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

Abstract

In this paper, a problem of multi-fontand multi-size offline printed character recognition of Sindhi language is 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 andmulti-size character recognition is still a major challenge. This is due to the various varietiesin the shape, style,and layout of the character.A synthetic dataset with background color image consisting Sindhi characters with multi-fonts, multi-size,and multi-colors is created. Three types of experiments withConvolutional Neural Networks (CNN) are performed separately. The first CNN network uses max-poolinglayer after every two convolutional layers, the second network applies max-pooling layer afterthe last convolutional layer and the third network is created without applying any max-pooling layer. The experimental results demonstratethat the max-pooling layers used after every two convolutionallayers improve the performance significantly. The recognition results of 99.96%, 97.94%,and98.72% are achieved with first, secondandthird networks respectively, which shows that CNN outperforms than the traditional machine learning algorithms