Object-Based Classification of Mixed-Mode Images

Abstract

This paper presents an efficient algorithm of classifying mixed-mode images into “objects” of rectangular shape and with parent-children relationship. We consider four different classes of “objects”: background, text, graph, and photograph. The classification algorithm has the hierarchical nature, i.e., it tries to (1) classify background regions and non-background regions, (2) classify bi-level objects and multi-level objects (in non-background regions), and (3) classify graph and photograph (in multi-level objects). During the classification, a merging-and-splitting refinement is used for boundaries and small fragments. Excellent classification results have been observed in all experimental tests.

This work has been supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China.