CycleGANs Based Social Image Stylization by Emotion
Jie Nie, Lei Huang, Zhen Li, Meng Yuan, Zhiqiang Wei · 2018
Social posts are always accompanied with one or more images to express enhanced emotion. It requires the image could produce a correct mood as text needs. However, existing image stylization methods either are not emotion-oriented or set a high threshold for user's aesthetic background and manipulation of image editing tools. Thus, we propose a novel image stylization model to transfer social images refers to emotion embedded in posting text. Firstly, this model is trained by pairwise image and Emoji label from social networks, and the learned model could stylize the input image to create a match-able mood request by the text. Secondly, the model is built on Cycle-consistent Generative Adversarial Networks. It learns a probability distribution model to replace an art-theory driven model and leads to a data-driven solution. Comprehensive experiments demonstrate the effectiveness of proposed method.