RestoreCUFormer: Transformers to Make Strong Encoders via Two-stage Knowledge Learning For Multiple Adverse Weather Removal

Removing bad weather effects from images is crucial for environmental perception, as it can provide clear and high-quality input for various downstream computer vision tasks. However, existing approaches are constrained to single weather removal or relying on multiple pre-trained weight sets to tackle different types of adverse weather scenarios. To addresses this challenge, We propose a two-stage knowledge learning mechanism utilizing knowledge distillation and feature alignment including knowledge teaching and knowledge examination. We also proposed RestoreCUFormer, an innovative model which merges the capabilities of CNN and Transformer to more accurately estimate the statistical patterns between the original and restored images. Experiments on multiple datasets achieve satisfactory results. The experimental findings demonstrate that the RestoreCUFormer model exhibits excellent performance in simultaneous adverse weather removal.