The First IAPR TC3 Workshop on Neural Networks and Learning in Document Analysis and Recognition (NNLDAR05) was held in conjunction with the Eighth International Conference on Document Analysis and Recognition (ICDAR).  NNLDAR05 was co-sponsored by IAPR TC3 Neural Networks and Computational Intelligence, the University of Florence (Italy), and Hitachi Ltd. (Japan).  Dr. Simone Marinai and Dr. Hiromichi Fujisawa served as co-chairs, and a number of active researchers joined the program committee.

 

The background of NNLDAR05 is the widespread and successful applications of artificial neural networks (ANNs) and machine learning paradigms in document analysis.  ANNs have been applied to all the processing steps of document analysis, including image pre-processing, layout segmentation, text area classification, character segmentation, word and string recognition, and character classification.  The great success of ANNs is attributed to their flexible learning capability from examples.  The aim of NNLDAR05 was to provide a forum for discussing the start-of-the-art, existing problems, and potential improvements and new applications of neural networks and learning paradigms in document analysis.

 

The program of NNLDAR05 included seven oral presentations and one demonstration session.  About 30 participants attended the sessions and actively participated in discussions.  The topics of presentations covered the review of methods, data pre-processing and selection for improving generalization or learning speed, and improving rejection.  In addition, applications of learning methods to segmentation, document retrieval, and text classification were proposed.  Each presentation was assigned 30 minutes, including 20 minutes for speaking and 10 minutes for questions and discussions.  The demonstration session showed a complete document understanding system incorporating machine learning techniques in many steps.

 

The discussions at NNLDAR05 were really stimulating. Issues included the benchmarking of learning and recognition methods, reliable rejection, and potential new applications.  Particularly, the urgent need of common datasets, especially large-scale and difficult-to-recognize ones, was raised.  The recognition rate on the widely used MNIST handwritten digit dataset is already very high, but it was remarked that details of errors are required for analysis and comparison.  Error reduction via reliable rejection was considered to have received much less attention than improving recognition rate.  Hence, moving the objective from a high correct rate to low error and reject rates was emphasized.  Other discussed topics concerned the comparison between ANNs and support vector machines (SVMs), improving recognition performance via utilizing new learning schemes, combining different learning methods, incorporating human knowledge, and so on.  The discussions at NNLDAR05 thus encouraged the continuation of this forum.

Workshop Report:  NNLDAR05

Co-Chairs:

Simone Marinai

Hiromichi Fujisawa

Return to Home Page