GLA-YOLO: A Lightweight Solar Cell Defect Detection Network Based on Spatial-Channel Collaborative Attention

Abstract

With the rapid development of the photovoltaic industry, higher requirements are imposed on the accuracy and efficiency of quality inspection in large-scale manufacturing and application. However, defects such as cracks, fragments and black cores are prone to occur during production, transportation and under complex working conditions, which not only deteriorate the photoelectric conversion efficiency but may also induce potential safety hazards. To address the challenge of achieving model lightweight while improving detection accuracy in edge deployment, this paper proposes a lightweight spatial enhancement detection model, namely GLA-YOLO. Based on YOLOv5s, GhostConv and C3Ghost are introduced to reduce computational complexity and parameter scale. To handle the small size, complex morphology and background interference of photovoltaic defects, a Lightweight Spatial-Channel Collaborative Attention (LSCA) module is designed and embedded into the backbone to enhance fine-grained feature representation. Meanwhile, a lightweight residual attention module (LRA) is further proposed and introduced into key layers to strengthen defect-related features and suppress background interference. In addition, the WIoU bounding box regression loss is adopted to improve localization stability and regression accuracy of small defect targets. Experimental results show that, compared with the original model, the proposed method improves mAP@0.5 and mAP@0.5:0.95 by 2.18 and 2.48 percentage points, respectively, while reducing computational load, model size, and parameter count by 36.7%, 37.7%, and 39.4%, respectively. The model achieves 123 FPS on an NVIDIA RTX 4080 and a compute-pipeline throughput of 35.60±0.27 FPS on an NVIDIA Jetson Orin Nano Super under FP16 TensorRT inference, demonstrating both high desktop-GPU efficiency and practical real-time edge-deployment capability.