Rough Set Based Analysis of Document Images

Rough set is a well-studied subject with a theoretical foundation and many applications. However, its usage in image processing has been very sparse. Most of the well-known algorithms for document image processing related to character recognition, character spotting, and logo retrieval resort to supervised classification, causing the system to slow down in the speed with increasing diversity in the documents, as well as the need to have a large training dataset. Hence, with an aim to resolve the tediousness and pitfalls of training, but without compromising on the efficiency, we introduce a rough-set-theoretic model. It is designed to perform an unsupervised classification of optical characters and logos with a small subset of attributes, called the semi-reduct. The semi-reduct attributes are mostly geometric and topological in nature, each having a small range of discrete values estimated from different combinatorial characteristics of rough-set approximations. This eventually leads to quick and easy discernibility of almost all the characters and logos. In this thesis, we first explain the basics of rough set theory. Subsequently, we propose various attributes that can be easily computed from the binary representation of the images. In subsequent chapters we show how one can select an appropriate subset of such attributes, known as semi-reduct, to perform a document processing task. We demonstrate in this thesis that using the above attributes one can design a character recognition system that is both computationally and storage efficient. Using a different semi-reduct, we show that one can also solve the very delicate task of character spotting in ancient inscriptions. Additionally, we propose appropriate pre-processing steps to binarize the old and dilapidated inscriptions. Finally, we propose a novel technique for logo retrieval using a suitably prepared semi-reduct. Comparison with other existing techniques substantiates our claim that attributes from the rough set are indeed good candidates for document image processing.