
where \(\rho_s\) is the particle density, \(\rho\) is the fluid density, \(g\) is the gravitational acceleration, \(r\) is the particle radius, and \(\mu\) is the dynamic viscosity of the fluid. Dust particles in the natural environment often have irregular shapes and are charged. They strongly adhere to the solar panel surface due to electrostatic forces. The accumulation of dust reduces the transmittance of sunlight and decreases the power output. Dust deposition can be classified into dry loose dust, spot-like dust, and salt-like dust. The latter two are usually caused by rainfall after long-term dust accumulation. Early detection and cleaning of dry loose dust can prevent more serious problems.
3.2 Color Characteristics in YCbCr Space
In order to separate dust from the background of the solar panel, I analyze the color properties of dust images. The original RGB images are converted to the YCbCr color space, where Y represents luminance and Cb, Cr represent color differences. I observed that in the YCbCr space, the dust region has a higher Y value, a lower Cb value, and a higher Cr value compared to the clean solar panel background. Based on this observation, I design a segmentation rule:
$$ R(x,y) = \begin{cases} I(x,y), & \text{if } Y(x,y) > Y_{mean} \text{ and } Cb(x,y) < Cb_{mean} \\ 0, & \text{otherwise} \end{cases} $$
However, global thresholds cannot handle the situation where dust is non-uniformly distributed on the solar panel. To adapt to local variations, I propose an adaptive threshold strategy. Instead of using a fixed global mean of the Cb channel, I compute the threshold for each row based on the previous row’s valid pixels. The adaptive threshold for the \(i\)-th row is:
$$ Cb_{mean}^{i} = \frac{1}{M} \sum_{j=1}^{M} Cb(i-1,j) + Cb_{mean} – Cb_{mean}^{i-1} $$
where \(M\) is the number of valid pixels in the previous row, and \(Cb_{mean}\) is the global average of the Cb channel. The first row threshold is initialized to the global mean. This adaptive method improves the segmentation result for non-uniform dust distribution. The segmentation flow chart is as follows: input image, convert to YCbCr, set adaptive thresholds, compare each pixel, and output the segmented image.
After segmentation, dust areas are preserved while the solar panel background is set to black as shown in the experimental results. The segmented image is then used for feature extraction.
3.3 Texture Feature Extraction Using LBP
Texture features are important for recognizing dust states. The Local Binary Pattern (LBP) operator is a powerful texture descriptor with good illumination invariance and low computational complexity. The original LBP operator can be defined as:
$$ LBP_{P,R} = \sum_{i=0}^{P-1} s(g_i – g_c) 2^i $$
where \(g_c\) is the value of the center pixel, \(g_i\) are the values of neighboring pixels, and \(s(x)\) is the threshold function:
$$ s(x) = \begin{cases} 1, & x \ge 0 \\ 0, & \text{otherwise} \end{cases} $$
To achieve rotation invariance and reduce the number of patterns, the uniform rotation-invariant LBP is used:
$$ LBP_{P,R}^{riu2} = \begin{cases} \sum_{i=0}^{P-1} s(g_i – g_c), & \text{if } U(LBP_{P,R}) \le 2 \\ P+1, & \text{otherwise} \end{cases} $$
where the uniformity measure \(U\) is defined as:
$$ U(LBP_{P,R}) = \left| s(g_{P-1} – g_c) – s(g_0 – g_c) \right| + \sum_{i=1}^{P-1} \left| s(g_i – g_c) – s(g_{i-1} – g_c) \right| $$
The uniform mode reduces the number of feature dimensions from 256 to 9 for a 3×3 neighborhood, significantly improving computational efficiency while preserving discriminative information. In this work, I use the LBP operator with radius \(R=1\) and \(P=8\) neighbors in the uniform rotation-invariant mode.
3.4 Recognition Process
The recognition process for dust state on solar panels consists of the following steps:
- Preprocess the captured image by filtering noise and resizing all images to a fixed size.
- Convert the image from RGB to YCbCr color space.
- Segment the dust regions using the adaptive threshold rule.
- Extract LBP features from the segmented image.
- Compute the histogram of the LBP image as the feature vector.
- Feed the feature vector into a Softmax classifier to determine the dust density level.
