A multiscale domain-independent algorithm for document image segmentation

Document Image Segmentation is a crucial step in the conversion process for paper document images into electronic documents. Entities in a document image, such as text blocks, tables and figures need to be separated before further document analysis and recognition can occur. Many Document Segmentation algorithms are designed exclusively for a few specific document types, utilizing highly-specialized document models.; This thesis presents a domain independent segmenter which does not assume specific document layout models in its segmentation. The segmenter utilizes a minimal amount of image domain knowledge. Segmentation of graphic and text entities is based purely on their geometric attributes and tonal values. Entities from the document images are extracted as non-overlapping sub-images by the segmenter.; The segmenter is a general-purpose tool, which can be used for segmentation tasks when domain specific models would be inappropriate, for example, in the purposes of image retrieval. The output of the segmenter can also be used to identify the domain of a document. Subsequently an algorithm specific for that domain may be applied to the image to produce a refined segmentation. The segmenter can also act as a pre-segmenter to separate out document entities so that they can be resegmented by domain specific segmenters. Due to the general nature of the segmenter, it can also be used for segmenting natural images. Results of segmentation are shown on a diverse set of test images.