In order to mitigate feature extraction degradation of conventional convolutional neural networks in asphalt pavement distress detection caused by fixed kernels’ poor adaptability to sinuous distresses and confusion between low-contrast distresses and complex textures, a robust algorithm RD-Net (Road Distress-Net) was proposed. First, targeting the multi-directionality and irregular topologies of linear distresses, the C3k2_DySnakeConv module was constructed by integrating the continuity deformation constraint of dynamic snake convolution into the bottleneck structure, breaking fixed sampling grid limits and enhancing subtle crack detection capability. Second, the SPPF_LSKA module was built by embedding Large Separable Kernel Attention into the pooling network. Through spatial decoupling, the receptive field was expanded and long-range dependencies were established, suppressing background interference and strengthening distress semantic features. Finally, the feature fusion network was reconstructed via dynamic content-aware upsampling to alleviate spatial misalignment of heterogeneous features and mitigate information degradation of small and low-contrast distresses. The results show that on the RDD2022 and UAV-PDD2023 datasets, RD-Net achieves mAP@0.5 of 86.5% and 79.3%, respectively, representing improvements of 1.8% and 4.1% over the baseline YOLOv11n, with an inference speed of 476 FPS, which strikes a favorable trade-off between detection accuracy and real-time performance. It is concluded that the proposed multi-module framework provides a reliable technical solution for intelligent asphalt pavement distress inspection in complex field scenarios.