Unsupervised image-to-image translation with multiscale attention generative adversarial network

Unsupervised image-to-image translation refers to translating images from the source domain to the target domain, assuring that the translated images have the style of the target domain while retaining the content of the source domain. Although existing image-to-image translation methods can map an image from the source domain to the target domain, the translation results are prone to visual artifacts, and the texture and shape of the input image cannot match the target domain well. The reason for this phenomenon is that the generator ignores the most differential information between the source and target domains, preventing the extraction of the rich image feature information. In this paper, we propose a multiscale attention-generative adversarial network (MSA-GAN) for unsupervised image-to-image translation. In MSA-GAN, we design a multiscale attention network (MSANet) as the backbone of the generator, which consists of the Res2Net block and convolutional block attention module (CBAM). MSANet can extract global and local features and effectively alleviate the detail missing and blurry problems in image translation. It also focuses on the important image features and improves the ability of the network to extract features from the most distinguishing regions between the source and target domains, which allows it to better translate the texture details and object shape. In addition, to generate high-quality images, we introduce the perceptual loss to constrain high-level feature information. Extensive experimental results show that the proposed MSA-GAN achieves competitive performance in image-to-image translation. Our model outperforms several advanced models on several public benchmark datasets.