A Compositional Transformer Based Autoencoder for Image Style Transfer

Abstract

Image style transfer has become a key technique in modern photo-editing applications. Although significant progress has been made to blend content from one image with style from another image, the synthesized image may have a hallucinatory effect when the texture from the style image is rich when processing high-resolution image style transfer tasks. In this paper, we propose a novel attention mechanism, named compositional attention, to design a compositional transformer-based autoencoder (CTA) to solve this above-mentioned issue. With the support from this module, our model is capable of generating high-quality images when transferring from texture-riched style images to content images with semantics. Additionally, we embed region-based consistency terms in our loss function for ensuring internal structure semantic preservation in our synthesized image. Moreover, information theory-based CTA is discussed and Kullback–Leibler divergence loss is introduced to preserve more brightness information for photo-realistic style transfer. Extensive experimental results based on three benchmark datasets, namely Churches, Flickr Landscapes, and Flickr Faces HQ, confirmed excellent performance when compared to several state-of-the-art methods. Based on a user study assessment, the majority number of users, ranging from 61% to 66%, gave high scores on the transfer effects of our method compared to 9% users who supported the second best method. Further, for the questions of realism and style transfer quality, we achieved the best score, i.e., an average of 4.5 out of 5 compared to other style transfer methods.