Performance Analysis of Deep Neural Network and Stacked Autoencoder for Image Classification

There are many challenging impacts on real-time image classification problems like extraction of features from a noisy and uncertainty existence. Task-based feature extraction is not possible for all the cases; to overcome this automatic feature extraction is included in the layers of the deep neural network and stacked autoencoder (SAE), which improves classification accuracy and speed. In this paper image datasets such as MNIST are taken, and it is trained and tested using two different networks. The time consumed and accuracy during the training period are calculated for the MNIST images applying the DNN algorithm. On the other hand, a stacked autoencoder (SAE) is constructed which is trained one layer at a time. Here the SAE consist of three layers which are stacked together, and its parameters are varied in such a way that the constructed SAE outperforms the DNN model. The SAE model improves the validation set accuracy by a noticeable margin. This paper demonstrates the effectiveness of using the SAE model over DNN with the performance analysis of binary handwritten image with time and accuracy trade-off.