A lightweight task-adaptive dynamic fusion model for lithium-ion battery electrode sheet defect detection

Abstract Defect detection in lithium-ion battery electrode sheets is a critical step in quality assurance. In complex industrial environments, existing detection methods are prone to missed detections and false alarms due to significant variations in defect scale and the low contrast of defect edges. To address these challenges, this paper proposes a lightweight defect detection model, termed MEAF-YOLO. Specifically, a Dynamic Task-aware Fusion Head (DTFHead) is designed to enable adaptive adjustment of feature fusion paths by introducing a learnable fusion-depth mechanism and a continuous cross-scale feature fusion strategy. In addition, a task-feedback modulation mechanism is incorporated to enhance the representation of complex defects. On this basis, a lightweight backbone network, namely the multi-scale edge-aware network (MSAENet), is constructed to improve the representation of small-scale and low-contrast defects through multi-scale convolutions and edge-enhancement operations. Meanwhile, a context-aware dynamic fusion (CADF) structure is introduced to achieve efficient integration of multi-level features. Experimental results show that, compared with YOLOv11m, MEAF-YOLO achieves 88.2% mAP@0.5 on the lithium-ion battery electrode sheet defect dataset for detection dataset while maintaining a lightweight architecture with only 12.75 M parameters, 51.1 G floating-point operations, and a real-time inference speed of 168.13 frames per second. Moreover, precision and recall are improved by 3.1% and 3.7%, respectively. Evaluation on the public NEU-DET dataset further yields 76.8% mAP@0.5 and 44.2% mAP@0.5:0.95, confirming the generalization capability of the proposed method. Overall, MEAF-YOLO achieves a favorable balance between detection accuracy and computational efficiency, providing an effective solution for lithium-ion battery electrode sheet defect detection in complex industrial environments.