DDRNet-SDR: A Spatial-Frequency Refinement Network for Real-Time Semantic Segmentation

Real-time semantic segmentation aims to balance accuracy and inference effi-ciency, which remains a key challenge in computer vision. Although dual-resolution networks such as DDRNet achieve competitive performance, they are still limited by insufficient multi-scale context modeling, loss of high-frequency details caused by repeated downsampling, and suboptimal static fea-ture fusion strategies.
To address these issues, we propose an enhanced dual-resolution network, termed DDRNet-SDR. Specifically, a Spatial-Frequency Fusion Module (SFFM) is introduced to exploit frequency-domain priors for attention genera-tion, enabling effective spatial feature refinement and preservation of high-frequency information. In addition, a DWR-Conv module is designed to inde-pendently model high- and low-resolution branches, facilitating efficient multi-scale context encoding through multi-branch and multi-dilation structures. Fur-thermore, a dynamic synergistic supervision strategy that combines Online Hard Example Mining (OHEM) and Dice loss is adopted to balance pixel-level accuracy and region-level consistency.
Extensive experiments on the Cityscapes dataset show that DDRNet-SDR achieves 78.43% mIoU at 66.1FPS on a single 2080Ti GPU. Compared with the baseline, it yields a 1.03% absolute improvement in segmentation accuracy while maintaining real-time performance, demonstrating its effectiveness for latency-sensitive applications such as autonomous driving and UAV vision.