Abstract:

Topics cannot be labeled in the unsupervised topic model, while the labeling work in supervised topic models is tedious and subjective. To solve these problems, a semisupervised topic model was proposed. First, the locationirrelevant local features were detected and described by the scaleinvariant feature transform (SIFT), based on which images were represented by a bag of visual words. Then partial labels were introduced to the topicword level distribution in the latent Dirichlet allocation (LDA) model to guide the classification of the unlabeled data, which resulted in a semisupervised LDA (SSLDA) model. The validation on head pose estimation showed the classification rate of the proposed method was 9.0％～24.7％ higher than that of LDA. And the pose classification rate on partially occluded and misaligned face images was 8.8％ and 21.5％～39.8％ higher than multi-pose PCA method. With a small amount of labeled images, the proposed SSLDA model approaches the fully supervised LDA method. And it is applicable to other image classification problems.
 

Key words: pose estimation, bag of visual words model, latent Dirichlet allocation, semi-supervised learning