3.5 Dataset and Experimental Analysis
I built a dataset by simulating natural dust deposition in the laboratory. Dust samples were collected from the outdoor environment in Urumqi, China. The particle size distribution was analyzed using a standard sieve. The following table shows the particle size distribution (in percentage) for several districts.
| District | 2.5–5 μm | 5–10 μm | 10–50 μm | 50–100 μm | >100 μm |
|---|---|---|---|---|---|
| Xinshi | 4.08 | 6.42 | 50.82 | 16.05 | 22.63 |
| Midong | 4.39 | 6.97 | 52.16 | 17.58 | 20.90 |
| Tianshan | 3.79 | 6.07 | 50.65 | 15.95 | 23.54 |
Particles below 100 μm were selected as dust samples. A monocrystalline silicon solar panel with dimensions 1580 mm × 810 mm and a maximum power of 200 W was used. Different dust densities were uniformly sprayed onto the solar panel using a simulated natural deposition process. I set five dust density levels: 0 g/m², 5 g/m², 10 g/m², 15 g/m², and 20 g/m². For each level, RGB images were captured. The total number of images was 2000, with 1400 for training and 600 for testing. All images were resized to 300 × 400 pixels.
I evaluated different LBP configurations to select the best one. The results are shown below.
| LBP Type | Feature Dimension | Accuracy / % | Time / s |
|---|---|---|---|
| Rotation invariant (R=1, P=8) | 324 | 74.23 | 70.73 |
| Rotation invariant (R=2, P=8) | 324 | 83.56 | 62.92 |
| Uniform rotation invariant (R=1, P=8) | 90 | 88.56 | 70.38 |
| Uniform rotation invariant (R=2, P=8) | 90 | 82.61 | 61.33 |
The uniform rotation-invariant mode with R=1, P=8 achieved the highest accuracy of 88.56% and a relatively short recognition time. I compared my proposed algorithm with other methods: the gray-level co-occurrence matrix (GLCM) and ResNet50. The comparison is presented in the following table.
| Algorithm | Accuracy / % | Recognition Time / s |
|---|---|---|
| Proposed adaptive segmentation + LBP | 88.56 | 70.38 |
| GLCM | 86.73 | 138.09 |
| ResNet50 | 82.53 | 615.23 |
The proposed algorithm outperforms both GLCM and ResNet50 in terms of accuracy and speed. It has a simple structure and strong real-time performance, making it suitable for embedded inspection systems. However, its accuracy is still insufficient for some high-precision applications, which motivates the use of deep learning methods.
4. Deep Learning for Dust Recognition on Solar Panels
4.1 Overview of Deep Learning
Deep learning, especially convolutional neural networks (CNNs), has achieved remarkable success in many image processing tasks. A CNN typically consists of convolutional layers, pooling layers, activation functions, and fully connected layers. The convolution operation is defined as:
$$ y(p_0) = \sum_{p_n \in R} w(p_n) \cdot x(p_0 + p_n) $$
where \(x\) is the input feature map, \(w\) is the convolution kernel, \(R\) is the receptive field, and \(y\) is the output feature map. CNNs can automatically extract hierarchical features from images, enabling high-level semantic understanding.
However, training CNNs requires large datasets. In many industrial applications, such as solar panel inspection, collecting thousands of labeled images is challenging. Moreover, outdoor illumination variations often degrade the performance of CNN models. In this section, I address these two problems by improving the DenseNet-121 architecture.
4.2 Dense Convolutional Network (DenseNet)
DenseNet connects each layer to every other layer in a dense block. The output of the \(l\)-th layer is:
$$ x_l = H_l([x_0, x_1, \dots, x_{l-1}]) $$
where \([x_0, x_1, \dots, x_{l-1}]\) denotes the concatenation of feature maps from all preceding layers, and \(H_l\) is a composite function including batch normalization, ReLU activation, and convolution. DenseNet alleviates the vanishing gradient problem, encourages feature reuse, and has fewer parameters than many other deep networks. DenseNet-121 has four dense blocks with growth rate \(k=32\). Each block consists of multiple bottleneck layers: a 1×1 convolution followed by a 3×3 convolution. Transition layers between blocks reduce the spatial dimension.
