A Lightweight Multi-scale Dilated Fusion Network for Real-Time Semantic Segmentation

In applications like autonomous vehicles and robotics, real-time semantic segmentation plays a critical role. While current lightweight models prioritize parameter reduction, this often compromises their segmentation performance. Striking an optimal trade-off between precision and model size is essential for real-time segmentation tasks. This paper proposes a lightweight multi-scale dilated fusion network (LMDFNet), which consists of two modules. One is the multi-scale dilated fusion attention (MDFA) module and another is the multi-scale attention lightweight decoder (MALD) module. To capture more comprehensive contextual information, the MDFA module enhances the receptive field in the encoder’s final layer. The MALD module aims to fuse multi-scale feature maps and generate attention-guided recovery of input images. On Cityscapes, LMDFNet attains 74.2% mIoU at 136.7 FPS (RTX 3090) with merely 0.81M parameters, balancing accuracy, speed, and model size.