Journal of Intelligent Systems (Jun 2019)
M-HMOGA: A New Multi-Objective Feature Selection Algorithm for Handwritten Numeral Classification
Abstract
The feature selection process is very important in the field of pattern recognition, which selects the informative features so as to reduce the curse of dimensionality, thus improving the overall classification accuracy. In this paper, a new feature selection approach named Memory-Based Histogram-Oriented Multi-objective Genetic Algorithm (M-HMOGA) is introduced to identify the informative feature subset to be used for a pattern classification problem. The proposed M-HMOGA approach is applied to two recently used feature sets, namely Mojette transform and Regional Weighted Run Length features. The experimentations are carried out on Bangla, Devanagari, and Roman numeral datasets, which are the three most popular scripts used in the Indian subcontinent. In-house Bangla and Devanagari script datasets and Competition on Handwritten Digit Recognition (HDRC) 2013 Roman numeral dataset are used for evaluating our model. Moreover, as proof of robustness, we have applied an innovative approach of using different datasets for training and testing. We have used in-house Bangla and Devanagari script datasets for training the model, and the trained model is then tested on Indian Statistical Institute numeral datasets. For Roman numerals, we have used the HDRC 2013 dataset for training and the Modified National Institute of Standards and Technology dataset for testing. Comparison of the results obtained by the proposed model with existing HMOGA and MOGA techniques clearly indicates the superiority of M-HMOGA over both of its ancestors. Moreover, use of K-nearest neighbor as well as multi-layer perceptron as classifiers speaks for the classifier-independent nature of M-HMOGA. The proposed M-HMOGA model uses only about 45–50% of the total feature set in order to achieve around 1% increase when the same datasets are partitioned for training-testing and a 2–3% increase in the classification ability while using only 35–45% features when different datasets are used for training-testing with respect to the situation when all the features are used for classification.
Keywords