An end-to-end shadow removal framework with an intuitive interaction scheme

Abstract

Shadow removal plays a crucial role in enhancing image quality by restoring the color and texture details of the shadow regions, thereby improving the performance of downstream visual tasks. Although recent shadow removal algorithms have achieved impressive results on benchmark datasets, shadows in such datasets are typically centralized and captured in relatively straightforward scenes. In contrast, real-world shadows tend to exhibit complex and irregular patterns due to the random distribution of objects, causing global processing methods to produce false positives and missed corrections. To address these challenges, this paper presents an end-to-end shadow removal framework leveraging Human-Computer Interaction (HCI), allowing simple bounding boxes to annotate targeted shadows. Our approach employs a novel chunked processing training strategy, which decomposes global shadow removal into iterative local refinements. Additionally, a Split-Channel module and an Edge-Weighted loss are incorporated to maintain consistent color and smooth edge transitions during restoration. Furthermore, an HSI-based shadow detection algorithm is proposed to generate shadow masks, facilitating end-to-end shadow removal. Experimental results demonstrate that our approach outperforms state-of-the-art methods on ISTD and SRD datasets, and exhibits robust performance on real-world images, effectively reducing restoration errors.