Multi-Font and Multi-Size Printed Sindhi Character Recognition using Convolutional Neural Networks

Abstract

In this paper, a problem of multi-font, multi-color and multi-size printed character recognition of Sindhi language are addressed. Although previous studies for offline handwritten isolated Sindhi character recognition with unique font and size have achieved satisfactory results, the problem of multi-fonts, multi-size and multi-color character recognition is still a major challenge. This is due to the various varieties in the shape, style, and layout of the character. A synthetic dataset with background color image consisting of Sindhi characters with multi-fonts, multi-size, and multi-colors is created. Three types of experiments with Convolutional Neural Networks (CNN) are performed separately. The first CNN network uses max-pooling layer after every two convolutional layers, the second network applies multi max-pooling layers after the last convolutional layer and the third network is created without applying any max-pooling layer. The experimental results demonstrate that convolutional neural network with max-pooling layers improves the performance significantly. The recognition results of 99.96%, 97.94%, and 98.72% are achieved with first, second and third networks respectively, which shows that CNN with pooling layers is more effective.