Summary: This README provides a comprehensive overview of a hybrid ancient document restoration framework. The project emphasizes a low-cost, verifiable approach that combines classical mathematical algorithms with targeted Deep Learning (DL) assistance, avoiding the "black-box" nature of end-to-end AI generation.

---

HYBRID FRAMEWORK FOR ANCIENT DOCUMENT RESTORATION

A Low-Cost Approach Combining Classical Geometry and Deep Learning

1. PROJECT OVERVIEW

This project presents a specialized pipeline for the restoration of degraded and physically distorted ancient documents. Unlike contemporary "End-to-End" AI models that generate images from scratch—often leading to historical hallucinations—this framework utilizes AI strictly for structural feature extraction.

The core philosophy is AI-as-an-Assistant:

• Minimal Computational Cost: Only a lightweight model is used for mask detection.
• Geometric Integrity: Restoration is handled by deterministic mathematical transformations.
• Verifiability: Each stage of the pipeline produces measurable, traceable results similar to a scientific verification study.

Design Philosophy: Why not End-to-End AI?

• Most modern restoration projects use Generative AI (GANs/Diffusion) which often "hallucinates" or creates fake details in ancient scripts. Our project uses a Geometric Constraint Approach:

  • AI for Perception: UNet only identifies where the text is.

  • Math for Transformation: TPS and Polynomial fitting ensure the original pixels are simply moved back to their rightful place, preserving 100% historical authenticity.

2. RESTORATION PIPELINE

The restoration process is executed through four distinct, sequential stages:

graph TD
    A[Degraded Image] --> B[Stage 1: Deskewing]
    B --> C[Stage 2: Dewarping]
    
    subgraph "AI-Guided Geometry"
    C --> C1[UNet Masking]
    C1 --> C2[Skeletonization]
    C2 --> C3[Curve Fitting]
    C3 --> C4[TPS Warp]
    end
    
    C4 --> D[Stage 3: Forensic Analysis]
    D --> E[Stage 4: Binarization]
    E --> F[Restored Document]

    style B fill:#ffffff,stroke:#333,stroke-width:2px
    style C fill:#ffffff,stroke:#333,stroke-width:2px
    style D fill:#ffffff,stroke:#333,stroke-width:2px
    style E fill:#ffffff,stroke:#333,stroke-width:2px

Loading

Stage 1: Deskewing

To correct global rotation, the system employs the Probabilistic Hough Transform. By detecting the dominant orientations of text-line segments, the algorithm calculates the precise skew angle and performs a compensatory rotation to align the document to a horizontal baseline.

Stage 2: Dewarping (AI-Guided Geometric Correction)

This stage rectifies non-linear distortions (e.g., page curls and folds) using a multi-step geometric process:

1. Mask Detection: A Deep Learning model (U-Net) identifies the precise pixel-area of text lines.
2. Skeletonization: The detected masks are reduced to one-pixel wide centerlines (skeletons).
3. Curve Fitting: Polynomial or spline-based functions are fitted to these skeletons to model the physical warp of the paper.
4. TPS Dewarp: Thin Plate Spline (TPS) transformation is applied to warp the entire image back into a flat, rectified plane based on the fitted curves.

Stage 3: Forensic Analysis (Background Separation)

To isolate text from stains, aging artifacts, and uneven lighting:

• Division Normalization (DN): Estimating the illumination layer and dividing the original image by it to achieve a uniform background.
• Enhancement: Implementation of CLAHE (Contrast Limited Adaptive Histogram Equalization) or ZCA Whitening (Zero-phase Component Analysis) to sharpen faint ink traces and decorrelate noise.

Stage 4: Mask-Driven Binarization

The final stage converts the image to high-contrast black and white:

• The system utilizes the AI-generated Mask from Stage 2 as a spatial filter.
• Content Preservation: Pixels within the mask boundaries undergo adaptive binarization to retain stroke details.
• Background Cleaning: All pixels outside the mask are programmatically set to pure white (), ensuring a perfectly clean output for OCR engines.

3. TRAINING PERFORMANCE

The AI component is trained solely to identify text-line masks, significantly reducing the required training data and time compared to generative models.

Metric	Value
Training Loss	0.0449
Validation Loss	0.0636
Mask mIoU	0.8562
Mask F1 Score	0.9223
Mask PA	0.988

4. EXPERIMENTAL RESULTS

MASK AI

Visual Restoration Results

The following section demonstrates the transition from a distorted, low-contrast original to a rectified, binarized output.

• (Recommended: Comparison of Original -> Dewarped -> Forensic -> Binarized)

Quantitative Evaluation

We evaluate the quality of the restoration by measuring the accuracy of text extraction using two primary metrics: Character Error Rate (CER) and Word Error Rate (WER).

Processing Stage	CER (%)	WER (%)
Original Scan	247	130
Post-Restoration	110	95

5. REPOSITORY STRUCTURE

• src/core/deskewer.py: Implementation of Probabilistic Hough Transform.
• src/core/dewarp.py: Logic for Skeletonization, Curve Fitting, and TPS.
• src/core/forensic.py: Division Normalization and ZCA modules.
• src/core/ai_model.py: Architecture for the Mask Detection model.
• src/utils/metrics.py: Calculation tools for CER and WER.

6. CONCLUSION

By constraining Deep Learning to structural detection and relying on classical mathematics for image transformation, this project provides a robust, low-cost, and transparent solution for document restoration. This hybrid approach ensures that the historical "truth" of the document is preserved without the artifacts typically introduced by purely generative AI.

7. Credits

• Developed by:
• Nguyen Minh Quang - University of Science, VNU. https://github.com/minhquang0407
• Dinh Nhat Tan - University of Science, VNU. https://github.com/Hecquyn175
• Nguyen Quoc Anh Quan - University of Science, VNU. https://github.com/nqaq2005
• Le Nguyen Bao Thi - University of Science, VNU. https://github.com/Wis2411

---