Advances in Emotional Picture Classification

One picture is worth more than thousand words. A picture contains not only visual features as well as object and scene information, but also emotional meanings that are usually represented in adjective form, for example, happy, sad, fear, etc. (Wang, 2008). In content-based visual information retrieval (CBVIR), emotional-based one is considered as in the highest level, as it involve the subjective perception and understanding of visual information. On the other side, emotion recognition is considered as one of the important part to develop in human-human and human-computer interaction. While the recognition of emotions revealed by facial expression has been mature enough to develop available emotion recognition software, automatic classification of emotions evoked by visual scenes in pictures is still a quite preliminary and challenging area of research (Li, 2010). Few works have been conducted in this field. One typical approach is based on holistic image feature (Yanulevskaya, 2008). It decomposed complex scenes according to an annotated object vocabulary, and assigned a similarity score to all words for each region in an image. Different combinations of visual words in a similarity histogram provided a sufficient characterization of a complex scene picture. Wiccest features and Gabor features were then used for regional feature extraction, and SVM framework was adopted for supervised learning of emotion classes. Two particular new works are presented here to show the recent advances in this field. One is focused on the middle level semantic extraction, in an intention to fill the gap between low-level visual features and high-level emotional concepts to connect human and machine in a uniform framework. Another is to combine low-level visual features and high-level text features to improve the efficiency of classification procedure.