Abstract: Aiming at the problem of difficulty of detecting surface defects of soft-packed lithium-ion batteries, a detection method based on Faster R-CNN was proposed. By introducing the feature pyramid network model into the backbone network ResNet50, the semantic information of the high-level feature layer and the detail information of the low-level feature layer were fused to improve the network's detection accuracy of defects on small targets; the cross-cross-attention mechanism was introduced to make the network focus on the defect features and to suppress the background information of reflections; and the ROI Align was used to avoid the two quantization errors generated by the pooling operation of the regions of interest affecting the network's detection accuracy of defects. The experimental results show that the mAP of the improved model reaches 80.2%, which is improved by 3.9% compared with Faster R-CNN, and has certain value for engineering applications.