A Region-based System for the Automatic Evaluation of Page Segmentation Algorithms

A method for automatically evaluating the quality of document page segmentation algorithms is described. Page segmentation involves decomposing a page into its structural and logical units such as paragraphs, halftones, captions and tables. These units are then ordered and logically associated. These two steps are very important in a document recognition strategy. Many di erent techniques have been proposed in literature to segment document pages, but since no robust method to benchmark them is available, it is impossible to evaluate them reliably. Our proposed page segmentation benchmarking strategy is a region-based approach, in which segmentation results are compared with manually generated "ground truth les", containing a synthetic description of all possible correct segmentations. It is described in detail and we give a set of guidelines to generate ground truth les and propose a general format to represent them. The algorithm is simple and fast, and provides a multi-level output for each segmentation, from the page level to the region level.