Abstract
Photovoltaic (PV) power generation, as a clean energy source, has promoted the transformation of the energy structure, making energy more diversified and sustainable. PV power stations are mainly distributed in open and remote areas. Due to the long-term exposure of solar panels to the outdoors, prolonged erosion from wind, sand, hail, snow, and other severe weather conditions can damage the surface of the solar panels. If not repaired in time, this can lead to the hot spot effect, which can damage panels lightly and cause fires severely. PV power stations now widely adopt unmanned aerial vehicles (UAVs) and thermal imaging technology to capture infrared images of solar panels and detect hot spots, ensuring timely repairs. However, existing hot spot detection algorithms may miss or falsely detect hot spots when applied to infrared images of solar panels with complex backgrounds. Therefore, this paper addresses this issue by studying a hot spot detection method for solar panels based on enhanced infrared images obtained through aerial photography.
Keywords: solar panel; infrared image of solar panels; image enhancement; solar panel detection; hot spot detection
1. Introduction
1.1 Background and Significance
The significance of this research lies in enhancing the accuracy and efficiency of hot spot detection in infrared images of solar panels, which is crucial for the safe and stable operation of PV power stations.
1.2 Research Status and Existing Problems
Many scholars have proposed corresponding solutions to the existing problems of solar panel and hot spot detection in infrared images. Existing methods are mainly divided into traditional image processing methods and deep learning methods. Traditional image processing methods perform well in single application scenarios, with detection results highly matching the target contours and high algorithm execution efficiency. However, these methods have poor generalization performance and cannot detect solar panels in special scenarios. In addition, when applied to hot spot detection, traditional image processing methods often struggle to accurately distinguish noise from valid features, resulting in the filtration of some key information and affecting hot spot detection results. Deep learning methods have better generalization performance and can achieve real-time detection after training. However, current networks struggle to effectively fuse multi-scale features, leading to suboptimal detection effects.
1.3 Main Research Content
The main research content of this paper is summarized as follows: designing an infrared image enhancement algorithm for solar panels based on improved homomorphic filtering and Contrast Limited Adaptive Histogram Equalization (CLAHE); designing a solar panel detection algorithm based on CFG-YOLO; designing a hot spot detection algorithm for solar panels based on RCE-YOLO.
2. Infrared Image Enhancement Algorithm for Solar Panels Based on Improved Homomorphic Filtering and CLAHE
2.1 Introduction
Infrared thermography is widely used in the PV field, reducing accident probabilities and losses at PV power stations. However, solar panels are susceptible to external factors such as infrared camera performance and extreme weather during imaging, leading to captured infrared images with low contrast and illumination, which is detrimental to subsequent solar panel detection. To address this issue, this paper proposes an infrared image enhancement algorithm for solar panels based on improved homomorphic filtering and CLAHE.
2.2 Homomorphic Filtering Algorithm
2.2.1 Theoretical Basis
Homomorphic filtering is a method that combines frequency domain processing and spatial domain processing to separate and enhance the illumination and reflection components of an image.
2.2.2 Selection of Transfer Function
The transfer function is designed to enhance the low-frequency components (illumination) and high-frequency components (reflection details) of the image.
2.3 Infrared Image Enhancement Algorithm for Solar Panels Based on Improved Homomorphic Filtering and CLAHE
2.3.1 Homomorphic Filtering Algorithm Based on Improved Transfer Function
An improved transfer function is constructed, similar in structure to the homomorphic filtering profile, to enhance the brightness and contrast of solar panel infrared images.
2.3.2 CLAHE Algorithm with Gamma Correction
To overcome the deficiencies of faint detail features and uneven brightness distribution in images enhanced by homomorphic filtering, a CLAHE algorithm fused with gamma correction is designed to optimize the image, enhancing its resistance to external interference and improving image clarity.
2.4 Design of Infrared Image Enhancement Algorithm for Solar Panels
The complete algorithm combines the improved homomorphic filtering and CLAHE with gamma correction.
