Doctoral Seminar: Mohammad Reza Ameri

Speaker: Mohammad Reza Ameri

Supervisor: Dr. T. D. Bui

Supervisory Committee:
Drs. N. Kharma, A. Krzyzak, C. Y. Suen

Title:  Graph-based Keyword Spotting in Handwritten Documents

Date: Thursday, March 17, 2016

Time: 11:40am

Place: EV 3.309

ABSTRACT

Handwritten document analysis is one of the most challenging tasks in pattern recognition. The problem comes from the large number of word classes and the high variability of character shapes. Full recognition of handwritten documents is not as practical as it is for machine printed documents. To recognize the content of these documents the word spotting approach is suggested rather than transcription based methods.

Dynamic time warping (DTW) is one of the most widely used state-of-the-art approaches to template based word spotting. In keyword spotting by DTW, the two-dimensional word images are converted to one-dimensional feature vector sequences. The keyword template and the candidate image segments are then compared with the DTW distance of their corresponding feature vectors.

We propose template based word spotting method using a graph representation of the words. Graph representation would preserve the structure of handwritten words that might be ignored in feature extraction process of classic methods. Our goal is to preserve the two-dimensional structure of the handwritten script. Graph-based representation of word images and graph edit distance algorithms are proposed alternatives of feature vectors and the DTW distance in this context.