Handwritten Tibetan Character recognition based on ELM using modified HOG features

Abstract

Among the recent variants of neural networks, Extreme learning machine (ELM) is widely popular in the domain of computer vision due to having short training time. Handwritten character recognition is one of the typical problem which has shown promising results using several neural networks. Classification of handwritten characters is primarily a two-step problem where we need to obtain the features from the data set and then applying a classifier to recognize the characters. We have proposed a Center of Gravity (CG) based quad-partitioning of images to extract Histogram of Oriented Gradients(HOG) features which are normally calculated over equally partitioned images. These modified HOG features are used to extract more local shape of the handwritten Tibetan characters, a newly introduced dataset. ELM is then applied for classification the characters. We have also compared our result with traditional HOG features across five popular classifiers for handwritten Tibetan characters. In every cases modified HOG features performed better than the traditional HOG features. Even if we combine modified HOG features along with traditional HOG features the performance show improvement for some classifiers.