Multi-scale Dynamic Network for Document Shadow Removal

Document clarity is pivotal for reliable information transmission, yet shadows remain a pervasive degradation that harms visual quality and downstream processing efficiency. Existing approaches often falter under extreme illumination, struggle with multi-scale shadow patterns, and exhibit limited generalization across document types and capture conditions. We present an end-to-end multi-modal, multi-scale framework that integrates advanced architectures with dynamic feature fusion to adaptively suppress shadows of varying sizes while preserving fine textual details. Concretely, we introduce a Multi-scale Dynamic Network (MDN) that performs scale-aware gating and cross-branch aggregation, enabling the model to emphasize informative cues and attenuate shadow bias at each resolution level. The pipeline is streamlined to reduce manual pre-/post-processing, thereby improving both efficiency and effectiveness in practical document workflows. Experimental results show consistent gains over traditional and deep learning baselines on the Jung and Kligler datasets, confirming superior accuracy and robustness under challenging lighting.