Recursive residual Fourier transformation for single image deraining

Zhiyuan Bao, Mingwen Shao, Yecong Wan, Yuanjian Qiao

Abstract: Reconstructing a rain-free image from its degraded counterpart requires transforms regarding information from diverse frequency levels. On the frequency domain perspective, despite current CNNs-based methods exhibit excellent abilities in capturing the high-frequency components of images, they often fail to adequately consider or overlook the low-frequency information. To address this challenge, we introduce a fast Fourier transform block (FFTB) which can effectively capture both long-term and short-term interactions, while integrating high- and low-frequency residual information. Our FFTB is a conceptually simple yet computationally efficient block, leading to remarkable performance gains. Based on FFTB, we further develop a multi-stage architecture termed recursive residual fourier network (RRFNet) to enhance the ability of capturing and modeling spatial and frequency domain visual cues. To fully maximize the performance of RRFNet, a novel global–local convert test strategy is employed to alleviate the training–testing inconsistency. Experimental results on the synthetic and real-world datasets demonstrate that our RRFNet performs favorably against state-of-the-art methods while enjoying faster inference speed.