CRPGAN: Learning image-to-image translation of two unpaired images by cross-attention mechanism and parallelization strategy

• DOI: 10.1371/journal.pone.0280073
• Publication Date: 2023-01-06T18:33:55Z
• Authors: Long Feng; Guohua Geng; Qihang Li; Yi Jiang; Zhan Li; Kang Li
• Keywords: Image translation; Convolution (computer science); Feature (linguistics); Distortion (music)

Unsupervised image-to-image translation (UI2I) tasks aim to find a mapping between the source and target domains from unpaired training data. Previous methods can not effectively capture differences domain on different scales often leads poor quality of generated images, noise, distortion, other conditions that do match human vision perception, has high time complexity. To address this problem, we propose multi-scale structure progressive growth generator method solve UI2I task. Our refines images global structures local details by adding new convolution blocks continuously shares network parameters in also same scale network. Finally, Cross-CBAM mechanism (CRCBAM), which uses multi-layer spatial attention channel cross generate more refined style images. Experiments our collected Opera Face, open datasets Summer↔Winter, Horse↔Zebra, Photo↔Van Gogh, show proposed algorithm is superior state-of-art algorithms.