Word spotting and character recognition of handwritten Hindi scripts by Integral Histogram of Oriented Displacement (IHOD) descriptor

This paper proposes efficient and robust segmentation-based word spotting and character recognition systems for offline handwritten Hindi scripts using a newly proposed shape descriptor called integral histogram of oriented displacement (IHOD). IHOD descriptor is easy to construct, has high discriminating power and minimizes intra-class variance while enhancing inter-class variance. IHOD descriptor has been used to train multilayer perceptron (MLP) to obtain MLP-based word spotting and character recognition models. The word spotting model has been evaluated with 3 datasets of handwritten Hindi word images whereas the character recognition model has been evaluated with 2 datasets of handwritten Hindi character images. The proposed word spotting technique attained a mean average precision of 0.9736 and best k-precision of 0.9758. The proposed character recognition technique achieved best recognition accuracy of 97.45%. Performance of the proposed word spotting technique has been compared with CNN-based state-of-art techniques like AlexNet, DenseNet, VGG and Inception. In this work also, a new dataset of segmented handwritten Hindi word images has been created for use in development and testing of word spotting systems. The dataset consists of 43,572 instances of words from 415 different writers.