A Quality-improved Method based on Attention Mechanism and Contrastive Learning for Image Style Transfer

Abstract

With the significant advances in deep learning, arbitrary style transfer (AST) methods are easy to achieve. However, artifacts and distortions occur constantly. The image quality improvement for AST has become a hot topic to research. In this paper, we propose a quality-improved method based on AesUST. First, a visual enhancement module guided by an attention mechanism is creatively presented to implement shallow and deep features integration on a per-point basis. Then, a contrastive coherence preserving loss is built to have a patch-wise calculation between content images and results, making the outputs softer and smoother. After the ablation study and the comparison with previous methods, the results show that our model visually produces more pleasant images and has a prosperous performance on quality evaluations.