Deep Image Retargeting Network with Multi-loss Functions

Image retargeting aims at displaying an image on a serious of device screen with different sizes, which has been widely applied in computer graphics and other fields. At present, many deep learning-based image retargeting methods implement an encoder-decoder retargeting network to resize attention map. Then, various types of loss functions are presented to preserve salient contents and reduce structure distortions. In this survey, we first review three types of loss functions utilized in deep image retargeting that consists of pixel-based, probability-based, and perception-based. Furthermore, we explore the baseline encoder-decoder retargeting network with three types of loss functions, and conduct the experiments on two public datasets, Retargetme and Pascal Voc 2007 datasets to verify their impact on the deep image retargeting.