Projection-based features: A superior domain for handwritten Bangla basic characters recognition

Abstract

In spite of wide use of projection-based features in handwritten character recognition of several languages, its implementation was somewhat scanty in Bangla handwritten character recognition. This paper introduces the usage of projection profile features in recognizing handwritten Bangla basic characters. Alongside it also demonstrates a qualitative and quantitative analysis to visualize the effect of using projection based features on accuracy of recognition of Bangla handwritten characters through a number of approaches. In fact, this particular effort comprises of five different approaches where first one used longest-run, quad-tree and octant centroid features, second one adopted additional shadow features in association with the features of first approach, third one used longest run, quad-tree, shadow and chain code histogram features, next approach used longest-run, quadratic center of mass, shadow and left projection profile features and finally fifth approach with additional right projection profile features along with other features involved in the fourth approach. Throughout this analysis, neural network (trained via back-propagation algorithm) acted as classifier to observe the change in accuracy of recognition. It is seen that, with the increase in number of projection-based features, percentage of accuracy enhances at a greater rate than in case of inclusion of other features. This effective analysis can certainly assist a researcher to choose the optimal feature vector (consisting of several feature sets) for handwritten Bangla basic characters recognition.