Smart document restoration that brings faded history back to life, pixel by pixel with DiffHDR
Paper propose to restore damaged historical documents using diffusion models and a custom dataset of 28,552 damaged-repaired image pairs.
-----
https://arxiv.org/abs/2412.11634
🔍 Original Problem:
Historical documents suffer severe damage over time through character missing, paper damage, and ink erosion. Existing methods focus only on basic document processing like binarization and enhancement, without addressing comprehensive repair.
-----
🛠️ Solution in this Paper:
→ Introduces HDR28K dataset containing 28,552 damaged-repaired image pairs with character-level annotations
→ Proposes DiffHDR, a diffusion-based network that uses semantic and spatial information for document repair
→ Implements three key degradation types: character missing, paper damage, and ink erosion
→ Uses character perceptual loss to ensure content preservation in repaired regions
→ Incorporates classifier-free guidance for attribute-sensitive repair
-----
💡 Key Insights:
→ Synthetic damage can effectively train models for real document repair
→ Character-level annotations crucial for accurate restoration
→ Diffusion models excel at maintaining visual coherence in repairs
-----
📊 Results:
→ 12.7% lower FID score than previous methods
→ 11.7% lower LPIPS score for visual quality
→ 81.9% character recognition accuracy, 11.9% higher than baseline
------
Are you into AI and LLMs❓ Join my daily AI newsletter. I will send you 7 emails a week analyzing the highest signal AI developments. ↓↓
🎉 https://rohanpaul.substack.com/
Share this post