MulTIR: Deep Multi-Target Image Retargeting

Di Sun, Yitong Guo, Chaojie Yao, Yijing Mei, Dufeng Chen, Gang Pan

Abstract: Image retargeting aims to resize images to fit various devices while maintaining good viewing experiences. Normally, multi-operator image retargeting shows better performance than single operator strategy, however, there is still no single method that performs well on all cases. This inspires us to provide a general image retargeting framework that can adaptively learns from multiple methods. We present a multi-target image retargeting model named MulTIR, which learns the deformation process from multiple diverse outputs and automatically pick the optimal target in feature space. We also introduce a Mean-GAN-Min-Task loss to adapt the additional targets in each training example. Experimental results indicate the superiority of MulTIR against representative methods.