Detection of indian classical asanyukta mudras using convolution neural network

Abstract

In this fast- changing world, the work system is constantly evolving and undergoing continuous improvement. In recent times, education has become a crucial aspect for individuals in every field. This project utilizes machine learning to provide highly profitable classical dance education. Classifying mudras on a large dataset is a multi-class classification problem. Here, a large dataset of 10 different mudras has been created. The dataset is further enlarged by implementing data augmentation. Some of the parameters in the layers of the CNN have been modified. It has helped improve the model's detection and recognition capabilities. The existing model is improved by changing the final dropout layer to 40% and introducing the Selu activation in the dense layer. After comparing the results, it was found that changing the optimizer from SGD to Adam reduced overfitting. It is observed that the existing architecture exhibited higher accuracy when evaluated on the dataset. In the experiment, a large dataset is being considered. The accuracy of the existing model, when compared to the proposed model, has decreased from 96.87% to 89.77% due to a change in the dataset.