<p>Single-image deraining aims to recover clean images from rain-degraded inputs, yet remains challenging due to cross-scale, anisotropic streaks and spatially varying structures. Prior-driven methods are often brittle under complex rain; CNNs are efficient but limited in long-range dependency modeling; transformers capture global context with quadratic complexity and may oversmooth fine details; and recent state-space models (SSMs) provide near-linear scaling but can underutilize high-frequency cues. To address these issues, we propose DFMIR-Net, a dual-branch deraining network that couples efficient global dependency modeling with explicit frequency guidance. In the main spatial branch, a selective SSM backbone performs multi-directional scanning to capture long-range streak structures under near-linear scaling, while a lightweight local enhancement module strengthens textures and edges. At the bottleneck, an FFT-based auxiliary branch builds a phase-preserving spectral representation and adaptively fuses it with spatial features, enabling spatial–spectral cues to jointly constrain restoration and improving high-frequency fidelity. Experiments on Rain100L/Rain100H and Test100/Test1200/Test2800 show that DFMIR-Net achieves competitive or improved PSNR/SSIM and produces clearer visual details compared with representative CNN-, Transformer-, and SSM-based baselines, with a small practical overhead from the frequency branch. We further provide computational profiling (Params, FLOPs and latency) and discuss scalable multi-GPU training and practical single-GPU inference for high-throughput deployment.</p>