DASFNet: A Direction-Aware Spatial-Frequency Network for Single Image Deraining
TL;DR
A direction-aware spatial-frequency network (DASFNet), an encoder–decoder framework with two modules: the cross-scale direction-aware convolution (CSDAC) module and the spatial-frequency collaborative channel attention (SFCCA) module, which outperforms state-of-the-art methods in quantitative metrics and visual quality.
Abstract
Image deraining is a low-level vision task that removes rain streaks and restores scene details. Balancing rain removal with the preservation of background detail remains challenging due to directional anisotropy, multiscale distribution, and complex interactions between rain and image textures. To address these issues, we propose a direction-aware spatial-frequency network (DASFNet), an encoder–decoder framework with two modules: the cross-scale direction-aware convolution (CSDAC) module and the spatial-frequency collaborative channel attention (SFCCA) module. CSDAC uses uniformly distributed directional branches with alternating kernels to model directional and multi-scale rain patterns. For precise rain-background separation, SFCCA first captures local context via a spatial feature refinement (SFR) branch. In parallel, a learnable frequency enhancement branch, fused with SFR, encodes global structural patterns. A sparse prompt channel attention mechanism then recalibrates the fused cross-domain features. Experiments on five synthetic datasets and two real-world benchmarks show that DASFNet outperforms state-of-the-art methods in quantitative metrics and visual quality, achieving an average PSNR of 29.55 dB and SSIM of 0.88, and improving PSNR by 0.44 dB on Rain100L and 0.26 dB on Test100 compared with the best competitor, while removing diverse rain streaks and preserving background details.