Predicting Image Emotion Distribution by Emotional Region

Recent studies on image emotion prediction mainly focus on classifying images into a certain emotion category, but single label cannot reflect peoples' multiple emotion to image. To get more realistic results, we study image emotion distribution problem. In the most image emotion tasks, features are extracted from the whole image, but not each part makes contribution to emotion, so features from the whole image contain noises. In order to get discriminative features, we propose to leverage the heatmap generated by Fully Convolutional Networks (FCN) to select the Region of Interest (ROI) from an image which represents the image emotion most. Both high-level features and hand-crafted features from ROI are fused to train Support Vector Regressors (SVRs) to predict emotion distribution. Extensive experiments conducted on two widely used datasets demonstrate that emotional region is selected out through our method so that emotion distribution prediction performances are improved.