Enhancing YOLOv5 with Custom Attention Mechanism for Defect Detection in Photovoltaic Power System Thermal Imaging

Existing deep learning models for defect detection in photovoltaic system thermal imaging may face issues such as suboptimal detection performance and limited defect types. Therefore, this paper proposes an enhanced YOLOv5n model based on a one-step strategy, incorporating a custom attention mechanism to improve the focus on critical information. The model features a network structure designed for lightweight processing and integrates custom attention modules and optimization techniques for object detection and post-processing, which improves model accuracy to some extent. Experimental results indicate that the model with the attention module achieved a minimum loss value of $\mathbf{22.225} \%$ during training, significantly lower than the 29.279% achieved without the module. The optimized model attained a standard deviation of 2.4772 on a public dataset, demonstrating higher stability and accuracy, and effectively performing defect detection and real-time monitoring tasks in photovoltaic system thermal imaging.