Automatic recognition of off-line handwritten Arabic (Indian) numerals using support vector and extreme learning machines

This paper describes a technique using Support Vector (SVM) and Extreme Learning Machines (ELM) for automatic recognition of off-line handwritten Arabic (Indian) numerals. The features of angle, distance, horizontal, and vertical span are extracted from these numerals. The database has 44 writers with 48 samples of each digit totaling 21120 samples. A two-stage exhaustive parameter estimation technique is used to estimate the best values for the SVM parameters for this application. For SVM parameter estimation, the database is split into 4 subsets: three were used in training and validation in turn, and the fourth for testing. The SVM and ELM classifiers were trained with 75% of the data (i.e. the first 33 writers) and tested with the remaining data (i.e. writers 34 to 44) by using the estimated parameters. The recognition rates of SVM and ELM classifiers at the digit and writers' levels are compared. The training and testing times of SVM and ELM indicate that ELM is much faster to train and test than SVM. The classification errors are analyzed and categorized. The recognition rates of SVM and ELM classifiers are compared with Hidden Markov Model (HMM) and the Nearest Mean (NM) classifiers. Using the SVM, ELM, HMM and NM classifiers the achieved average recognition rates are 99.39%, 99.45%, 97.99% and 94.35%, respectively. ELM and SVM recognition rates are better for all the digits than HMM and NM. The presented technique, using simple features and SVM and ELM classifiers, are effective in the recognition of handwritten Arabic (Indian) numerals, and it is shown to be superior to HMM and NM classifiers for all digits.