Style Transfer Based on Normalization Loss

Abstract

Gatys et al. firstly proposed artistic style transfer. Based on their research, Luan et al. proposed a method stylizing a photo using the style of a reference photo. Many algorithms have shown great progress in style transfer. The combination of content loss and style loss is commonly used as objective function in style transfer. This paper proposes a new loss term called normalization loss which can replace the classic combination. Normalization loss is easier to converge. Therefore, the generation process requires less iterations. Experiments prove that our method is effective and the result images have better visual effect.