DSAFNet: Dilated–Separable Convolution and Attention Fusion Network for Real-Time Semantic Segmentation

Real-time semantic segmentation has been widely adopted in resource-constrained applications such as mobile devices, autonomous driving, and drones due to its high efficiency. However, existing lightweight networks often compromise segmentation accuracy to reduce parameter count and improve inference speed. To achieve an optimal balance among accuracy, latency, and model size, we propose the Dilated–Separable Convolution and Attention Fusion Network (DSAFNet), a lightweight real-time semantic segmentation network based on an asymmetric encoder–decoder framework. DSAFNet integrates three core components: (i) the Double-Layer Multi-Branch Depthwise Convolution (DL-MBDC) module that fuses channel splitting and multi-branch depthwise convolutions to efficiently extract multi-scale features with minimal parameters; (ii) the Multi-scale Dilated Fusion Attention (MDFA) module that utilizes factorized dilated convolutions and channel-spatial collaborative attention to expand the receptive field and reinforce key contextual features; (iii) the Multi-scale Attention Lightweight Decoder (MALD) that integrates multi-scale feature maps to generate attention-guided segmentation results. Experiments conducted on an RTX 3090 platform demonstrate that DSAFNet, with only 1.00 M parameters, achieves 74.78% mIoU and a frame rate of 74.74 FPS on the Cityscapes dataset, while 70.5% mIoU and a frame rate of 89.5 FPS on the CamVid dataset.