关键词: 缺陷检测, YOLOv8n, 深度学习, SE, CA, Focal Loss

Abstract:

Due to the fine scale of surface defects on photovoltaic cells and their high degree of overlap with background features, traditional detection algorithms frequently suffer from missed detections and false positives. The robustness of existing methods in complex scenarios needs to be improved. This paper aims to enhance the YOLOv8n model by proposing SBC-YOLOv8n (SE-BiFPN-CA-YOLOv8n)to improve the detection ability of micro-defects on photovoltaic cells and increase the recognition accuracy and stability in complex industrial environments. The model is optimized in three aspects: First, dual attention mechanisms, namely CA and SE, are integrated before and after the SPPF module in the backbone network to enhance the feature extraction ability of micro-defects from both spatial location perception and channel weight adjustment. Second, the BiFPN is adopted to replace the original neck network structure, which simplifies nodes and uses a bidirectional weighted fusion mechanism to strengthen multi-scale feature interaction and suppress background interference. Third, to address the class imbalance issue, the Focal Loss function parameters are optimized. The focusing parameter γ is increased from 2 to 3 to enhance the attention to hard-to-classify samples, and the class balance factor α is adjusted from 0.25 to 0.5, significantly increasing the loss weight of defect samples and reducing the background weight. The improved SBC-YOLOv8n model achieves an mAP@0.5 of 80.2% on the test set, a 4.2 percentage point increase over the original YOLOv8n model. At the same time, precision, recall, and F1 score have also significantly improved. While maintaining the real-time performance of the model, it effectively enhances the detection ability of micro-defects and overall robustness. The SBC-YOLOv8n model proposed in this study significantly improves the comprehensive performance of photovoltaic cell defect detection through the collaborative optimization of attention mechanisms, feature fusion structures, and loss functions, meeting the real-time detection requirements of industry and providing reliable technical support for photovoltaic manufacturing quality control.

Key words:

defect detection, YOLOv8n, deep learning, SE, CA, focal loss