Emotional image color transfer via deep learning

Different colors in images convey different emotions, e.g., artists often use different color combinations to convey different emotions in their creation. Traditional image color emotion transfer methods do not take the semantic information of images into account, which may lead to unnatural transfer results. To this end, this paper proposes a new emotional image color transfer framework by exploiting deep learning which can process images in an end-to-end fashion. Our network contains four main components, namely a low level feature network, an emotion classification network, a fusion network and a colorization network. The low level feature network extracts the semantic information which is designed to prevent the anti-natural phenomenon. The emotion classification network is used to constrain the color to make the enhancement results meet user's emotion. Then, the final deep learning framework combine the emotion classification network and the low level feature network by a fusion network. Finally, we use the colorization network to get the enhanced images. Various experiment results and comparisons validate our method.