Recognition of Handwritten Gujarati Conjuncts Using the Convolutional Neural Network Architectures: AlexNet, GoogLeNet, Inception V3, and ResNet50

Abstract

A methodology for recognizing offline handwritten Gujarati conjuncts has been introduced in the proposed article. This article includes 767 types of frequently utilized conjuncts. The convolutional neural network (CNN) architectures AlexNet, GoogLeNet, Inception V3, as well as ResNet50, are used to recognize handwritten Gujarati conjuncts. The performance of these CNN architectures is systematically evaluated. Gujarati conjuncts are required to segment from the appropriate place before recognition. Conjuncts are segmented into preceding components and succeeding components before recognition. These preceding components and succeeding components can be then individually recognized by the system. 19694 sample images of preceding components of the conjuncts and 28050 samples images of the succeeding components of the conjuncts are utilized for the proposed research work. Maximum accuracy of 93.41% for the preceding components and maximum accuracy of 89.01% for the succeeding components have been achieved by using GoogLeNet.