Extracting Maya Glyphs from Degraded Ancient Documents via Image Segmentation

Rui Hu, Jean-Marc Odobez, Daniel Gatica-PerezACM Journal on Computing and Cultural Heritage 2017 · DOI

TL;DR

This paper introduces an automatic system for extracting Maya hieroglyph strokes from severely degraded ancient documents, a task previously requiring extensive manual effort. By combining multi-resolution super-pixel segmentation with a Conditional Random Field (CRF) model, the system achieves robust glyph extraction, enabling comparable performance in glyph retrieval tasks to manually segmented data.

What Problem Does This Paper Address?

Ancient Maya codices, invaluable historical artifacts, are often in poor condition due to age, fading ink, and staining. Extracting legible glyphs from these degraded documents is crucial for their analysis and decipherment, but traditional image processing techniques struggle with the high variability in stroke width, brightness, and the pervasive background noise. Epigraphers currently spend considerable time manually segmenting these glyphs, a process that is not scalable for large datasets. This paper aims to automate this labor-intensive process, making Maya hieroglyphic analysis more accessible and efficient.

Methods

The proposed system employs a multi-stage approach:

  • Super-pixel Segmentation: The image is first over-segmented into super-pixels using the Simple Linear Iterative Clustering (SLIC) algorithm. This creates compact, homogeneous regions that help preserve local details while smoothing background noise. Crucially, the system utilizes multiple resolutions of super-pixel segmentation to accommodate the wide range of stroke widths present in the glyphs.
  • Super-pixel Classification: A Support Vector Machine (SVM) classifier is trained to assign a probability to each super-pixel region belonging to a foreground glyph stroke. This classification is performed across multiple super-pixel resolutions.
  • Probability Map Aggregation: Pixel-wise probability maps generated from different super-pixel resolutions are aggregated to create a more robust representation that balances stroke detail preservation with background noise reduction.
  • Fully Connected CRF: A fully connected Conditional Random Field (CRF) model is then applied to enforce label consistency across pixels, further refining the segmentation and smoothing out imperfections. This is a key component for improving the overall quality of the extracted glyph strokes.

Datasets

The paper utilizes two primary datasets:

  • Maya Glyph Block Binarization (MGBB) dataset: This dataset consists of cropped raw images of Maya glyph blocks and their corresponding manually generated "clean raster" forms. It includes 414 glyph blocks from the Dresden, Madrid, and Paris codices, with varying quality rankings. This dataset is used for training and evaluating the segmentation accuracy.
  • Maya Codex (MC) dataset: This dataset comprises 759 individual glyphs, segmented from 296 blocks, and annotated with labels from the Thompson catalog. This dataset is used for evaluating the glyph retrieval performance.

The authors also mention sharing these datasets for academic research, serving as benchmarks for image segmentation and shape analysis.

Key Insights and Ideas

  • Multi-resolution Super-pixels are Key: The inherent variability in Maya glyph stroke widths necessitates a multi-resolution approach. Using a single super-pixel scale leads to either losing fine details or retaining too much noise.
  • CRF for Label Consistency: The fully connected CRF model significantly improves the final segmentation by ensuring that neighboring pixels are labeled consistently, effectively cleaning up the output from the appearance-based classification.
  • Segmentation Quality Directly Impacts Retrieval: The paper demonstrates that the quality of glyph segmentation has a direct impact on downstream tasks like glyph retrieval. Their automatically segmented glyphs achieve retrieval performance comparable to manually segmented ones, validating the system's utility.
  • Challenging Data Requires Sophisticated Methods: Simple thresholding methods fail on these degraded documents. The combination of super-pixels, SVM, and CRF is essential for handling the complex noise and variations.

Key Results

  • Segmentation Accuracy: The proposed method achieves the best segmentation results (highest F-measure) compared to baseline methods like K-means, GMM, Otsu's algorithm, and a saliency detection framework. The F-measure for the proposed method reaches up to 93.76%.
  • Retrieval Performance: When using the automatically segmented binary images, the glyph retrieval system achieves comparable results to using manually segmented clean raster images. Specifically, with glyph co-occurrence information, around 30% of queries achieve their correct match in the first returned result, and 63% in the top 10, using the proposed method. This is a significant improvement over thresholding-based methods.
  • Qualitative Improvements: The system effectively removes background noise and extra strokes while preserving delicate glyph stroke details, leading to visually superior results compared to other methods.

Limitations and Future Work

  • Dataset Size: The learned model is trained on a relatively small, manually cleaned glyph collection.
  • Problematic Queries: Retrieval performance can be significantly affected by queries with very low visual quality or large visual differences from catalog examples.
  • Future Work: The authors suggest comparative analysis with other data sources (e.g., Maya monument data, Chinese calligraphy), automatic Maya document layout analysis, and segmenting individual glyphs from blocks for more advanced functionalities.

Summary updated

BibTeX
@article{hu2017extracting,
  author = {Rui Hu and Jean-Marc Odobez and Daniel Gatica-Perez},
  title = {{Extracting Maya Glyphs from Degraded Ancient Documents via Image Segmentation}},
  journal = {ACM Journal on Computing and Cultural Heritage},
  year = {2017},
  doi = {10.1145/2996859},
}
Figure