RAP-DETR: Enhancing RT-DETR for Railway Track Defect Detection

Abstract

To safeguard the smooth operation of railway transportation, this paper proposes an improved RT-DETR railroad track defect detection algorithm, RAP-DETR, which detects defects such as scratches, burns, surface wear, and fractures that may occur during railway operation. Firstly, the backbone network is streamlined by integrating the CSP-RAB module for multi-scale feature fusion, which not only enhances denoising performance but also lowers resource consumption. . Secondly,to augment the Attention-based Intra-scale Feature Interaction (AIFI) module, learnable positional coding is introduced to enhance the detection efficiency by dynamically adapting the positional encoding. Finally, the Pinwheel-shaped Convolution (PSConv) module, which is based on a novel windmill-style convolution, is put forward. It effectively refines the standard convolution operation, substantially expands the receptive field, and facilitates the enhancement of feature extraction. As a result, a higher detection accuracy is achieved. Experimental validation conducted on the RailDefect dataset demonstrates that the [email protected] of RAP-DETR reaches 84.8%, representing a 4.4% improvement compared to the original RT-DETR. Moreover, the precision and recall rates of the model have increased by 2.3% and 4.7%, respectively. Meanwhile,its parameter count has decreased by 35.2%, and the number of floating-point operations (FLOPs) has been reduced by 7.54%. These notable improvements underscore the robust capability of the proposed model to effectively detect defects on railway tracks. The RailDefect dataset is publicly available at https://github.com/0317cellxie/RailDefect.