Document Enhancement for Improving Post-processing Tasks in Palm Leaf Manuscripts

Advanced Techniques for Image Restoration and Binarization

Nimol Thuon*, Jun Du*, Zhenrong Zhang, Jiefeng Ma

National Engineering Laboratory for Speech and Language Information Processing (NEL-SLIP), USTC, China

iFLYTEK Research, Hefei, Anhui, China

The Challenge of Manuscript Degradation

Palm leaf manuscripts are vital historical artifacts, but their organic nature makes them prone to significant degradation. Over time, factors like light exposure, handling, and storage conditions lead to a range of issues, including faded ink, physical damage, low contrast, and various forms of noise. These problems are often exacerbated by the limitations of digitization equipment.

The inherent complexity of Southeast Asian scripts, with their extensive character classes and intricate layouts, combined with poor image quality, makes automated analysis exceptionally difficult. Our research focuses on developing robust enhancement techniques to restore legibility and prepare these valuable documents for post-processing tasks like classification and recognition.

Low contrast manuscript

Our Enhancement Methodologies

The GTC (Generate, Transform, Clean) Strategy

GTC Framework Diagram

To address the dual problems of data scarcity and image degradation, we developed the GTC strategy. This framework uses two Generative Adversarial Networks (GANs) for data augmentation and image cleaning.

  • Generate: An enhanced DCGAN creates diverse and realistic manuscript backgrounds, providing varied contexts for training.
  • Transform: A multi-task algorithm artfully combines cleaned text with the newly generated backgrounds, creating a rich augmented dataset that captures real-world variations.
  • Clean: A customized transformer-based GAN, which we call PALM-GAN, performs the final binarization and enhancement, effectively removing noise and producing a clean, high-quality image ready for analysis.

IEPalm: A Multi-Task Image Processing Approach

The IEPalm framework is a targeted, multi-step process designed to improve character-level quality through careful normalization and thresholding.

  • Normalization: We first focus on balancing image contrast. Using an enhanced version of Contrast Limited Adaptive Histogram Equalization (CLAHE), we correct inconsistencies between text and background, reduce noise, and sharpen edges without over-enhancement.
  • Thresholding: Standard binarization methods often fail on low-contrast historical documents. We use an improved WAN thresholding technique that dynamically adjusts to the image's characteristics. By calculating a "maximum-mean" value, our method can better separate text from the background, even in faded or damaged areas, restoring lost features and reducing artifacts.
IEPalm Framework Diagram

How to Cite

Generate, Transform, and Clean... (PALM-GAN)

@article{thuon2024generate,
  title={Generate, transform, and clean: the role of GANs and transformers in palm leaf manuscript generation and enhancement},
  author={Thuon, Nimol and Du, Jun and Zhang, Zhenrong and Ma, Jiefeng and Hu, Pengfei},
  journal={International Journal on Document Analysis and Recognition (IJDAR)},
  pages={1--18},
  year={2024},
  publisher={Springer}
}

Improving Isolated Glyph Classification... (IEPalmV1)

@inproceedings{thuon2022improving,
  title={Improving isolated glyph classification task for palm leaf manuscripts},
  author={Thuon, Nimol and Du, Jun and Zhang, Jianshu},
  booktitle={International Conference on Frontiers in Handwriting Recognition},
  pages={65--79},
  year={2022},
  organization={Springer}
}