A Lightweight Multi-Scale Feature Enhancement Network for Real-Time Sementic Segmentation

Semantic segmentation, as a core task of pixel-level image understanding, faces two main challenges: limited computational resources and real-time requirements in practical applications such as autonomous driving and medical diagnosis. This paper proposes a lightweight multi-scale feature enhancement network (LMFENet) to address the issue of representation gap between semantic and detailed information caused by simple operations in existing feature fusion methods. The network is based on an encoder-decoder architecture: (1) In the encoding phase, it employs MobileViTv3 as the backbone network, which combines the advantages of CNNs and Transformers to balance global receptive field and computational efficiency; (2) In the decoding phase, it introduces a multi-scale feature enhancement module (MFEM), which uses a dual-path information enhancement mechanism to achieve cross-layer feature complementarity. Based on the idea of frequency domain separation, we propose a detailed feature fusion module(FFM-D) and a semantic feature fusion module(FFM-S). Specifically, FFM-D injects detailed low-level feature information into high-level features, while FFM-S reverses this by fusing high-level semantic information into low-level features, effectively narrowing the representation gap of multi-scale features. Experiments show that LMFENet achieves 75 % mIoU on the Cityscapes benchmark dataset with only 3M parameters and reaches an inference speed of 38 FPS at a resolution of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$512 \times 1024$</tex>, achieving a better trade-off between accuracy, inference speed, and model size.