A Pattern-Based Method for Document Structure Recognition

One of the main goals of the CIDRE1 project is the design of an interactive document recognition system able to improve with use. In previous work, members of this project have already described software architecture issues [1, 6] as well as font and logical structure recognition algorithms [8, 2]. In this paper, we present a new method for document structure recognition based on geometrical and logical layout analysis. Most traditional methods for layout analysis are restricted to simple document structures (see [4] for an overview of these methods). However, recent works show that complex layout analysis such as newspaper page decomposition have also to be considered [3] and, in particular, illustrate the needs of using learning-based algorithms [5] for these tasks. The method we propose in this paper is both adaptive and interactive. As we will see, this combination of learning techniques and user-interaction helps preventing repetitive errors and supports the handling of complex and previously unknown document structures. The paper is organized as follows. In Section 2, we describe our new method for document structure recognition. In Section 3, we describe two potential applications of this method for classification and segmentation of document structure objects. Section 4 gives some preliminary results concerning the evaluation of the method for classification tasks. Finally, we conclude with directions for future research.