4.3 Parameter-Free Attention Module
To enhance the feature extraction capability of DenseNet for small sample datasets, I introduce a simple, parameter-free attention mechanism called SimAM. It estimates the importance of each neuron based on the energy function:
$$ e(w_t, b_t) = \frac{1}{M-1} \sum_{i=1}^{M} \left(1 – (\hat{t} – x_i)\right)^2 + \left(1 – (\hat{t} – t)\right)^2 + \lambda \hat{t}^2 $$
where \(t\) is the target neuron, \(x_i\) are other neurons in the same channel, \(M = H \times W\) is the number of neurons, and \(\lambda\) is a weighting factor. The closed-form solution yields a weight matrix:
$$ E = \frac{4(\hat{\sigma}^2 + \lambda)}{(\hat{t} – \hat{\mu})^2 + 2\hat{\sigma}^2 + 2\lambda} $$
where \(\hat{\mu}\) and \(\hat{\sigma}^2\) are the mean and variance of all neurons in the channel. Then the attention-enhanced feature is obtained by:
$$ \tilde{X} = X \cdot \text{sigmoid}\left(\frac{1}{E}\right) $$
Since SimAM does not introduce any additional parameters, the network complexity remains unchanged while the representation ability is improved.
4.4 Central Difference Convolution
Standard convolution is sensitive to illumination changes. To improve the illumination robustness, I replace the standard convolution in DenseNet with central difference convolution (CDC). The CDC operation is formulated as:
$$ y(p_0) = \sum_{p_n \in R} w(p_n) \cdot \left[ x(p_0 + p_n) – x(p_0) \right] \cdot \theta + \sum_{p_n \in R} w(p_n) \cdot x(p_0 + p_n) \cdot (1 – \theta) $$
where \(\theta\) is a hyperparameter controlling the contribution of the gradient information. When \(\theta = 0\), CDC reduces to vanilla convolution. In this work, I set \(\theta\) to a positive value to incorporate the central difference information, which is analogous to the LBP operation and helps to extract gradient features that are less affected by illumination.
4.5 Improved Network Architecture
I integrate SimAM into each dense block of DenseNet-121 and replace all standard convolutions with CDC. The improved network structure is:
- Input layer: 300 × 400 RGB image.
- Initial convolution and pooling.
- Dense Block 1: 6 layers of CDC bottleneck.
- Transition Layer 1.
- Dense Block 2: 12 layers of CDC bottleneck.
- Transition Layer 2.
- Dense Block 3: 24 layers of CDC bottleneck.
- Transition Layer 3.
- Dense Block 4: 16 layers of CDC bottleneck.
- Global average pooling, fully connected layer, and Softmax classifier.
Each dense block is followed by the SimAM module to reweight the feature maps. The loss function is the cross-entropy loss:
$$ L = -\sum_{i} y_i \log(\hat{y}_i) $$
where \(y_i\) is the true label distribution and \(\hat{y}_i\) is the predicted probability. I use the Adam optimizer, which combines the advantages of AdaGrad and RMSProp. The Adam update rule is:
$$ m_t = \beta_1 m_{t-1} + (1 – \beta_1) \nabla f(\theta_t) $$
$$ v_t = \beta_2 v_{t-1} + (1 – \beta_2) (\nabla f(\theta_t))^2 $$
$$ \hat{m}_t = \frac{m_t}{1 – \beta_1^t}, \quad \hat{v}_t = \frac{v_t}{1 – \beta_2^t} $$
$$ \theta_{t+1} = \theta_t – \frac{\alpha}{\sqrt{\hat{v}_t} + \epsilon} \hat{m}_t $$
where \(\alpha\) is the learning rate, \(\beta_1 = 0.9\), \(\beta_2 = 0.999\), and \(\epsilon = 10^{-8}\).
4.6 Data Augmentation
To alleviate the small-sample problem, I use data augmentation techniques to expand the dataset. The augmentation operations include:
- Geometric transformations: horizontal flip, vertical flip, rotation by various angles, scaling, and cropping.
- Intensity transformations: brightness adjustment, contrast adjustment, and sharpening.
- Noise addition: positive salt-and-pepper noise.
Through offline augmentation, the original 2000 images were expanded to 20,000 images. Among them, 14,000 were used for training and the rest for testing.
