Abstract
Remote sensing images (RSIs) exhibit complex spatial patterns due to diverse object scales, orientations, and distributions. These characteristics pose significant challenges for accurate interpretation. While existing methods focus on spatial feature extraction or rotational constraints, they often overlook the intrinsic relationship between frequency-phase information and spatial structures, which is critical for precise localization and classification. In this paper, we propose a generic and flexible feature fusion framework, called Phase-aware Mamba Fusion Mechanism (PMFM), which jointly captures frequency-phase patterns and global spatial dependencies. PMFM consists of two core modules, i.e., Frequency-Phase Position Encoding (PhasePE), which enhances positional encoding by embedding frequency-phase cues, and Cross-attention–Mamba Progressive Fusion (CaMPF), which integrates global-local information across multi-scale feature maps using cross-attention and Vision Mamba. Furthermore, we design a PMFM-based feature pyramid architecture (PMFM-FPN) via a cross-scale concatenation strategy, enabling seamless integration into various detection and segmentation frameworks. Extensive experiments on five public RSI benchmarks (DOTA, DIOR-R, SISP, ISPRS Potsdam, and Vaihingen) across three major tasks, i.e., horizontal detection, oriented detection, and semantic segmentation, demonstrate that PMFM consistently improves performance and achieves competitive accuracy and robustness. In addition, an auxiliary experiment on ExDark further explores the behavior of phase-aware representations under general low-light visual conditions.
Funding Information
  • National Natural Science Foundation of China (62476040)
  • Fundamental Research Funds for the Central Universities (DUT24GF311)
  • Beijing Nova Program (20230484261)

This publication has 99 references indexed in Scilit: