Get our free extension to see links to code for papers anywhere online!

Chrome logo  Add to Chrome

Firefox logo Add to Firefox

Constraining Multi-scale Pairwise Features between Encoder and Decoder Using Contrastive Learning for Unpaired Image-to-Image Translation

Add code

Nov 20, 2022
Xiuding Cai, Yaoyao Zhu, Dong Miao, Linjie Fu, Yu Yao

Share this with someone who'll enjoy it:

Contrastive learning (CL) has shown great potential in image-to-image translation (I2I). Current CL-based I2I methods usually re-exploit the encoder of the generator to maximize the mutual information between the input and generated images, which does not exert an active effect on the decoder part. In addition, though negative samples play a crucial role in CL, most existing methods adopt a random sampling strategy, which may be less effective. In this paper, we rethink the CL paradigm in the unpaired I2I tasks from two perspectives and propose a new one-sided image translation framework called EnCo. First, we present an explicit constraint on the multi-scale pairwise features between the encoder and decoder of the generator to guarantee the semantic consistency of the input and generated images. Second, we propose a discriminative attention-guided negative sampling strategy to replace the random negative sampling, which significantly improves the performance of the generative model with an almost negligible computational overhead. Compared with existing methods, EnCo acts more effective and efficient. Extensive experiments on several popular I2I datasets demonstrate the effectiveness and advantages of our proposed approach, and we achieve several state-of-the-art compared to previous methods.

* 16 pages, 10 figures 

   Access Paper Source

Share this with someone who'll enjoy it: