Arbitrary Keyword Spotting in Handwritten Documents

Despite the existence of electronic media in today’s world, a considerable amount of written communications is in paper form such as books, bank cheques, contracts, etc. There is an increasing demand for the automation of information extraction, classification, search, and retrieval of documents. The goal of this research is to develop a complete methodology for the spotting of arbitrary keywords in handwritten document images. We propose a top-down approach to the spotting of keywords in document images. Our approach is composed of two major steps: segmentation and decision. In the former, we generate the word hypotheses. In the latter, we decide whether a generated word hypothesis is a specific keyword or not. We carry out the decision step through a two-level classification where first, we assign an input image to a keyword or non-keyword class; and then transcribe the image if it is passed as a keyword. By reducing the problem from the image domain to the text domain, we do not only address the search problem in handwritten documents, but also the classification and retrieval, without the need for the transcription of the whole document image. The main contribution of this thesis is the development of a generalized minimum edit distance for handwritten words, and to prove that this distance is equivalent to an Ergodic Hidden Markov Model (EHMM). To the best of our knowledge, this work is the first to present an exact 2D model for the temporal information in handwriting while satisfying practical constraints. Some other contributions of this research include: 1) removal of page margins based on corner detection in projection profiles; 2) removal of noise patterns in handwritten images using expectation maximization and fuzzy inference systems; 3) extraction of text lines based on fast Fourier-based steerable filtering; 4) segmentation of characters based on skeletal graphs; and 5) merging of broken characters based on graph partitioning. Our experiments with a benchmark database of handwritten English documents and a real-world collection of handwritten French documents indicate that, even without any word/document-level training, our results are comparable with two state-of-the-art word spotting systems for English and French documents