Integrated DR-YOLOv8-Based Equipment Recognition Method for Regional Rail Transit Network

This paper introduces a DR-YOLOv8-based equipment recognition method for regional rail transit network by incorporating DC_C2f, RC_FPN, and PLI modules into the original YOLOv8 network architecture. The deformable convolution (DC_C2f) module enables adaptive adjustment of the network's receptive field. By integrating the RC_FPN concept, the neck structure of the YOLOv8 model is restructured to achieve multi-scale integrated extraction of key features. Additionally, PLI units with prior positional information are utilized, and both positional and classification information are combined to optimize the object search process of the model. Experimental results demonstrate that the proposed method increases precision by 6.3% on the train control onboard equipment image dataset compared to the original algorithm, reduces FPS by 7.3%, and decreases FLOPs by 6.6G, thus achieving superior detection performance in train onboard equipment recognition.