Spatial attributes-based segmentation and topological attributes-based recognition algorithm for Myanmar OCR

In this paper, we propose a novel segmentation and character recognition algorithm for printed offline Myanmar documents. The main contribution of this paper is threefold: 1) it firstly presents a segmentation algorithm based on spatial attributes of the characters; 2) it then extracts the most relevant features from segmented images using determinant and trace values of Hessian feature matrix. The feature vectors are fed into fully connected self-organisation map (SOM) computational network for recognition of those segmented images; 3) the system finally assembles partially recognised characters into a complete compound character depending on their topological attributes. To prove the performance of the system, we have conducted experiments with a dataset with 40,878 images and evaluate the performances in terms of accuracy, error rate and computational time by comparing with contemporary works CNeT and OCRMPD. Our system proves that we outperform 97.5% in overall accuracy than those in compared works.