Dynamic convolution and multi scale attention for small defect detection in photovoltaic modules with PED-YOLO

Defect detection in photovoltaic modules is crucial for extending their service life. To address the challenges of detecting small and blurry targets, this study proposes PED-YOLO, an improved model based on YOLOv11. The standard convolution is replaced with PConv, which leverages a larger receptive field to capture small defects and contextual information for more accurate feature extraction. The backbone incorporates DyC3K2, enabling adaptive convolutional parameters to enhance detection precision and robustness. Additionally, the C3K2-EMA module applies a multi-scale attention mechanism to redistribute feature weights, emphasising critical details and improving small-object detection, such as hot spots. A fourth detection layer with DyHead-detect further strengthens classification and localisation for small targets, optimising output efficiency. Experiments show that PED-YOLO achieves an overall mAP of 94.2% and a small-target defect mAP of 84.2%, representing gains of 2.6% and 9% over YOLOv11. Compared with RT-DETR, YOLOv12, Improved Faster-RCNN, and other state-of-the-art enhanced models, PED-YOLO demonstrates superior accuracy in both general and small-object defect detection. These results confirm the effectiveness of the proposed enhancements and highlight PED-YOLO’s potential for practical photovoltaic module inspection applications.