GCEI-YOLO: a lightweight model applied to the detection of bearing surface

Abstract As a key component of modern industrial equipment, bearings are susceptible to various surface defects during the manufacturing process. Based on the YOLOv8 architecture, this study has developed a new single-stage object detection model, GCEI-YOLO. By adopting the lightweight feature extraction network GhostConv-C2Flight, the redundancy in the computation process and memory access frequency are effectively reduced. The dynamic grouping strategy and multi-scale branch processing were introduced to obtain the efficient channel attention mechanism of EGCS-EMA, enhancing important channel feature information. By adopting the Shape-IoU loss function and a reasonable gradient gain allocation strategy, the model pays more attention to ordinary quality samples. The resulting GCEI-YOLO model has 2.46M parameters and 7.4GFLOPS of computational cost, balancing compactness and performance. Compared with the benchmark model, the improvement rates of mAP50, mAP50-95 and Precision in GCEI-YOLO were 2.98%, 4.43% and 1.25% respectively, and the number of parameters and the amount of calculation were reduced by 18.27% and 9.75% respectively. The ablation experiments show that GCEI-YOLO achieves synergy through hierarchical division of labor, dynamic resource allocation and explicit training adaptation, and is more suitable for bearing surface defect detection tasks and embedded platform deployment and inference.