Periodic Report Summary - ADAO (Administrative Document Automate Optimization) Public organisations, social security services and large companies every day handle a large volume of various administrative documents: ID cards, forms, mails, etc. They look for help in process optimization, cost reduction and customer satisfaction. Automation is the answer. The French company ITESOFT has developed products to automate the classification and the data extraction from administrative documents. These products rely on key-technologies: Graph matching algorithms, Handwritten recognition, Document classification, Document segmentation, Colour image processing, etc. The Document Analysis and Pattern Recognition Group of the Computer Vision Center (CVC) in Barcelona has large R&D experience in the field of Document Image Analysis and Recognition. The group is active in research in symbol recognition, indexing and browsing by graphical content, sketch interfaces, diagrammatic reasoning and visual languages for graphic documents, graphic recognition architectures, reading systems for forms and structured documents, camera-based OCR. ITESOFT and the CVC develop together a knowledge transfer program driven by the optimization of an automatic document processing system. The goal of the program can be summarised in three main objectives: - transfer the knowledge of the ITESOFT application domain and open problems to the CVC by delivering each time a dataset of thousands of real samples (reflecting the statistics and variance of the real world); - implement dedicated demonstrators embedding CVC existing algorithms to evaluate how they behave and how they perform on the provided ITESOFT dataset; - transfer theoretical knowledge from CVC researchers to ITESOFT researchers who then learn new algorithms. We identified five knowledge areas to focus our work: Graph matching algorithms, Logo classification, Cursive classification, Document (graphical) classification, Graphic / Text overlapping segmentation in binary images and colour images. During the first period of the project we have successfully delivered four large datasets, with ground truth information and developed four demonstrators: - demonstrator to both classify and cluster documents based on a graph representation; - demonstrator to spot logos within a page, implementing two different logo classification approaches; - demonstrator combining ITESOFT and CVC algorithms for page segmentation and handwriting recognition; - demonstrator to classify structured documents by a BSM approach. The technical results up to now have been very interesting. In some cases, they demonstrate relevant technologies that are candidate to be integrated in ITESOFT next product release. This will contribute to maintain a technology leadership over the worldwide market. In other cases, they demonstrated technologies that cannot be directly applicable. This result is positive too because it answers questions that we would never be able to evaluate without this knowledge transfer program. The experience has also been very positive. This program mainly involved staff exchanges between the two partners. Both CVC and ITESOFT researchers remain enthusiastic in sharing their work and collaborating through the project. CVC researchers have learnt about new real-world problems and ITESOFT researchers have been trained to new algorithms and methodologies. In terms of dissemination the project is also progressing well, based on different actions: publications have been submitted to international conferences and journals, a workshop will be organized during the 11th International Conference on Document Analysis and Recognition (ICDAR), the website of the project is available at http://www.cvc.uab.es/adao. The next steps of the project will be focused on the difficult problem of Graphic / Text overlapping segmentation in binary and colour images.
