Multi-scale Attention Enhancement for Arbitrary Style Transfer via Contrast Learning
Lei Zhou, Tao Zhang · 2023
Arbitrary style transfer is to transfer any artistic style to the content image while preserving the content structure as much as possible. Although there is currently a lot of work to improve the effect of image migration, there are still problems, such as the content structure being destroyed or the image quality not being guaranteed. In this paper, we combine multi-scale attention with contrastive learning methods to propose an efficient model called AttCST. The critical part is to fuse the current features extracted by the pre-trained deep convolutional network with the previous shallow features. Make full use of the previous shallow features to supplement the texture information lost by the in-depth features. A contrastive learning model is then used to distinguish the same style from others, thereby enhancing the network’s learning ability. At the same time, we add a local loss to constrain the transfer module to improve the quality of generated images. We conduct extensive qualitative and quantitative analyses, the results of which demonstrate the effectiveness of our approach.