ShadowAug: A Multi-Strategy Data Augmentation Method for Image Shadow Removal

Shadow removal is a fundamental yet challenging problem in low-level vision. Although deep learning-based methods have achieved significant progress, their performance remains constrained by the limited diversity and scale of existing shadow datasets. In particular, current datasets often lack variation in shadow shapes and quantities, which hampers generalization to real-world scenarios. This paper introduces ShadowAug, an online data augmentation framework tailored for shadow removal tasks. ShadowAug incorporates two complementary strategies: ShadowCut, which randomly removes rectangular regions within shadow areas to introduce shape variation, and ShadowSplit, which partitions contiguous shadow regions into multiple segments through horizontal and vertical cuts, enhancing instance-level diversity. Experimental results on the ISTD+ dataset show that ShadowAug yields a PSNR increase of 0.43 dB compared to the baseline. These results demonstrate the effectiveness of ShadowAug in generating diverse and challenging training examples that improve shadow removal models’ robustness and generalizability.