Région Lorraine,
CNET,
Institut
National Polytechnique de Lorraine, and
Communauté urbaine du Grand Nancy.
This book is a collection of selected papers from the workshop. At the end of the workshop, every presented paper was evaluated by two reviewers. As a result of this process, many of the papers were thoroughly revised and improved, so much so that we feel that this book has become much more than a mere post-workshop proceedings volume. However, we have mostly limited our work to the scientific editing; basically, the authors provided camera-ready copies and we haven't done much copy-editing.
The book is organized into several main topics. The first deals with vectorization and segmentation of scanned graphics documents. After a short introduction by Doermann, the paper by Wenyin and Dori presents the authors' work on generic recognition algorithms; the potential of their MDUS system was largely demonstrated during the contest held at the workshop. The two following papers explore the use of directional morphological filters for segmentation of graphics documents; these filters give promising results for separation of text and graphics. J. Y. Ramel proposes an original vectorization method, based on the decomposition of the image into quadrilaterals, which avoids the well-known artifacts introduced by thinning. Finally, Anne Vialard shows how the concept of Euclidean paths can be used for geometrical transformations of scanned characters.
The second topic of the book is symbol recognition. The first paper on this topic provides an overview of the state of the art, and lists a number of issues which are still open. The two following papers propose methods for recognizing symbols in architectural drawings. Finally, Francesconi et al. present their system for recognizing logos, using recursive neural networks.
Our impression is that the area of forms processing has matured a lot in recent years. This was clear from the very good results presented at the workshop. The first paper on this third topic lists a number of practical applications developed in Korea. The two following papers present very interesting results on text/graphics separation when the text is touching the lines of the forms. Another application, with many similarities, is that of bar-charts analysis, presented by Yokokura and Watanabe.
Significant progress has also been made in map processing. A team from the French national geographic institute (IGN) has been working for several years on the analysis of topographic maps; some of their results are presented in the first two papers on this topic. Equally promising results have been achieved in Switzerland, as shown by Frischknecht and Kanani. Color segmentation is often a problem in map analysis; Centeno proposes an approach for solving it. Schavemaker and Reinders deal with the problem of conflict resolution when several pieces of information must be merged in map interpretation. Finally, Roux and Maître show the potential of map analysis as a component of an aerial image interpretation system.
The interpretation of engineering drawings is also an important domain in graphics recognition. Tombre presents the state of the art in this area, and lists a number of challenges. The two other papers on this topic deal with the difficult problem of 3-D reconstruction from several views.
At the workshop, several systems were presented which solve complex graphics recognition problems. It is an indication of the maturity of a field that state of the art algorithms can be implemented and integrated into complete systems, working on real-life applications. Baum et al. present a graphics recognition systems for indexing of a large document database, developed at Boeing. Nagy et al. clearly show the potential of implementing the best possible methods from various sources, and integrating them into a partially-automated system. Arias et al. and Yagi et al. describe two other applications, in two very different areas: telephone company drawings on one hand, and artistic drawings on the other.
Performance evaluation was at the heart of many discussions during the workshop. It has become clear to everybody that significant progress in the robustness of graphics recognition algorithms necessitates the ability to clearly assess and evaluate the performance of methods and systems. Smeulders and de Boer present the general problem of performance evaluation in their first paper, and propose a method which combines real and synthetic data in their second article. The next two papers propose two different protocols for performance evaluation.
This leads us to the graphics recognition contest which was held during the workshop. The paper by Chhabra and Phillips, who organized this contest, introduces the challenges the participants had to meet, and presents the results of the contest. The book ends with a chapter summarizing the conclusions and recommendations from the numerous discussions and exchanges during the workshop. This chapter also includes a detailed report from a vivid debate about performance evaluation and the contest.
The workshop was supported by grants from
Région Lorraine,
CNET,
Institut
National Polytechnique de Lorraine, and
Communauté urbaine du Grand Nancy.
We owe special thanks to the contributing authors and to the reviewers of the papers. Anne-Lise Charbonnier and Armelle Demange, who were responsible for local arrangements, worked hard to make everything run very smoothly and pleasantly. Many thanks also to Isabelle Herlich and to Fabienne Perani, for efficient help with all the secretarial duties, both for the organization of the workshop and during the preparation of this book.
February 1998