MSA-YOLO: Multi-scale Adaptive Network for surface defect detection of cylindrical lithium battery

Surface defects are unavoidable during the manufacturing of cylindrical lithium batteries and may compromise battery performance and safety. However, existing detection methods still face considerable challenges in dealing with complex defect morphologies, low contrast between defects and backgrounds, and severe class imbalance, which limits detection accuracy and robustness. To address these issues, this paper proposes MSA-YOLO (Multi-Scale Adaptive Network for Cylindrical Lithium Battery Surface Defect Detection). Built upon YOLOv9s, the proposed method introduces a Re-parameterised Dynamic Selective Kernel module (Rep-DSK) to enhance the extraction of weak and fine-grained defect features. In the feature fusion stage, an Adaptive Multi-scale Feature Fusion module (AMFF) is incorporated to improve the representation of multi-scale defects under complex backgrounds. In addition, a Dynamic Adjustment Loss (DALoss) is developed by combining Slide Loss and WIoU to alleviate the learning bias caused by class imbalance. Experimental results demonstrate that MSA-YOLO improves mAP50 by 3.0% over baseline on the Battery-Defects dataset and achieves gains of 2.4% and 1.8% in cross-domain evaluations on GC10-DET and NEU-DET, respectively. These results indicate that the proposed method can effectively enhance the accuracy and robustness of industrial battery surface defect detection, providing an efficient and reliable solution for practical applications.