A lightweight adaptive network with context fusion for battery defect detection

Detecting defects in lithium batteries is critical to ensuring product consistency. However, existing visual models face significant challenges in identifying minute defects, dealing with intricate defect structures, and overcoming background interference in industrial environments. To address these practical challenges while meeting industrial production demands, a lightweight context fusion adaptive network for defect detection, named LACF-YOLO, is proposed in this work. LACF-YOLO replaces the conventional concatenation operation in the Path Aggregation Feature Pyramid Network (PAFPN) structure with a new context fusion method, effectively alleviating the degradation and loss of local and global features during feature fusion. Moreover, an adaptive defect detection module incorporated into the backbone network utilizing deformable convolution to accurately extract complex defect features. To further reduce the model’s parameter and computational load, a lightweight detection head based on depth-wise separable convolution is designed. Evaluated on a custom battery surface defects dataset, LACF-YOLO achieved a mean Average Precision (mAP) of 87.6%, outperforming other models. Moreover, LACF-YOLO effectively balance accuracy, real-time performance, and parameter efficiency, establishing itself as the state-of-the-art technology for battery defect detection.