CRFormer: A cross-region transformer for shadow removal

Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Zhihao Liu, Song Wang

Abstract: Highlights•CRFormer adeptly restores image signals damaged by shadows.•A hybrid CNN-Transformer model is proposed to exploit regional context for shadow removal.•A region-aware cross-attention is proposed to aggregate non-shadow features into shadow regions.•Superior performance can be achieved in both image and video shadow removal tasks.