A novel facial expression database construction method based on web images

Facial expression recognition plays an important role in web images understanding. Since current databases for expression learning are usually achieved in experimental environment which are different from actual emotions, and the samples in each database is small, it is hard to train good expression classifiers with them, and which leads to a low expression classification ability, especially when recognizing expressions in web images. In this paper, a novel facial expression database construction method is suggested: First, a large-scale social label images are obtained by the Google web search with the keywords of happiness, sadness, surprise, anger, disgust and fear respectively; Then, unrelated images are filtered as junk images interactively by the hyperbolic visualization technique and the expression database is constructed. All the images in the database are from real social network, so the database is more proper to train classifiers for recognizing expressions in web images.