Unpaired Image-To-Image Translation Using Generative Adversarial Networks With Coordinate Attention Loss

Abstract

Image stylization is an important research direction in image processing, graphics, and computer vision. At present, methods based on deep learning, especially generative adversarial network, have made great progress in image stylization migration. However, there are several limitations to the current mainstream methods, the biggest of which is the inability to perform geometry changes, remove large objects, or ignore irrelevant textures in unpaired scenarios. This paper proposes a style transfer algorithm CAGAN based on Adversarial Consistency Loss Generative Adversarial Network and Coordinate Attention. The stylized transfer of high perceptual quality in mismatched scenes is achieved by combating consistency loss and attention mechanism, and the Laplacian noise module is added to generate multi-modal output. Through a lot of experiments, it is verified that the algorithm can achieve high quality stylization effect.