A Single-Stage Photovoltaic Module Defect Detection Method Based on Optimized YOLOv8
Yihong Gao, Chengxin Pang, Xinhua Zeng, Pengyi Jiang
Abstract
Defect detection in photovoltaic (PV) modules presents significant challenges. In the presence of inconspicuous or small-scale defects, downsampling operations can cause features to vanish, and computational complexity increases as the model deepens, resulting in reduced detection accuracy, higher latency, and challenges in deploying the model on edge devices with limited computational resources. In response, this article proposes a single-stage object detection model based on YOLOv8, the PSA-PVdetector (PSA-det). The core innovation of PSA-det is the novel Partial Spatial Attention (PSA) mechanism, which integrates Partial Convolution (PConv) with Spatial Attention (SA) to optimize feature extraction and reduce computational overhead. For small-scale defect detection, a Multi-Channel Feature Fusion (MCFF) detection head is designed to enable finer-grained center prediction. Additionally, ShapeIoU is employed as the bounding box regression metric, enhancing the classical IoU by incorporating bounding box shape, thereby improving defect localization accuracy. Experimental results demonstrate that PSA-det achieves an mAP50 of 87.2% on the Panel-2 dataset and 72.0% on the Solar dataset. On the Nvidia Jetson Xavier NX DK platform, PSA-det achieves an inference latency as low as 2.6 ms, effectively balancing accuracy and real-time performance.