A Lightweight and Accurate Defect-Detection Framework for Battery Cells Based on EfficientNet-YOLOv9

作者：Yixiang Liu, Jinlong Wu, Bing Leng, Haihua Guo, Yibin Li · 年份：2026 · DOI：10.1109/icara69401.2026.11480321 · 研究领域：Advanced Battery Technologies Research、Industrial Vision Systems and Defect Detection、Advanced Neural Network Applications

In this study, a lightweight and high-accuracy visual inspection framework is developed for distinguishing defective and normal prismatic lithium-ion battery cells under non-standardized recycling conditions. An EfficientNet-YOLOv9 variant is constructed by replacing the baseline CSP backbone with a compound-scaled feature extractor. Experimental results show that the proposed EfficientNet-YOLOv9 achieves$\text {1. 1 1 M}$parameters and 6.4 GFLOPs, representing a 46.6% reduction in parameters and a 24.7% reduction in computation compared with YOLOv9-t(2.08M, 8.5 GFLOPs). And the EfficientNet-YOLOv9 model achieves strong detection results, with Precision reaching 0.946, Recall reaching 0.960, mAP50 reaching 0.986, and mAP50-95 reaching 0.915. These results indicate that the proposed framework maintains high discriminative capability while substantially reducing model size and computational cost. Overall, EfficientNet-YOLOv9 provides an effective, lightweight, and computation-efficient solution for large-scale battery-cell inspection and offers strong potential for deployment in automated recycling systems.