4.7 Experimental Results
I first compared the improved network with the original DenseNet-121 on the enhanced dataset. The experimental settings were: 100 epochs, learning rate 0.1, growth rate \(k=32\), and \(\lambda = 10^{-4}\). The performance metrics are presented below.
| Model | Params | FLOPs | Accuracy | Loss | EER | TPR |
|---|---|---|---|---|---|---|
| DenseNet-121 | 7.351M | 2.524G | 0.9273 | 0.1082 | 1.0115 | 0.6424 |
| Improved model | 7.351M | 2.524G | 0.9501 | 0.0974 | 0.9057 | 0.8156 |
The improved model achieves an accuracy of 95.01%, which is much higher than the original DenseNet-121 (92.73%). The parameters and FLOPs remain the same because SimAM is parameter-free. The central difference convolution also does not increase the parameter count. This demonstrates that the improvements significantly enhance the feature extraction and illumination robustness.
I also compared the improved model with VGG19 and ResNet50. The accuracy and loss curves over training epochs are shown in the following figures. The improved model converges faster and achieves higher accuracy than the other two networks. The receiver operating characteristic (ROC) curves also confirm the superiority of the proposed model, as the curve is closer to the top-left corner.
These results indicate that the improved DenseNet model is well-suited for solar panel dust recognition under small-sample and uneven illumination conditions.
5. Hot Spot Detection on Solar Panels
5.1 Hot Spot Formation
Hot spots are localized high-temperature areas on solar panels caused by partial shading, dust, or cell damage. When a solar cell is shaded, it stops generating electricity and acts as a load that dissipates power from other cells. The dissipated power is converted into heat, increasing the temperature of the shaded cell. This may lead to solder joint failure, encapsulation degradation, or even fire. The heat generation can be modeled by the power dissipation in the shaded cell:
$$ P_{\text{hotspot}} = I^2 \times R_{\text{shunt}} $$
where \(I\) is the current flowing through the shaded cell and \(R_{\text{shunt}}\) is the shunt resistance. Hot spots are usually invisible in ordinary RGB images but are clearly visible in infrared (IR) images, where the hot regions appear as bright spots.
5.2 Traditional Detection Methods
Template Matching
Template matching is a classical object detection method. It compares the input image with a predefined template using similarity functions such as normalized cross-correlation:
$$ R(x,y) = \frac{\sum_{x’,y’} T(x’,y’) \cdot I(x+x’, y+y’)}{\sqrt{\sum_{x’,y’} T(x’,y’)^2 \cdot \sum_{x’,y’} I(x+x’, y+y’)^2}} $$
Template matching is simple but sensitive to image rotation, scaling, and illumination changes. In UAV inspection, the image perspective and size vary, so template matching often fails to detect small hot spots.
Edge Detection
Edge detection methods, such as the Canny operator, are commonly used to extract boundaries of hot spots in IR images. The Canny algorithm performs Gaussian filtering, gradient calculation, non-maximum suppression, and double-threshold detection. The gradient magnitude is computed as:
$$ G = \sqrt{G_x^2 + G_y^2} $$
where \(G_x\) and \(G_y\) are the image gradients along the \(x\)- and \(y\)-axes. Edge detection works well for clear hot spot boundaries, but it can produce many false edges due to background noise, reflection, and uneven thermal distribution.
