Image memorability prediction using deep features

Abstract

Image memorability is an intrinsic characteristic of an image that is predictable. Prediction of image memorability measure is a challenging task in visualization and photography. In this study, the effectiveness of deep features on measuring of image memorability is investigated. The term of ‘deep features’ refers to the features which are extracted from a deep convolutional neural network (CNN). Within the framework of the CNN, we use output of the last three fully connected layers, fc6, fc7 and fc8, as the features. For this purpose, we use three common types of convolutional networks, Alex-Net, Hybrid-CNN and VGG-Net. We show that the derived deep features outperform some other commonly used features such as the SIFT, GIST, SSIM and HOG 2×2. We also show that the prediction accuracy can be more improved by proper fusing of the deep features.