Abstract
The relevance of feature selection technique is found in decreasing the dimensionality of the feature set by rejecting the irrelevant and redundant features. The feature selection technique considers only relevant and significant features which are required to classify the word image into one of the multiple classes. In this article, four feature selection techniques such as principal component analysis (PCA), correlation feature set (CFS), consistency-based analysis (CBA) and chi squared attribute (CSA) are used to optimize the boundary extent features extracted from the word images. The proposed approach is used to recognize offline handwritten Gurumukhi place names based on holistic approach which recognize the whole word as an independent entity without going for its explicit segmentation. The experiments are evaluated on benchmark dataset of 40,000 handwritten words of Gurumukhi script using two classification techniques, namely decision tree and random forest classifiers. To evaluate the performance of different feature selection techniques, various evaluation parameters are used. Based on experiments, CSA feature selection technique achieved the best results using random forest classifier and surpassed the other three feature selection techniques in terms of various considered evaluation parameters. CSA feature selection technique along with random forest classifier achieved best recognition rate of 87.42%.
\
