YOLOv8n-RSDD: A High-Performance Low-Complexity Rail Surface Defect Detection Network

Zhanao Fang,Liming Li,Lele Peng,Shubin Zheng,Qianwen Zhong,Ting Zhu

DOI: https://doi.org/10.1109/access.2024.3466559

IF: 3.9

2024-01-01

IEEE Access

Abstract:Detecting surface defects on railway tracks is of significant importance for reducing the risk of safety incidents in high-speed railways. In response to the challenges in the field of railway track surface defect detection, such as insufficient detection performance, high model complexity, and difficulties in terminal device deployment, this study proposes a new type of railway track surface defect detection network named YOLOv8n-RSDD. The network uses a minimalist VanillaNet as the backbone network for feature extraction, effectively simplifying the model structure and accelerating inference speed. Further, the study introduces a slim-neck module based on GSConv convolution to replace the original C2f module, thereby enhancing performance and improving the detection capability for small targets. Additionally, the integration of the CBAM attention mechanism significantly enhances the network’s ability to capture key information on the railway track surface, strengthening perceptual performance. To obtain actual railway track images, this study developed an image acquisition system and constructed the Rail-1600 dataset specifically for the detection of railway track surface defects. Experimental results show that YOLOv8n-RSDD improved by 2.3% in the mAP@0.5 metric compared to YOLOv8n, while maintaining the stability of mAP@0.5:0.95. In terms of computational resource consumption, GFLOPs were reduced by 44.6%, the number of parameters decreased by 58.1%, the model size was reduced by 56.5%, and inference speed was increased by 17.8%. YOLOv8n-RSDD also demonstrated outstanding performance on the RSDDs and NEU RSDDs-113 datasets, indicating its potential for practical application.