READ - Recognition and Document Analysis
by Liliane Peters and Ashutosh Malaviya
The project READ aims at increasing the efficiency of the current "Recognition and Document Analysis" technology. The main goal in READ is to combine and refine document analysis techniques, ranging from low level picture processing over document structure analysis to linguistic extraction, into a general framework.
The objects to be recognized in READ are examples from the three main domains of document analysis: addresses, forms and documents. The combination and integration of the expertise of the project partners in these domains is both a promising but also a challenging task. READ is funded by the German Federal Ministry of Education, Science, Research and Technology.
To achieve the pursued goals, the implementation of intelligent systems is required. These systems should ensure robustness towards data errors and adaptivity towards unknown data. The progress in document recognition and analysis methodologies is expected to be realised through the following research activities:
- acquisition of documents
- object extraction and recognition
- modeling and interpretation.
The outcome of these cooperative research activities will be integrated into a prototype application.
GMD is mainly involved in the activities related to object recognition, and benchmarking of object recognition systems modeling and interpretation.
The contribution of GMD Institute for System Design Technology is the development of robust object recognition methodologies. Soft computing methods, such as the combination of fuzzy logic and neural networks, are used to recognize cursive handwriting, as well as isolated words. The cognitive processing of segmented document objects is the focus of the proposed method. Fuzzy grammars are used to represent the unconnected and incomplete feature information of hand-written documents.
The FOHDEL language, which was developed by GMD scientists for on-line character recognition, is extended and applied to represent the fuzzy rules for word recognition. Neural networks supplement the fuzzy set theoretic techniques to generate the expert information automatically and thus generate the rule-bases for the recognition process. This capability is developed also to adapt the recognition system on-the-fly to new document environments.
A major activity of this work package will be the integration of the various object recognition approaches developed by GMD, Siemens EletroCom and University Koblenz into a unique toolbox.
The benchmarking activities of the project were divided into several working packages, corresponding to the various levels of document recognition. GMD Institute for System Design Technology will participate in the benchmarking activities related to object recognition. Therefore a server with reference data collected from the partners was made available.
By defining a common data interface and common test and verification data sets, all partners can test their algorithms, even during the development and implementation phases. This should reduce the complexities of final integration phase which comprises the combination of various object recognition approaches and will offer a better overview of the advantages of each algorithmic strategies. At the end of the project an improved recovery of classification errors is expected.
Liliane Peters - GMD
Tel: + 49 2241 14 2332
Ashutosh Malaviya - GMD
Tel: + 49 2241 14 2751