5.3 SSD Network for Hot Spot Detection
To overcome the limitations of traditional methods and to detect small hot spots in high-altitude UAV images, I apply the Single Shot Multibox Detector (SSD). SSD is a deep learning object detection algorithm that predicts bounding boxes and class probabilities directly from feature maps. The architecture is based on VGG16 with additional convolutional feature layers. The network predicts detections at multiple scales, making it effective for small objects. The default boxes (anchors) in SSD are defined with different scales and aspect ratios. The scale of the \(k\)-th feature map is computed as:
$$ s_k = s_{\min} + \frac{s_{\max} – s_{\min}}{m – 1} (k-1), \quad k \in [1, m] $$
where \(m\) is the number of feature maps, \(s_{\min} = 0.2\) and \(s_{\max} = 0.9\). The aspect ratio \(a_r\) is selected from \(\{1, 2, 3, \frac{1}{2}, \frac{1}{3}\}\). For each default box, the width and height are computed by:
$$ w_k = s_k \sqrt{a_r}, \quad h_k = s_k / \sqrt{a_r} $$
The total loss of SSD is a weighted sum of localization loss and confidence loss:
$$ L(x, c, l, g) = \frac{1}{N} \left( L_{\text{conf}}(x, c) + \alpha L_{\text{loc}}(x, l, g) \right) $$
where \(N\) is the number of matched default boxes, \(c\) is the predicted class, \(l\) is the predicted bounding box, and \(g\) is the ground truth bounding box. The localization loss is the smooth L1 loss:
$$ L_{\text{loc}}(x, l, g) = \sum_{i \in Pos} \sum_{m \in \{cx, cy, w, h\}} x_{ij}^k \cdot \text{smooth}_{L1}(l_i^m – \hat{g}_j^m) $$
The confidence loss is the softmax loss over multiple classes:
$$ L_{\text{conf}}(x, c) = – \sum_{i \in Pos} x_{ij}^p \log(\hat{c}_i^p) – \sum_{i \in Neg} \log(\hat{c}_i^0), \quad \hat{c}_i^p = \frac{\exp(c_i^p)}{\sum_p \exp(c_i^p)} $$
5.4 Dataset and Experiment
I built a hot spot dataset using an infrared camera at a photovoltaic power station. A total of 1500 IR images were collected. The images contained solar panels and hot spots. I manually annotated each image with bounding boxes using LabelImg. Two classes were defined: “board” for solar panels and “hot_spot” for hot spots. The dataset was split into 1000 training images and 500 testing images. All images were resized to 300×300 pixels.
The SSD network was trained with a learning rate of \(10^{-4}\) for 150 epochs. I compared SSD with Faster-RCNN. The evaluation metrics include precision, recall, and average precision (AP). The results are shown below.
| Network | Precision | Recall | AP |
|---|---|---|---|
| Faster-RCNN | 0.8723 | 0.8544 | 0.8361 |
| SSD | 0.8965 | 0.8831 | 0.8593 |
SSD outperforms Faster-RCNN in all metrics. Moreover, SSD is faster and more suitable for real-time inspection. The detection visualization shows that SSD successfully locates the solar panels and hot spots, even when the image contains strong reflections. The algorithm can distinguish hot spots from areas of high reflection on the solar panel surface.
To investigate the small-object detection capability, I evaluated SSD on images captured at different UAV flight heights. When the flight altitude is high, the solar panels appear small in the image, and hot spots are even smaller. SSD still managed to detect most hot spots. In extremely high altitude images, some solar panels may be merged into one bounding box, and some reflections may be misclassified as hot spots. This indicates that there is still room for improvement, but the overall performance is acceptable for practical inspection. In such cases, the UAV flight altitude can be adjusted to obtain clearer images.
6. Conclusion and Outlook
In this thesis, I studied image recognition techniques for solar panel inspection in photovoltaic power stations. I focused on two critical faults: dust accumulation and hot spots. For dust recognition, I proposed an adaptive image segmentation algorithm based on the YCbCr color space, combined with LBP texture feature extraction. This traditional method achieved an accuracy of 88.56% with high speed and low complexity. To further improve accuracy, I developed an improved DenseNet model that integrates a parameter-free attention module and central difference convolution. The improved model achieved 95.01% accuracy on the augmented dataset, demonstrating superior performance under small-sample and uneven illumination conditions. For hot spot detection, I applied the SSD network, which performs well in detecting small hot spots in infrared images. The experimental results confirm that SSD outperforms Faster-RCNN in precision, recall, and average precision.
There are several directions for future work. First, more open datasets of solar panel images are needed to advance deep learning research in this field. Second, the generalization ability of recognition models should be further improved to handle a wider range of environmental conditions. Third, the integration of multiple sensor modalities, such as fusing RGB and infrared images, may lead to more robust and accurate inspection systems. Finally, lightweight network architectures designed for edge computing devices could enable real-time onboard analysis on UAVs, reducing the need for data transmission to ground stations.
In summary, the proposed methods contribute to the intelligent operation and maintenance of photovoltaic power stations, helping to ensure the safe and efficient operation of solar panels. With continuous improvements, image recognition technology will play an increasingly important role in the renewable energy industry.