A Generative Adversarial Network with Attention Module for Unpaired Image-to-Image Translation

Abstract

The Image-to-Image Translation (I2IT) is a challenging image processing task that can be applied to many aspects, such as super-resolution and style transfer. Although several image translation algorithms based on Generative Adversarial Network (GAN) have been proposed, achieving better translation effects still remains a problem worthy of attention. This work proposes a model that fuses an attention module and ResNet-based generator to enhance the performance of I2IT. Using an attention module after the first downsampling, our model can focus more on important low-level semantic features. After the downsampling, the residual blocks provide contextual supplementary information of the photos. The qualitative and quantitative experimental results on unpaired datasets show that our model is better than the SOTA methods, which further confirms the robustness and effectiveness of the proposed model.