CycleGAN-based unsupervised image smoothing framework with wavelet downsampling and multi-scale spatially-adaptive attention

Edge-preserving smoothing is an important image processing operation designed to enhance low-frequency structural components while suppressing high-frequency textures and noise. However, existing methods entail high costs for parameter tuning and dataset requirements, and lack generalization across different images. In response, this paper proposes an unsupervised image smoothing framework based on a cycle-consistent adversarial network (CycleGAN). It learns smoothing relationships from unpaired, unlabeled data and uses adversarial training to generate high-quality smoothing results. To better leverage image information, this paper designs a wavelet-based downsampling module to extract key features from subbands in different frequency bands of the image. Furthermore, a multi-scale spatially-adaptive attention module is proposed, which dynamically adjusts the importance of spatial features and facilitates comprehensive information interaction by fusing image features at different scales. Additionally, a composite loss function is employed to guide network optimization and improve the quality of generated results. Qualitative and quantitative experimental results demonstrate that, compared to state-of-the-art smoothing methods, the proposed approach achieves both effective smoothing performance and computational efficiency.