Defects in photovoltaic (PV) modules, including hotspots, shading, and diode failures, significantly reduce power-generation efficiency and pose safety risks. This study proposes a real-time detection and localisation framework for PV defects based on infrared images acquired by unmanned aerial vehicles (UAVs). A dedicated dataset of 5583 infrared/visible images was constructed under standardised acquisition conditions. An improved rotating-bounding-box detector, termed YOLO-CLO, was developed upon YOLOv8-OBB by introducing a lightweight C3m module and a shared-convolution LSCD-OBB detection head. The proposed detector attains 99.1% mAP@0.5, 96.7% mAP@0.5:0.95, and 59.88 FPS with only 8.52 M parameters and 23.6 GFLOPs, outperforming the baseline in both accuracy and efficiency. A multi-feature image-processing pipeline combining gradient, grayscale, temperature, and morphological cues identifies hotspots, diode failures, and obstructions with detection accuracies of 96.97%, 100%, and 88.89%, respectively. A component-level localisation strategy integrating GNSS metadata, the Hough transform, and an improved K-means clustering algorithm accurately recovers the row–column index of each defective module within an array. Comparative experiments against YOLOv5 and Faster R-CNN confirm the superiority of the proposed framework. The method offers low hardware dependency and is suitable for engineering deployment in large-scale PV power stations.