Adaptive Threshold-Based Database Preparation Method for Handwritten Image Classification

Abstract

Problem-specific, well classified database is primary and most important requirement of all machine learning-based systems. Hand-written character classification and recognition system is also no exception to this. In terms of features the good database must have less intra-class variance and high inter-class variance. However, extracting unique features from digital handwritten character image (HCI) is one of the most challenging task. High variation in the writing style of the writers and similarity in the feature value between multiple classes of characters are the primary hurdles. Unfortunately this scenario leads to high false acceptance rate which interns results drastic decrease in accuracy of the classification and recognition system.

As far as handwritten character classification and recognition system is concerned to overcome this hurdle we have developed a simple but effective adaptive threshold-based database preparation method. In the proposed method, Adaptive Threshold Value ($AT_v$) is calculated based on the similarity score (SS) of an existing HCI images in the respective class. If the threshold value of a new sample is in acceptable range then it is added into the feature map of respective class. To verify the efficiency and accuracy of the proposed method, the series of experiments are conducted on two standard datasets (MNIST and VDM). For the experimentation high-level features extracted using Deep neural network (DNN) based architecture and the proposed adaptive threshold-based method is applied to place HCI to correct class. Experimental results state that, the proposed adaptive thresholding-based method produces promising results and reduced the false acceptance rate.