Image Classification on Low Resolution Images using Dilated Dense Pyramid Convolutions
Jenisha Thankaraj, Swarnalatha Purushothaman · 2020
Deep neural networks has the tendency to deploy a high image classification accuracy for high resolution images. Classifying the images captured by low resolution cameras are always a challenging task. Because of low resolution the extraction of good features from this type of image is still remaining as a challenging task for researchers. Moreover high resolution images have more layers to train that means they take more time to train. As the layers go deeper the problem of vanishing gradient increases. In this research work, low resolution images from scratch are trained using dilation kernels. By adding dilation the receptive field of the image gets increased. This boost the classification accuracy in Tiny ImageNet data set. This improved model shows top 1 accuracy of about 63.45% without using any ensemble approaches.