Single Image De-raining Guided by Fourier Transform Prior
Chaobing Zheng, Yao Yao, Wenjian Ying, Shiqian Wu · 2024
It is challenging to remove rain-steaks from a single image because the rain steaks are spatially varying in the rainy image. Data-driven rain removal algorithms, despite achieving good performance, still have some shortcomings, such as a heavy reliance on data and limited interpretability. A novel approach for single image de-raining guided by Fourier Transform prior knowledge is proposed in this paper. Leveraging the inherent frequency domain information, our method effectively reduces rain streaks and restores image clarity. Firstly, employing Fourier transform to decompose the rainy image into amplitude and phase components, with raindrops predominantly present in the amplitude image. Subsequently, applying data-driven algorithms separately to process the amplitude and phase components, and utilize Fourier inverse transform to obtain enhanced features with improved clarity. Finally, a neural network with attention mechanism is designed to enhance the processed features, thereby enhancing the robustness of the algorithm. Experiments show that the proposed algorithm significantly outperforms state-of-the-art methods in terms of both qualitative and quantitative measures.