BC-FR: Bijective Contrast and Fusion Reconstruction Networks for Unpaired Image-to-Image Translation

Image-to-image translation aims to learn some mapping relationships between two different domains to implement cross-domain conversion. And symmetric dual learning is one of the classic architectures. However, if the generators are expected to achieve the cycle-consistency, this high requirement frequently causes mode collapse, which constrains the performance. In this paper, to meet this challenge, we propose a novel dual learning framework, bijective contrast and fusion reconstruction (BC-FR) network. On the one hand, drawing inspiration from the widely adopted contrastive representation learning, we propose the bijective contrast (BC) network. Specifically, a reverse contrastive learning from output to input patches is designed, which enables two embedding spaces to learn the mapping relationship in a bijective way. This strategy provides a richer pixel-level domain shift for object. On the other hand, to further improve the visual performance, we also propose the fusion reconstruction (FR) network, which provides an unsupervised fusion approach to achieve the cycle-consistency. Specifically, it separates and reassembles different text elements of input and output images to achieve the reconstruction work. Experiments on various pixel-level benchmark datasets show that BC-FR can obtain comprehensive quantitative metrics and yield high-fidelity visual outputs. Furthermore, the sub-scheme FR can be extended to semantic-level datasets.