GDANet: Frequency-Space Collaborative Learning with Wavelet-Inspired Directional Perception for Single Image Deraining

Rain streaks are complex degradations in images, exhibiting multi-scale characteristics and diverse directional patterns, often spatially entangled with background textures. Therefore, effectively removing rain while preserving fine-grained image details remains a significant challenge in single image deraining. Traditional methods, whether based on pure spatial filtering or global frequency analysis, often fall short in simultaneously capturing both multi-scale and orientation-selective features crucial for rain removal. Inspired by the inherent ability of wavelet transforms to provide powerful multi-scale and directional feature representations, we propose a novel frequency-space collaborative learning framework, named GDANet. GDANet achieves content-adaptive rain suppression and detail restoration through the synergistic integration of three key components. Wavelet-inspired Directional Feature Module (WDFM) leverages multi-directional Gabor kernels to extract orientation-aware multi-scale rain features, effectively addressing the directional diversity of rain streaks. The Frequency-Spectrum Dynamic Gate Aggregation (FSDGA) dynamically modulates spectral components, reducing phase distortion while faithfully preserving structural details in the frequency domain. The Multi-scale Gated Fine-Grained module (MGFM) employs adaptive spatial attention to precisely separate complex rain patterns from background textures in the spatial domain. Furthermore, we design a multi-scale perception loss that jointly supervises content, edges, and log-frequency spectra across various resolutions, ensuring comprehensive detail preservation. GDANet demonstrates strong parameter efficiency: compared with the DMSR, it introduces only 0.255M additional parameters (from 0.838M to 1.093M). Extensive experiments on Rain13K show GDANet achieves 37.47 dB PSNR on Rain100L, with clear gains in detail preservation. On the real-world RealRain-1K-L dataset, it achieves 41.02 dB PSNR and 0.986 SSIM, demonstrating strong performance in practical scenarios.