From Attention Stacking to Architecture Rethinking: Dual-Path Decoupling for Efficient Remote Sensing Image Segmentation

Remote sensing image segmentation has long struggled with balancing efficiency and accuracy. Most current research focuses on hybrid architectures that combine state-space models with CNNs, with CM-UNet being a notable example. However, its single-path decoding mechanism creates a performance bottleneck, and its multi-level homogeneous attention mechanism introduces significant computational redundancy. This paper proposes a novel Dual-Path U-Net (DP-UNet), using CM-UNet as the baseline. Our key insight is to explicitly decouple the processing of global context and local details, which are inherently distinct tasks. First, we innovate by designing a Decoupled Visual State Space (DVSS) block: one path models long-range dependencies through a simplified state-space model, while the other path integrates Efficient Channel Attention (ECA) and a new Parameter-Modulated Convolution (PMC) module to enhance detail extraction. The features from both paths are fused through an Adaptive Path Fusion Gate (APFG) module. Second, we streamline attention by removing redundant units from the SS2D module in CM-UNet and simplifying its multi-scale fusion module into a Multi-scale Spatial Kernel (MSK) module, which refines its structure to focus solely on spatial attention, reducing model complexity. Experimental results on three benchmarks show that DP-UNet outperforms existing state-of-the-art methods in both accuracy and efficiency. This study demonstrates the potential of explicit decoupling architectures in visual tasks. The codes are available at https://github.com/sharkzzzy/DP-UNet.