Performance Analysis of Various Trained CNN Models on Gujarati Script

Abstract

In the age of digitization, it’s worthy to preserve ancient scripts, researches, and documents. A digital format of the mentioned documents would help in preservation and replication and therefore, the chief goal of this research is to analyze the performance of various trained CNN (Convolutional Neural Network) models to recognize Gujarati scripts. As the ancient scripts are somewhat degraded in terms of their physical aspects and are sometimes unrecognizable, CNN plays a vital role in digitizing printed texts so that they could be edited in an electronic format, stored more compactly and used in various machine processes. The models used for experimentation includes LeNet (Learning CNN), DenseNet (Dense CNN), and VGGNet (Visual Geometric Group Network), which are used for Optical Character Recognition (OCR), and have obtained the results on our synthetically generated dataset. The dataset comprises more than 400 classes with an 80-20 division of train and test images respectively. The performance parameters include validation accuracy, validation loss and the number of epochs required to train the CNN model.