DFMIR-Net: Dual-Frequency Mamba Network for Single-Image Deraining

Single-image deraining aims to recover a clean image from inputs degraded by rain streaks, yet remains challenging because the streaks are cross-scale, anisotropic, and spatially complex. Existing approaches share several limitations: prior-driven methods are brittle under complex rain; CNNs, though efficient, are confined to local modeling and lack long-range dependencies; Transformers capture global context but have O(N²) time and memory complexity with global self-attention and may oversmooth details; state-space models (SSMs) offer near-linear complexity but often underemphasize high-frequency details; and frequency-only designs risk ringing artifacts and unstable spatial–spectral fusion. To address these issues, we combine global dependency modeling, lightweight local texture enhancement, and an explicit frequency prior under near-linear complexity, balancing global consistency with fine-detail fidelity. Building on this idea, we propose DFMIR-Net: selective SSMs provide efficient long-range modeling to complement CNNs and Transformers; a lightweight local enhancement mechanism strengthens textures and edges, improving performance in dense rain; and a rain-aware frequency module at the bottleneck employs a phase-preserving spectral representation with adaptive fusion so that spatial features and high-frequency cues jointly constrain restoration, mitigating SSMs’ high-frequency underemphasis and stabilizing spatial–spectral fusion. Experiments on Rain100L/H and Test100/1200/2800 show that DFMIR-Net achieves competitive or improved PSNR and SSIM compared with CNN-, Transformer-, and SSM-based methods, yields clearer visual details, and maintains near-linear computation and memory scaling.