Resizing Tiny Imagenet: An Iterative Approach Towards Image Classification

Abstract

Deep neural networks have attained almost  human-level performance over several Image classification  and object detection problems, deploying robust and powerful  state-of-the-art networks. Stanford’s Tiny ImageNet dataset has been around for a while and neural networks have  struggled to classify them. It is generally considered one of the  harder datasets in the domain of image classification. The  validation accuracy of the existing systems maxes out at 61- 62% with a select few shooting beyond 68-69%. These  approaches are often plagued by problems of overfitting and  vanishing gradients. In this paper, we present a new method to  get above average validation accuracy while circumventing these problems. We use the resizing image technique which  trains multiple model times over different image sizes. This  approach enables the model to derive more relevant and  precise features as we move towards the original image size. It  also makes the training process adaptable to available  hardware resources. After reaching the image size, we hyper tune other parameters such as Learning Rate, Optimizers, etc.  to increase the overall validation accuracy. The final  validation accuracy of the model using resizing and hyper  tuning is 62.57%.