Contrastive Learning with Attention Mechanism and Multi-Scale Sample Network for Unpaired Image-to-Image Translation

Abstract

The aim of unpaired image translation is to learn how to transform images from a source to a target domain, while preserving as many domain-invariant features as possible. Previous methods have not been able to separate foreground and background well, resulting in texture being added to the background. Moreover, these methods often fail to distinguish different objects or different parts of the same object. In this paper, we propose an attention-based generator (AG) that can redistribute the weights of visual features, significantly enhancing the network's performance in separating foreground and background. We also embed a multi-scale multilayer perceptron (MSMLP) into the framework to capture features across a broader range of scales, which improves the discrimination of various parts of objects. Our method outperforms existing methods on various datasets in terms of Fréchet inception distance. We further analyze the impact of different modules in our approach through subsequent ablation experiments.