Resizing Tiny Imagenet: An Iterative Approach Towards Image Classification

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%.