Arbitrary Style Transfer with Style Enhancement and Structure Retention

Arbitrary style transfer is to transfer the style of any reference image to another image by a trained neural network, while preserving its content as much as possible. So far, a lot of work focused on reducing training costs without achieving sufficient style transfer, while some other work has neglected image frequency domain alignment. Balancing style presentation and content retention is no doubt challenging, we therefore propose a style transfer method that introduces frequency domain alignment and style secondary embedding, which is mainly embodied in two parts: style enhancement module (SEM) and content retention module (SRM). SEM aligns the stylistic image and stylized image statistics in the feature space. SRM reduces the loss of content by mapping the original and stylized images into the frequency domain and airspace for synchronous alignment. This new approach works well in terms of both style transfer and content retention. Experimental and questionnaire results show that this method can generate satisfactory stylized images without loss of content information.