Lightweight Dual-Resolution Network for Real-Time Semantic Segmentation

Semantic segmentation has become a crucial technique in various domains, including autonomous driving and medical image analysis. To enhance both accuracy and efficiency in real-world applications, we propose LiteDRNet, a novel lightweight and efficient dual-resolution network tailored for realtime semantic segmentation. The proposed architecture adopts a dual-branch design and incorporates a lightweight feature fusion module, termed the LiteFusion Module (LFM), which facilitates multi-path feature interaction during training to enhance representation capacity. During inference, the architecture is restructured into a single-path design to significantly improve runtime efficiency. Furthermore, we introduce the Multi-scale Attention Pyramid Module (MAPM), which integrates a pyramid structure with attention mechanisms to strengthen multi-scale context modeling and feature fusion. Extensive experiments conducted on the Cityscapes dataset demonstrate that LiteDRNet achieves a favorable balance between segmentation accuracy and inference speed, underscoring its potential for practical deployment in real-time semantic segmentation tasks.