Photovoltaic Module Surface Defect Detection Model Based on Improved YOLOv8s

Overcoming the limited accuracy and poor multi-scale feature capture of conventional vision methods, we present a lightweight, highly precise YOLOv8s detector tailored for PV surface defects. First, the C2f module in the backbone network is structurally reconstructed, integrating the SENetV2 enhanced channel attention mechanism to improve multi-scale feature integration capability and reduce the omission of multi-scale and irregular defect features. Second, an efficient multi-scale attention mechanism (EMA) with cross-space interaction is introduced, leveraging a lightweight design to significantly enhance the discriminative representation capability of contextual features. Finally, a composite bounding box loss function (Inner-ShapeIOU) combining InnerIOU and ShapeIOU is proposed to optimize localization accuracy and accelerate model convergence. Experimental results show that the improved model significantly enhances detection accuracy while maintaining model lightweightness, achieving an average precision (mAP) of 92.1% at 116 FPS, which is a 2.7% improvement over the baseline YOLOv8s model.