Improved YOLOv11s rail defect detection algorithm based on multiscale feature fusion

Aiming at the common problems in rail defect detection, such as excessive model parameters, poor detection effect of small targets, false detection and missing detection, this paper proposes an improved YOLOv11s lightweight algorithm. This algorithm effectively realizes multi-scale feature fusion by replacing the standard detector head of YOLOv11 with Lightweight Shared Detail-enhanced Convolutional Detection Head (LSDECD). In addition, the group normalization operation further improves the detection accuracy and stability of the model. On this basis, the Focusing Diffusion Pyramid Network (FDPN) is designed, which significantly optimizes the detection performance in the process of feature information fusion, especially in the recognition ability of small target defects. In order to improve the efficiency of feature extraction, this paper also improves the C3k2 module by adopting the Omni-dimensional Dynamic Convolution (ODConv). Experimental results demonstrate that the improved model achieves 4.3%, 2.6%, 2.4%, and 1.7% improvements in precision, recall, mAP50, and mAP50-95, respectively. Additionally, computational load is significantly reduced, achieving a balance between precision and efficiency, thus validating the algorithm's effectiveness and potential for application.