Dual-Stream Feature Fusion Network for Electrode Sheet Defect Detection in Lithium-Ion Batteries

This study addresses the challenges of insufficient local feature sensitivity and the absence of global semantic correlations in detecting micro-defects on lithium-ion battery electrode sheets. We propose a deep learning detection method based on a dual-stream feature fusion mechanism. By constructing a dual-branch architecture with parallel convolutional neural networks (CNN) and vision Transformers (ViT), the method independently extracts local texture features and global morphological features of electrode sheet defects. A dynamic attention fusion gating module is designed to achieve adaptive weighted fusion of cross-modal features, addressing the inadequate representation of complex defects in traditional singlemodality approaches. Furthermore, a joint optimization loss function with spectral clustering regularization constraints is introduced to simultaneously enhance feature discriminability and cluster compactness. Experimental results demonstrate that the proposed method achieves $98.7 \%$ detection accuracy on a selfbuilt multi-source lithium-ion battery electrode sheet dataset, with a 19% improvement in inference speed compared to the conventional YOLOv5s model and a reduced missed detection rate of $0.35 \%$. The method reliably identifies six types of electrode sheet defects, attaining a $95.2 \%$ detection rate for 0.2 mm-level scratches, providing a high-precision and efficient visual inspection solution for lithium-ion battery intelligent manufacturing.