A Printed Character Recognition System for Meetei-Mayek Script Using Transfer Learning

This paper presents a printed character recognition system for Meetei-Mayek script of Manipuri language. The input to the system is downloaded and cropped images containing printed characters of Meetei-Mayek. The images go through pre-processing which involves thresholding and Morphological Transformations. Segmentation of characters has been carried out using image processing techniques like dilation, thinning and text properties like height, width, and aspect ratio. After segmentation, the images are converted into three channeled black and white images. The work comprises development of standard database for printed characters for Meetei-Mayek. The classification task has been carried out using transfer learning using pre-trained VGG-16, VGG-19, and ResNet152-V2 convolutional neural networks (CNNs). The recognition results have been compelling, and VGG-16 has achieved better classification accuracy in comparison with other pre-trained CNN—VGG-19, ResNet152-V2. The classification accuracy obtained by VGG-16 is 99.27%.