Answers are grounded in this paper and cite their sources.
High-resolution remote sensing images typically exhibit significant intra-class variability and high inter-class similarity, while single-backbone networks struggle to capture both global semantics and local details simultaneously, thereby posing significant challenges for accurate semantic segmentation. In this paper, we propose a novel parallel branch cooperative enhanced visual Mamba semantic segmentation model, called CDDMamba. The CDDMamba model integrates two complementary branches: the channel-decoupled flow-matching branch enhances global context modeling and long-range dependency capture, and the dual-path multi-axis polynomial convolution branch captures fine-grained local details, thus reducing classification confusion among visually similar classes. The global branch decouples and groups deep features based on channel-wise spatial variation, then enhances global semantic representations through differentiated flow-matching paths. The local branch combines local DoubleConv with multi-axis polynomial strip convolutions parameterized by Legendre polynomials to extract fine-grained details. CDDMamba proposes a novel global–local feature discrepancy ranking mechanism based on spatial position, which then adjusts the difference-guided Hilbert-style scan order under spatial proximity constraints to adaptively fuse local details derived from two state streams with global contextual information, thereby achieving more refined segmentation results. Extensive experiments on the real datasets show that CDDMamba achieves highly competitive segmentation accuracy while maintaining outstanding computational efficiency.
No takes yet. Share an insight, caveat, or question.
Bai et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: