Document image analysis

Scientific goals

I am interested in models addressing fundamental problems in document image analysis: restoration, segmentation, and structure recognition. Machine learning plays a central role in this research, as do structured models and graphical models. One of our current goals concerns the integration of structured information and structured terms into deep networks.

Recognition of document structure

PhD thesis of Bastien Moysset (10/2014-). Goal: learning deep models for the automatic recognition of document structure taking into account context and auto-context. Collaboration with A2IA (co-supervised with Jérôme Louradour, A2IA).

Evaluation

We propose a performance metric for text detection and for document segmentation which is based on clear separation of quality measurements and quanitity measurements. The metric comes with easily interpretable measures as well as performance graphs which illustrate the behavior of an algorithm as graphs showing quantity as a function of quality.

The metric (and corresponding publically available software DetEval are widely used in the document community and have been used as standard metric to evaluate participants of the following competitions:

  • ICDAR 2015 Robust reading competition [WWW]
  • ICDAR 2013 Robust reading competition [WWW]
  • ICDAR 2011 Robust reading competition [WWW]
  • Image Eval 2007 text detection competition [REF]
  • ICDAR 2003 Robust Reading Competition [REF]
Papers:
  • Christian Wolf and Jean-Michel Jolion. Object count/Area Graphs for the Evaluation of Object Detection and Segmentation Algorithms, In International Journal on Document Analysis and Recognition , 8(4):280-296, 2006.
Software:

Document restoration

Restoration of degraded documents with computational models taking into account degradation processes and a priori knowledge on the structure. C++ code for dual layer MRFs is publically available.

Papers:
  • Christian Wolf. Document Ink bleed-through removal with two hidden Markov random fields and a single observation field. In IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 32(3):431-447, 2010.
  • Christian Wolf and Gerald Gavin. Inference and parameter estimation on hierarchical belief networks for image segmentation. In Neurocomputing 73(4-6):563-569, 2010.
  • Christian Wolf, Improving recto document side restoration with an estimation of the verso side from a single scanned page. In International Conference on Pattern Recognition (ICPR), 2008.
  • Christian Wolf and David Doermann. Binarization of Low Quality Text using a Markov Random Field Model. In International Conference on Pattern Recognition (ICPR), 2002.

Document segmentation and binarization

Our adaptive document binarization method has placed 5/43 in the 2009 DIBCO competition and is widely used in the community. C++ Code is publically available.

Papers:
  • Christian Wolf and David Doermann. Binarization of Low Quality Text using a Markov Random Field Model. In International Conference on Pattern Recognition (ICPR), 2002.
  • Christian Wolf , Jean-Michel Jolion and Francoise Chassaing. Text Localization, Enhancement and Binarization in Multimedia Documents. In International Conference on Pattern Recognition (ICPR), 2002.
  • Christian Wolf and Jean-Michel Jolion. Extraction and Recognition of Artificial Text in Multimedia Documents. In Pattern Analysis and Applications, 6(4):309-326, 2003.

Contact

Christian Wolf
LIRIS UMR CNRS 5205
INSA-Lyon / Université de Lyon