Classification of Computer Generated and Natural Images based on Efficient Deep Convolutional Recurrent Attention Model

Published in CVPR Workshops, 2019

Most state-of-the-art techniques of distinguishing natural images and computer generated images based on handcrafted feature and Convolutional Neural Network require processing of the entire input image pixels uniformly. As a result, such techniques usually require extensive computation time and memory, that scale linearly with the size of the input image in terms of number of pixels. In this paper, we deploy an efficient Deep Convolutional Recurrent Attention model with relatively less number of parameters, to distinguish between natural and computer generated images. The proposed model uses a glimpse network to locally process a sequence of selected image regions; hence, the number of parameters and computation time can be controlled effectively. We also adopt a local-to-global strategy by training image patches and classifying full-sized images using the simple majority voting rule. The proposed approach achieves superior classification accuracy compared to recently proposed approaches based on deep learning.

Download paper here