2.5 Experimental Results and Analysis
A comparison experiment is conducted with 150 infrared images of solar panels from different locations to verify the effectiveness of the proposed algorithm. Experimental data demonstrates that the proposed algorithm enhances the comprehensive evaluation value of solar panel infrared images, with obvious visual advantages compared to other algorithms, improving the completeness of manual annotation of solar panel images and enhancing detection accuracy.
Table 1. Comparison of Infrared Image Enhancement Effects
| Algorithm | Brightness Enhancement | Contrast Enhancement | Detail Preservation |
|---|---|---|---|
| Proposed | High | High | Good |
| Algorithm A | Medium | Medium | Fair |
| Algorithm B | Low | Low | Poor |
3. Solar Panel Detection Algorithm Based on CFG-YOLO
3.1 Introduction
After enhancing solar panel infrared images using the proposed algorithm, the accuracy of manual annotation of solar panels improves, but issues of misdetections and missed detections remain unresolved. To address these issues, a solar panel detection algorithm based on improved YOLOX-S, called CFG-YOLO (CA_Focal GIoU-YOLO), is proposed.
3.2 Algorithm Design
3.2.1 Integration of 3CA Attention Module
To effectively improve local representation capability and better fit the feature information of solar panels, the 3CA attention module is integrated into the CSPDarknet backbone feature network.
3.2.2 Use of Focal Loss
To address the imbalance of positive and negative samples in solar panel infrared images, the binary cross-entropy confidence prediction loss is changed to Focal Loss.
3.2.3 Improvement of Localization Loss Function to GIoU Loss
The original IoU Loss in the detection model cannot optimize the case where the predicted box and the ground truth box do not intersect. Therefore, it is improved to GIoU Loss to enhance the regression accuracy of the detection model.
Table 2. Comparison of Solar Panel Detection Accuracy
| Algorithm | Detection Accuracy | Inference Time (ms) |
|---|---|---|
| CFG-YOLO | 99.62% | 23.4 |
| YOLOX-S | 98.35% | 21.7 |
| Faster R-CNN | 97.89% | 34.5 |
4. Hot Spot Detection Algorithm for Solar Panels Based on RCE-YOLO
4.1 Introduction
Traditional image processing methods for hot spot detection are slow, have poor robustness, and need improvement in detection accuracy. Therefore, this paper studies deep learning methods applied to hot spot detection. Existing methods neglect to obtain deep spatial feature information, leading to missed or falsely detected small-scale hot spots. To address these issues, this paper proposes an RCE-YOLO (RFB_CA_EIoU-YOLO) algorithm for solar panel hot spot detection.
4.2 Algorithm Design
4.2.1 Replacement of Spatial Pyramid Pooling with Receptive Field Block (RFB)
To address the high computational cost and time consumption of spatial pyramid pooling in the model’s backbone feature network, it is replaced with the RFB module, which simulates the size of the human receptive field by dynamically adjusting the dilation rate, diversifying and refining the receptive field.
4.2.2 Addition of CA Attention Mechanism
The CA attention mechanism is added before upsampling in the network neck, allowing the algorithm to extract more deep feature information in the width and height directions of hot spots before upsampling, enhancing the algorithm’s sensitivity to hot spot locations and improving detection accuracy.
4.2.3 Replacement of Localization Loss Function with EIoU Loss
The original localization loss function is replaced with the EIoU function, which optimizes the issue of sample imbalance in bounding box regression, making the regression more focused on high-quality anchor boxes and reducing the contribution of error samples to loss regression, thereby improving the regression accuracy and detection accuracy of hot spots.
Table 3. Comparison of Hot Spot Detection Accuracy
| Algorithm | Detection Accuracy | False Detection Rate | Missed Detection Rate |
|---|---|---|---|
| RCE-YOLO | 98.76% | 1.24% | 0.98% |
| YOLOv5 | 96.32% | 2.87% | 2.11% |
| SSD | 94.58% | 4.23% | 3.49% |
5. Conclusion
The solar panels as the research object and conducts research on infrared image enhancement algorithms, solar panel detection algorithms, and hot spot detection algorithms for solar panels.