ReHiT: Retinex-guided Histogram Transformer for Mask-free Shadow Removal

Abstract

While deep learning techniques have significantly advanced the field of shadow removal, a considerable number of current methods depend on shadow masks, which are often challenging to acquire accurately. This reliance on masks restricts their ability to generalize effectively to unconstrained real-world scenarios. To address this limitation, we introduce ReHiT, an efficient mask-free shadow removal framework that leverages a hybrid CNN-Transformer architecture, guided by the principles of Retinex theory. Our approach begins with a dual-branch pipeline designed to model the reflectance and illumination components of an image separately. Each of these components is then processed and restored by our novel Illumination-Guided Hybrid CNN-Transformer (IG-HCT) module. Furthermore, in addition to incorporating CNN-based blocks that excel at learning residual dense features and performing multi-scale semantic fusion, we have developed the Illumination-Guided Histogram Transformer Block (IGHB). This specialized block is designed to effectively handle the complexities of non-uniform illumination and spatially intricate shadow patterns. Comprehensive experiments conducted on several standard benchmark datasets demonstrate the superior performance of our proposed method compared to existing mask-free techniques. Notably, our solution achieves competitive results while boasting one of the smallest parameter counts and fastest inference speeds among the state-of-the-art models. This highlights the practical applicability of our method for real-world applications where computational resources may be constrained.