Image Re-Emotionalizing

Abstract

In this work, we develop a novel system for synthesizing user specified emotional affection onto arbitrary input images. To tackle the subjectivity and complexity issue of the image affection generation process, we propose a learning framework which discovers emotion-related knowledge, such as image local appearance distributions, from a set of emotion annotated images. First, emotion-specific generative models are constructed from color features of the image super-pixels within each emotion-specific scene subgroup. Then, a piece-wise linear transformation is defined for aligning the feature distribution of the target image to the statistical model constructed from the given emotion-specific scene subgroup. Finally, a framework is developed by further incorporation of a regularization term enforcing the spatial smoothness and edge preservation for the derived transformation, and the optimal solution of the objective function is sought via standard non-linear optimization. Intensive user studies demonstrate that the proposed image emotion synthesis framework can yield effective and natural effects.