Utilization of DCT Coefficients for the Classification of CIFAR-10 & CIFAR-100 Datasets

Abstract

In this work we propose efficient deep neural networks for classification that are well suited to the edge computing and cloud computing environment. These environments inherently have to deal with bandwidth limitations and bounded computational resources. Our proposed methods tend to reduce the bandwidth requirements and reduce the computational costs for running these deep learning algorithms. As an extensively used image compression algorithm, DCT (Discrete Cosine Transform) is used to reduce image information redundancy because a limited number of DCT coefficients can preserve the most significant image information. In this thesis, a novel frame work is presented by joining DCT coefficients and deep neural networks. We have targeted the deep neural network that can predict the most important DCT coefficients for an image and we have utilized those important DCT coefficients for classification purpose. As part of its training process, the proposed DCT model eliminates the input information which mostly represents the high frequencies. After achieving the important DCT coefficients from images we have applied the classification models on the compact representation of Grey scale and RGB datasets (MNIST, CIFAR-10 and CIFAR-100). We have used two approaches for classification first is classification by important DCT coefficients and second is classification by low resolution images. We have used VGG-16 and purposed CNN architecture for classification purpose. Additionally we have also proposed the prediction model for predicting the important DCT coefficients by using MultiLayered Perceptron (MLP) model. The experiments has shown the promising results and we have found out that we can achieve almost the same classification accuracy with compact representation of Grey Scale and RGB datasets as we were achieving with raw pixels.