Lightweight and High-Precision Detection Model for Battery Terminals in Complex Environments

Accurately locating the welding area of the battery module pole is the key to achieving high-precision and high-quality welding. Given the strict requirements of the welding environment for battery module pole detection and the shortcomings of the current visual inspection method in real-time performance, an improved pole detection method based on YOLOX is proposed. The introduced densely connected bottleneck module is applied to the CSPLayer in the output part of the backbone, and a parallel connection method is adopted. The residual connection is introduced to build a more efficient CSP module, enrich the feature gradient flow, and enhance the feature extraction capability. A lightweight neck is introduced, and the original feature processing module is replaced by the VoV-GSCSP module. The depthwise separable convolution and channel shuffling strategies can reduce the model calculation complexity and maintain high accuracy. Experimental results show that compared with other detectors, our method improves mAP by 4.45%, increases F1 score by 5.92%, reduces inference time by 1.92 ms, and reduces the number of parameters by 0.22 M.