Performance evaluation of classifiers for the recognition of offline handwritten Gurmukhi characters and numerals: a study

摘要

Classification is a process to pull out patterns from a number of classes by using various statistical properties and artificial intelligence techniques. The problem of classification is considered as one of the important problems for the development of applications and for efficient data analysis. Based on the learning adaptability and capability to solve complex computations, classifiers are always the best suited for the pattern recognition problems. This paper presents a comparative study of various classifiers and the results achieved for offline handwritten Gurmukhi characters and numerals recognition. Various classifiers used and evaluated in this study include k-nearest neighbors, linear-support vector machine (SVM), RBF-SVM, Naive Bayes, decision tree, convolution neural network and random forest classifier. For the experimental work, authors used a balanced data set of 13,000 samples that includes 7000 characters and 6000 numerals. To assess the performance of classifiers, authors have used the Waikato Environment for Knowledge Analysis which is an open source tool for machine learning. The performance is assessed by considering various parameters such as accuracy rate, size of the dataset, time taken to train the model, false acceptance rate, false rejection rate and area under receiver operating characteristic Curve. The paper also highlights the comparison of correctness of tests obtained by applying the selected classifiers. Based on the experimental results, it is clear that classifiers considered in this study have complementary rewards and they should be implemented in a hybrid manner to achieve higher accuracy rates. After executing the experimental work, their comparison and analysis, it is concluded that the Random Forest classifier is performing better than other recently used classifiers for character and numeral recognition of offline handwritten Gurmukhi characters and numerals with the recognition accuracy of 87.9% for 13,000 samples.