YOLOv5 Optimization for the Identification of Surface Defects in Solar Cells

Abstract

In light of the dynamic solar cell picture background information, varied problem morphology, and large-scale differences, a solar cell defect identification approach utilizing an enhanced YOLO v5 algorithm is suggested. A tiny predicted defect head is added to the model's network structure, which improves target detection accuracy at various scales. Lastly, the model's feature extraction capability is further enhanced through the addition of the ECA-Net attention mechanism. Initially, the CSP module incorporates the deformable convolution to attain an adaptable learning scale as well as perceptual field size. The K-means++ clustering anchoring box algorithm, the CIOU loss function, and the Mosaic & Mix-up fusion data improvement are used in this research to better optimize & enhance the YOLO v5 algorithm. As demonstrated by the experimental results, the new YOLO v5 method accomplishes 89.64% map for a model trained on a solar cell EL image the data set, a percentage 7.85% higher than the original algorithm's map. It also reaches a speed of 36.24 frames per second, which allows it to meet real-time requirements while still completing the solar cell defect identification task with greater accuracy.