Lightweight and real-time detection model for railway fasteners based on feature divide-and-conquer and fusion

YAN Huabiao; LIN Chuxin; HUANG Lve; LI Dongli; LIU Cibo; XU Fangqi

doi:10.11860/j.issn.1673-0291.20230145

Abstract：To address the challenge of balancing accuracy and detection speed in processing large-scale visual image data of railway fasteners on embedded devices in real time,a lightweight real-time detec-tion model based on attention divide-and-conquer and feature fusion is proposed.First,a hybrid divide-and-conquer attention module,leveraging both spatial and channel features,is introduced to en-hance the model's feature extraction capability and reduce interference from complex backgrounds in the image.Second,a dual divide-and-conquer feature fusion method is designed to improve detection performance across targets of varying sizes.In the construction of the cost volume detection head(YOLO Head),a Varifocal Loss(VFL)function is employed to replace the binary cross-entropy loss used in YOLOX-Nano,thereby enhancing the accuracy of lightweight real-time detection.Further-more,a Random Alpha-IoU(RAL)loss function is adopted to dynamically adjust parameters,slow down convergence,and optimize the training curve,thus preventing the model from falling into local optima.Finally,a dataset of 10,233 annotated fastener targets categorized into six types is used for evaluation.Comparative experiments are conducted using mainstream object detection models,includ-ing YOLOX-Nano,Faster R-CNN,and YOLOv8n.The research results indicate that the proposed model achieves a frame rate of 60.24 Frames Per Second(FPS)and an Average Precision(AP)of 83.40％,representing a 3.24％improvement over the baseline.The parameter count is 2.31 M,which is 54.08％fewer than YOLOX Tiny,and the floating-point operations is 1.99 G,a 69.15％de-crease compared to YOLOX Tiny.These findings provide valuable insights for the development of lightweight real-time detection models and embedded computing systems.