Meitei Mayek handwritten dataset: compilation, segmentation, and character recognition

A peculiar Indian Script Meitei Mayek has experienced a resurgence in the last few years and gets very little attention in handwriting research due to recently insurgence and limited sources. The objective of this paper is two folds; firstly, develop two different datasets: Mayek27 having 4900 isolated Meitei Mayek alphabets and MM (Meitei Mayek) dataset of 189 full-length handwritten text page. Secondly, develop a recognition system on the Mayek27 dataset using convolutional neural network and segmentation algorithms (text-lines, words, and characters) on the full-length Meitei Mayek handwritten text. A recognition rate of 99.02%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$99.02\%$$\end{document} is achieved using three layers of convolutional layers with a filter size of 3×3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3 \times 3$$\end{document} with 16, 32, and 96 kernels. In MM text dataset, the text-line and word segmentation are performed concurrently on 809 lines by tracking space between lines in a novel approach based on horizontal projection histogram and monitoring vertical projection histogram along the run-length of segmentation. Various constraints like skew, curve, close, and touching text-lines are incorporated, and the segmentation algorithm results are 91.84% and 88.96% for text-line and word, respectively. Furthermore, characters are segmented by headline removal, and connected component analysis achieves an accuracy of 91.12%.