A dynamically adaptive and high-efficiency small object detection network for infrared thermographic images in online monitoring of solar photovoltaic panel defects

Photovoltaic systems are often exposed to complex environmental conditions, which can significantly affect their performance and lifespan. As the core component in converting solar energy to electrical power, photovoltaic (PV) panels may experience performance degradation due to operational wear and damage. Therefore, real-time monitoring systems are essential for ensuring the efficient operation of solar power systems and preventing potential energy loss. Efficient and reliable online monitoring of photovoltaic panel failures under complex conditions has always been a huge challenge. This study utilizes infrared thermography UAV to capture both surface and internal defect images of solar panels, constructing a dataset containing three common defect types. To mitigate background interference in large-scale images, we propose a Dynamically Adaptive and High-Efficiency Small Object Detection Network in Infrared Thermographic Images, enabling rapid and accurate extraction of multi-scale defects, even the small ones. A super-resolution weight matrix module is added to the network's frontend to enhance infrared thermography image resolution, filtering spatial and channel dimensions to extract defect regions and reduce computational complexity. A parallel efficient cross-channel fusion module refines multi-scale defect features, preventing information loss and enhancing feature extraction across scales. For small pitting defects, a dynamic local morphological dilation module is introduced to improve detection accuracy and reduce false positives. Lastly, a dynamic region proposal network adjusts the classifier and regressor based on prior detection results, improving overall detection accuracy. The model achieves a mAP of 98.1 %, outperforming others in speed and complexity, making it ideal for online defect detection applications.