A Multi-blocked Image Classifier for Deep Learning

Abstract

onvolutional Neural Networks (CNN) have been very successful in classification and object recognition. A lot of related work has been done to modify the performance of the networks, but they could either perform better with accuracy at high cost or decrease the time taken by a model on training datasets. We propose a deep neural network model ’Multi-Blocked Model’ which tends to decrease this gap and is efficient and accurate with less convolutional layers, easy to deploy online or embedded systems. We choose three datasets that are publicly available and popular for their own uniqueness among the datasets in deep neural networks. These three diverse datasets are: Modified National Institute of Standards and Technology (MNIST), Street View House Number (SVHN) and the Canadian Institute for Advanced Research with 10 Cases (CIFAR-10). Our stateof-the-art Multi-Blocked model is presented well on all three data sets. Dropout is added to overcome the overfitting problem. The multi-blocked model is designed in a way that it uses a minimum number of parameters so that it is able to run on a Graphical Processing Unit (GPU), which requires less power. The experimental results show that our proposed Multi-Blocked model tends to achieve the accuracy of these datasets by 99.40%, 90.8%, 88.07% consuming under 2 GB of graphical memory.