BiRGAN: Bi-directional Deep Image Retargeting

Di Sun, Yunxiang Wang, Tingting Yang, Yijing Mei, Gang Pan

Abstract: Current single retargeting operators perform poorly on diverse images and varying target sizes, rendering them unsuitable for both image reduction and expansion simultaneously. In this paper, we present a deep bi-directional image retargeting network, BiRGAN. The network performs two opposite processes: one training a generator to learn the process of downsizing images, and another learning the upsizing process. The output from the first process is fed into the second, creating a comprehensive closed loop. This network is designed to comprehend the deformation process of retargeted images by employing multiple methodologies and executing retargeting operations within the feature space. Experimental results demonstrate that BiRGAN outperforms previous methods in terms of overall image retargeting results.