Feature separation and non-shadow information-guided shadow removal network

Abstract

To tackle the performance bottlenecks and color deviation issues stemming from current shadow removal methods, a feature separation and non-shadow information guided shadow removal network (FSNIG-ShadowNet) was constructed. In the separation and reconstruction stage, the shadow image was separated into direct light and ambient light using self-reconstruction supervision, with decoupling of lighting types and reflectance. Subsequently, a decoder was employed to re-couple the separated features to yield shadow-free images. In the refinement stage, the network focused on the adjacent regions of shadow and non-shadow, incorporating a local region adaptive normalization module to transfer the color priors of local non-shadow region to shadow regions for mitigating color deviation between the two regions. Experimental results demonstrate that the proposed FSNIG-ShadowNet achieves competitive results compared to other state-of-the-art methods.

Keywords

• shadow removal;feature separation;self-reconstruction;color prior