Handwriting recognition by analyzing topological features and geometrical properties of characters

Abstract

With the widespread use of computers in this digital world, handwriting recognition is important as it enables us to still use handwriting as a means of communication. In this work an off-line handwriting recognition technique that identifies characters based on their topological features and geometrical properties, is described. The technique is able to create graph templates of all characters and then use these templates to identify unseen characters. For the training and testing of characters, the MNIST character database is used. The idea is to skeletonize a digitized image by applying line thinning techniques and then by using feature templates, the edges and other topological features that uniquely identify a character, are identified. With the help of these features which are treated as nodes and edges, a planar graph is created. These graphs are compared with each other using graph isomorphism. In the training phase, if a graph template is isomorphic with another graph template, it is ignored, otherwise if it is unique, it is stored. On the other hand, in the testing phase, if the tested graph is isomorphic with another template created during the training phase, the character class type is determined. A number of problems including isomorphic clashes are diagnosed in this work and a number of clash resolution techniques are suggested and implemented. A 62.23 recognition rate is achieved, but the results discussed during evaluation are quite promising.