Enhanced YOLOv8 With VTR Integration: A Robust Solution for Automated Recognition of OCS Mast Number Plates

Abstract

Against the backdrop of advancing automation and intelligence in rail transit, the accurate detection and recognition of Overhead Contact System (OCS) mast number plates are crucial for ensuring train safety. Challenges such as scale variations, motion blur, and lighting changes have intensified the difficulty of feature extraction, revealing deficiencies in current algorithms regarding precision and processing speed. To address these challenges, we developed a YOLO-M4ST algorithm tailored for railway scenarios. This algorithm enhances the detection and recognition capabilities by refining YOLOv8 and integrating the VTR framework. YOLO-M4ST adopts lightweight MobileNetV4 as its backbone network, significantly boosting the inference speed without compromising detection accuracy. The algorithm incorporates the C2f-Star module, which replaces the bottleneck module in the C2f module, thereby notably enhancing the feature extraction capability. Furthermore, it employs a single-class target approach to reduce the input parameters. In addition, a three-stage VTR framework was utilized, eliminating the need for character segmentation, thereby shortening the training time and improving the network convergence speed and recognition accuracy. Experimental validation demonstrated that YOLO-M4ST achieved an average recognition accuracy of 99.1% on our railway dataset with an inference speed of 420 Frames Per Second (FPS).

Keywords

• OCS mast number plate
• deep learning
• target recognition
• YOLOv8