Registration is closed.


The following speakers have already confirmed their participation to the summer school

Academic talk

Andreas Fischer Diva Group, University of Fribourg, Switzerland

Structural Methods for Handwriting Analysis

In this talk, I will provide an overview of recent structural pattern recognition methods in the field of handwriting analysis. Topics include graph-based representation formalisms and graph matching methods for handwriting recognition, keyword spotting, interactive layout analysis, and signature verification. Afterwards, I will seize the opportunity to elaborate on the kinematic theory of rapid human movements and its ability to decompose complex handwriting signals into elementary movements. As an outlook, recent attempts towards deep learning for graph-based handwriting representations will be addressed.

The lecture is accompanied by a lab session, where the participants will get hands-on experience with graph-based keyword spotting in historical manuscripts.

David Doermann University of Maryland, USA

Computer Forensic in documents

description soon

Bart Lamiroy Loria, University of Loraine, France

Honest, Reproduciple Reporting of Results is Difficult ... an Analysis and some Good Practices

A big part of the research activity in Document Analysis relates to designing, improving and comparing classifiers of different types. Making sure the resulting work can be reused by the community in a knowledgeable way depends on reliable reporting, reproducible experimental protocols and a keen understanding of the limits of comparative tools and frameworks.
In this session we shall conduct a critical analysis of current practices, and introduce some ways of mitigate collateral damage resulting from inadequate experimental reporting.

Jean-Marc Ogier L3i, University of La Rochelle, France

General Introduction on TC10/TC11 research topics


Jean-Yves Ramel Lifat, University of Tours, France
  Interactive approaches and techniques for Document Image Analysis

After showing the numerous drawbacks of fully automatic systems, my talk will provide an overview of some approaches and techniques that can be used to introduce more interaction in document image analysis systems. First, based on research done on historical documents, I will explain how document content should be represented to allow a user-driven content extraction and layout analysis. Next, the interest of incremental learning and classification techniques will also be explained and illustrated with several examples. Formalisms and architectures that let users define the concepts to extract or recognize in document images, the associated training samples as well as the features to use will be introduced.
Finally the interest of anytime and budgeted methods, as well as future trends in interactive systems for document analysis will be discussed to conclude the talk.
Marçal Rusiñol CVC, Universitat Autònoma de Barcelona,Spain

Large-scale Document Indexing

Document indexing is the technique aimed at the retrieval of documents from repositories that might contain millions of documents. Within the field of Document Analysis and Recognition, documents are often indexed either by their full-text content or by their visual appearance. We will overview some basic textual and visual description techniques of documents, and have a glimpse on state of the art indexing strategies that will allow an efficient storage of indices for a later retrieval of documents.

Besides the state of the art overview, we will have a chance to do a hands-on session to implement an indexing strategy using Python and numeric libraries. 

Seiichi Uchida Human Interface Laboratory, Kyushu University, Japan

Machine learning for document analysis and understanding

As everyone knows, machine learning (ML) is one of the most important techniques for document analysis and understanding (DAR). In fact, DAR has been improved by many ML techniques. Recently, deep neural networks (DNN) and their versions have made drastic improvements not only quantitatively but also qualitatively. In other words, DNN extend the horizon of DAR research. In this lecture, various new ML applications for DAR tasks are introduced, along with a simple explanation of the mechanism of DNN. 

Dimosthenis Karatzas CVC, Universitat Autònoma de Barcelona,Spain

Scene text Understanding

The ability of machines to read text in unconstrained settings such as scene images and videos has been an important challenge for the computer vision for the past 20 years. This lecture will give an overview on robust reading systems, with a special focus on the current state of the art on scene text understanding.
During this lecture we will review the evolution of text understanding systems, with a special focus on current trends and open problems. We will study specific deep network architectures for text localisation, rectification, script identification, word spotting, open recognition, and retrieval. We will demonstrate the importance of visual context for scene text understanding, and study how we can jointly model visual and textual information - then we will briefly venture to other related application areas in computer vision.
Finally, the lecture will give an overview of the last seven years of Robust Reading Competition activity, reviewing existing datasets and their limitations, open tools and methodologies for performance evaluation. Lessons to be learnt after receiving over 17000 submissions to the Robust Reading Competition to date will be outlined and an outlook to the future of robust reading will be offered.


Industrial talks

Vincent Poulain d'Andecy YOOZ, France

Industrial perspectives in document analysis : Problematics and Best practices for partnerships

description soon



Application Colloque - L'application de gestion des colloques est un projet soutenu historiquement par l'Union Européenne dans le cadre du programme Innova-TIC.
Legal Notice