Multiscale Document Segmentation

In this paper, we propose a new approach to document segmentation which exploits both local texture characteristics and image structure to segment scanned documents into regions such as text, background, headings and images. Our method is based on the use of a multiscale Bayesian framework. This framework is chosen because it allows accurate modeling of both the image characteristics and contextual structure of each region. The parameters which describe the characteristics of typical images are extracted from a database of training images which are produced by scanning typical documents and hand segmenting them into the desired components. This training procedure is based on the expectation maximization (EM) algorithm and results in approximate maximum likelihood (ML) estimates of the model parameters for region textures and contextual structure at various resolutions. Once the training procedure is performed, scanned documents may be segmented using a fine-to-coarse-to-fine procedure that is computationally efficient.