Text line segmentation in indian ancient handwritten documents using faster R-CNN

Textline segmentation in ancient handwritten documents is still considered as a challenging task in document analysis and recognition field even though various rule-based methods exist. These methods succeed under constraint such as a roughly uniform background. They do not contribute well in case of variable inter-line spacing and overlapping characters. This article proposes faster region-convolution neural network (R-CNN) based robust method to segment the textlines in the ancient handwritten document in Devanagari script for the first time in literature. The feature matrix has been generated by residual network and proposals have been predicted through the region proposal network (RPN). A pooling layer has been used to extract regions of interest, known as region of interest pooling layer, to locate the textlines. The performance of the proposed textline segmentation system has been evaluated on self generated dataset of ancient handwritten documents in Devanagari script and it has achieved the f-measure of 99.98%. Experimental results demonstrate that the proposed system outperforms the existing state-of-the-art methods of textline segmentation.