A lightweight dynamic network framework based on feature complexity awareness and frequency-domain feature fusion for rail surface defect detection

In real rail surface defect images, defect targets are small and the background is complex. Traditional deep learning methods struggle to capture small defect features, leading to low recognition accuracy. However, improving small-target recognition performance requires higher model complexity, which reduces running speed on edge computing devices and fails to meet real-time detection requirements. To address these problems, we propose a lightweight dynamic network framework based on feature complexity awareness and frequency-domain feature fusion for rail surface defect detection. This framework consists of a Dynamic Network Module (DNM), a Fine-grained Channel Attention mechanism (FGCA), and a Lightweight Frequency-domain Feature Extraction Module (LFFEM). DNM adaptively allocates computing resources by input feature complexity, activating the optimal feature extraction branch dynamically for balanced inference speed and recognition accuracy. FGCA enhances the model’s focus ability on small details by the pixel-by-pixel feature focusing algorithm, improving small-target recognition accuracy under complex backgrounds. LFFEM further improves recognition performance through the multi-scale frequency-domain feature extraction algorithm, and significantly reduces model complexity by channel sparsification and feature reuse algorithms. Experimental results on the self-built rail surface defect dataset show that DNRD-Net attains a mAP of 84.46%, representing an improvement of 9.70% over the basemodel, while reducing parameters and FLOPS by 41.04% and 38.59%, respectively. Deployment on an edge computing device further shows a 253.49% increase in inference speed, enabling real-time inference. In addition, the model achieves mAP of 87.03% on the NEU-DET, indicating strong generalization capability. These results demonstrate that the proposed model effectively balances detection accuracy and computational efficiency, making it suitable for real-time, vehicle-mounted rail inspection applications.