Abstract: To address the limitations of incomplete rain streak removal and easy loss of texture details on hard sam-ples in existing deep learning based image deraining algorithms, this paper proposes an image deraining algorithm based on progressive frequency domain learning. The algorithm is composed of two cascaded U-Nets with identical structures. First, the input image is processed by the low-frequency correction U-Net to obtain a preliminary deraining result. A pixel-wise constraint is imposed in the spatial domain using the ground truth image, and a low-frequency constraint is applied in the wavelet domain. Subsequently, the preliminary result is fed into the high-frequency refinement U-Net to generate the final deraining output. On the basis of the pixel-wise constraint in the spatial domain and the high-frequency constraint in the wavelet domain, an additional contrastive constraint in Fourier domain is introduced. Furthermore, a novel spatial-frequency dual branch architecture is designed for the U-Nets. On one hand, an adaptive sparse self-attention mechanism is adopted to extract spatial domain features and suppress redundant interference. On the other hand, a multi-scale Fourier modulator is utilized to supplement frequency domain information. Experiments conducted on four public datasets demonstrate that the average PSNR and SSIM metrics of the proposed algorithm are improved by 1.10% and 0.27% respectively compared with existing state-of-the-art algorithms, achieving significant performance gains across diverse scenarios.