Infrared Defect Detection for Power System Photovoltaic Panels in UAV Inspection Using Dynamic Convolution

INTRODUCTION: This paper addresses the significant demand for non-contact, efficient, and large-area inspections in photovoltaic power system drone infrared inspections. It also improves the performance bottlenecks of existing lightweight models in handling the low-resolution, high-noise, and complex features characteristic of real-world infrared images from power plants. OBJECTIVES: The objective is to propose an Infrared Adaptive Dynamic Perceptron (IADP) that enhances defect recognition accuracy and hot spot detection precision for drone-based infrared inspections, while maintaining low computational complexity and achieving real-time performance on resource-constrained embedded platforms. METHODS: The core method employs a dynamic convolution mechanism that aggregates multiple heterogeneous kernels based on input-dependent attention. The key innovations focus on the resource limitations of drones: a global average pooling method tailored to infrared images optimizes the attention mechanism, and the introduction of the Softplus activation function mitigates the loss of low-contrast features. Dynamic convolution kernels (DK-conv) are used to effectively adapt to the regular textures and localized defects of photovoltaic panels. A hierarchical temperature annealing training strategy is introduced to address training stability issues. RESULTS: Experiments on a dataset of 11,608 drone infrared images show that the IADP-MobileNetV3-Small model improves defect recognition accuracy (TOP-1) by 4.0% and hot spot detection precision (mAP@0.5) by 4.7%, with only a 3.5% increase in computational complexity. It achieves 40.1 FPS on an embedded platform, exceeding the 20 FPS engineering requirement. CONCLUSION: The proposed IADP model effectively meets the demands for real-time, accurate, and efficient infrared inspection in photovoltaic power plants using drones, offering a good balance between performance and computational cost for embedded deployment.