Performance Evaluation of Optimised Backpropagation Algorithms for Yorùbá Character Feature Extraction and Recognition

Abstract

Character recognition has been an important area of research in the last few decades. It is basically divided into two major types namely online and offline (handwritten) character recognitions. Characters with tonal marks (diacritics) such as Yorùbá characters (orthography) had posed more challenges than their counterparts with no tonal marks and as a result require some optimization methods to improve the recognition rate and reduce the error rate. This study evaluated the performance of four optimized backpropagation algorithms, Levenberg-Marquardt, Quasi-Newton BFGS, Resilient Propagation and Scaled Conjugate Gradient, on Yorùbá character recognition. The method used in this study involves the five basic stages of image processing namely; image acquisition, image preprocessing, segmentation, feature extraction and classification. The performances of the algorithms were experimentally measured using mean squared error (MSE), epochs, accuracy and response time. From the experiments, it was observed that the Levenberg-Marquardt training algorithm has the best accuracy of 98.8%; Resilient Propagation and Scaled Conjugate Gradient are the fastest to converge with an average response time of 2 seconds. The results obtained can serve as a fundamental guideline in adopting the most relevant training algorithm for character image recognition.