Abstract
Can the abundant emotions reflected in pictures be classified automatically by computer? Only the visual features
extracted from images are considered in the previous researches, which have the constrained capability to reveal various
emotions. In addition, the training database utilized by previous methods is the subset of International Affective Picture
System (IAPS) that has a relatively small scale, which exerts negative effects on the discrimination of emotion
classifiers.
To solve the above problems, this paper proposes a novel and practical emotional picture classification approach, using
semi-supervised learning scheme with both visual feature and keyword tag information. Besides the IAPS with both
emotion labels and keyword tags as part of the training dataset, nearly 2000 pictures with only keyword tags that are
downloaded from the website Flickr form an auxiliary training dataset. The visual feature of the latent emotional
semantic factors is extracted by probabilistic Latent Semantic Analysis (pLSA) model, while the text feature is described
by binary vectors on the tag vocabulary. A first Linear Programming Boost (LPBoost) classifier which is trained on the
samples from IAPS combines the above two features, and aims to label the other training samples from the internet.
Then the second SVM classifier which is trained on all training images using only visual feature, focuses on the test
images. In the experiment, the categorization performance of our approach is better than the latest methods.
