AI Research Analysis

Unlike conventional binary masks, MatteViT utilizes a continuous luminance-based shadow matte for precise spatial guidance. This matte, generated from a custom dataset of paired shadow/shadow-free images, captures subtle luminance variations and soft transitions. This detailed guidance allows the model to accurately localize shadow regions and restore them with high fidelity from the earliest processing stages.

The core of MatteViT is a Vision Transformer (ViT), enhanced with the HFAM and shadow matte integration. The self-attention mechanism of the ViT allows the model to effectively focus on shadow-affected regions while preserving the overall document structure. The architecture combines spatial and frequency-domain information for comprehensive shadow elimination and detail restoration.

MatteViT's training employs a composite loss function that combines Edge-aware Charbonnier Loss for spatial fidelity with FFT Loss for spectral consistency. The Charbonnier loss, weighted by Laplacian-derived edge information, emphasizes high-frequency regions, while the FFT loss minimizes discrepancies in frequency components, ensuring global structural consistency and local texture preservation.

Extensive experiments on RDD and Kligler datasets demonstrate MatteViT's state-of-the-art performance in document shadow removal. Quantitatively, it achieves superior PSNR, SSIM, and RMSE. Qualitatively, it preserves text-level details vital for OCR accuracy, validating its practical utility for real-world document digitization and robust performance across diverse document types and illumination variations.