CMLNet: an improved lightweight high-precision network for solar cell surface defect detection

Electroluminescence (EL) imaging technology detects tiny defects at the production stage, which are small but have a significant impact on battery performance and life. In this paper, a CMLNet is proposed, which can effectively solve the problem of defect detection in EL images with complex backgrounds. First, the CSPHet network is introduced to replace the C2f structure in Yolov8. Second, a lightweight shared convolution detection head (LSCD-Head) is introduced. Then, a lightweight mixed local channel attention mechanism (MLCA) is integrated into the neck of the model to enhance the feature expression ability of the model. Finally view of Yolov8’s sensitivity to CIoU, an Inner-ShapeIoU solution is introduced to enhance model generalisation ability. In the experiment of the PVEL-AD dataset, the improved model has a 42.4% reduction in the number of parameters compared with Yolov8n. At the same time, map@50 and map@50–95 of this model can reach 92.6% and 64.7%, respectively, which are 3.1% and 2.4% higher than those of the original Yolov8n network, achieving lightweight and high-precision optimisation. In addition, the generalisation capability of the model is verified on the PV-Defect dataset. The results show that CMLNet has remarkable advantages in various performance indexes while maintaining lightweight performance.