Multimodal Style Transfer with Rotation-Attention Network
Bin Xie, Wenhao Kong, Xiang Li · 2024
Digital image style transfer has received more attention, which aims to combine the style representation of an image with the semantics of its content to generate visually beautiful stylized images. Most current methods can only generate one result from two images, while others use complex networks to generate multiple outputs. In this paper, we propose a method that can generate arbitrary outputs, called Feature Rotation Attention Network (FRA).Our network contains a channel attention module (CA) and a style-cooperation attention module (SCoA). FRA can generate multiple outputs while solving the color distortion caused by deep feature rotation, and improves network efficiency by optimizing the selection of convolutional layers. Experimental comparisons reveal that our network is able to generate more aesthetically pleasing results.