Affective Image Classification Using Multi-Scale Emotion Factorization Features

摘要

Image emotion analysis is a new and challenging research direction that gains more and more attention in the research community. Most previous works in this field only use common or generic features, and have hard restrictions on training images, such as scale, resolution, etc. Inspired by scale-space theories and psychology theories of color, we propose a procedure to extract interpretive features expressing human color emotions' three mayor variables' (activity, weight, heat) edge, ridge and blob structures described as pose vector in images, then use data fusion method to prune and unify the extracted raw data by the line-token representation. At last we construct a codebook representation and train SVM classifiers to implement affective image classification. We extensively demonstrate our proposed approach on two benchmark database, finally an improved classification results are obtained, compared to state of the art work.