Handwritten Text Segmentation Using Pixel Based Approach

This paper presents a simple approach for the character segmentation of the handwritten words by bounding box approach and pixel based approach. The handwritten character segmentation is a tedious process because of their unconstrained writing styles. The handwritten words are classified based on their writing method. The non-touching words are segmented by the bounding box approach and the touching words are segmented by the pixel-wise approach. The novel part of this paper is to avoid over segmentation by thresholding automatically. The characters which are separated are then subjected to recognition using KNN, the non-parametric classifier. The features extracted for the classifying the model are Histogram of Oriented Gradients (HOG), Log-Gabor filters, concatenation of both the features and some geometric features. The segmentation accuracy is also responsible for the performance of the recognition. A benchmark database of IAM handwritten words is used for the handwritten segmentation and recognition. A random subset of words are taken into account and carried out for the proposed work. This paper achieves the segmentation rate of 94.45% and the recognition rate for 50-50% training and testing ratio is 85.89%.