A Multi-Stage Deep Learning Framework for Document Image Restoration

Real-world camera-captured document images often exhibit complex degradations, including cast shadows, non-uniform illumination, and contrast distortion, which severely degrade visual quality and prevent robust document analysis. In this paper, we present an illumination estimation multi-stage deep learning framework for restoring document images that explicitly separates shadow suppression from illumination normalization. The proposed pipeline involves an initial deep network estimating and mitigating shadow-induced intensity variations before a refinement network corrects global illumination consistency while maintaining textual structure and fine document details. By decomposing the enhancement task into complementary stages, the framework effectively copes with both local shadow artifacts and global lighting imbalance in unconstrained document imaging scenarios. Extensive experiments on real-world camera-captured document images reveal that the proposed method provides visually coherent enhancement with more readable results compared to conventional image processing techniques and existing deep learning-based methods. Standard image quality metrics have been quantitatively evaluated, showing notable gains. The results indicate that the proposed framework offers a robust and practical preprocessing solution for analyzing camera-based document images.