Multilevel contrast strategy for unpaired image-to-image translation

Contrastive learning for unpaired image-to-image translation utilizes adversarial loss to ensure the realism of generated images in the target domain and incorporates pixel-wise contrastive loss to maximize the correlation between them. However, existing methods only contrast pixel-wise features, ignoring higher-level features, and the pixel-wise contrast is imperfect, which leads to poorer perceptual and visual results. In order to alleviate these problems, we propose an effective multilevel contrast strategy for unpaired image-to-image translation (MLCUT), which contrasts features at three levels to generate more harmonious and realistic images. Specifically, we strengthen the pixel-wise level contrast and introduce the contrasts of plane and voxel-wise levels. On the one hand, MLCUT enhances training effectiveness by picking over hard negative keys for each query at the pixel-wise level. On the other hand, we strengthen the learning preferences of generators on features of objects rather than backgrounds by contrasting the plane-wise discriminative matrices in adversarial loss. Furthermore, by contrasting voxel-wise global semantic vectors, MLCUT effectively improves the realism of generated images and avoids mode collapse. Qualitative and quantitative experiments demonstrate that our method effectively improves performance in perception and vision.