Image Annotation Research Based On The Probabilistic Topic Model

In the image annotation work, the image annotation method based on the probabilistic topic model is an important branch of image annotation study, which learns the semantic of the image to annotate images and is played more and more attention by researchers in recent years. This thesis studies image annotation based on probabilistic topic model in three aspects: Strengthen the relevance between image topic and text topic; use image proportion in the real images to construct image annotation model; class label information can provide the valuable information for image annotation, put label information fusing image annotation model. The main specific work is as follows:1. For the problem that the relevance between image topic and text topic is weak, this thesis attempts to strengthen the relevance between image topic and text topic by introducing intermediate variables, propose an image annotation probabilistic topic model(mm-LDA-C model) based on mm-LDA model. And it derives a parameters estimation algorithm based on the variational EM algorithm, as well as gives the method annotating the new images.2. For the existing image annotation works do not consider to put the information of image proportion to fuse the probabilistic topic model of image annotation, this thesis proposes a probabilistic topic model of image annotation based on Corr-LDA model, then derives a parameters estimation algorithm based on the variational EM algorithm, as well as gives the method of annotating new images according to the model.3. For class label information can provide the valuable information for image annotation, this thesis puts the class label information to fuse the Corr-LDA model, and proposes an image annotation probabilistic topic model fusing class information which uses class information to promote image annotation. And it derives a parameters estimation algorithm based on the variational EM algorithm, as well as gives the method annotating the new images.This thesis proposes three image annotation models based on the probabilistic topic model which are mm-LDA-C model, Corr-LDA-ITD model and Corr*-LDA-L model, as well as gives the method annotating the new images based on proposed model. The experimental results on Lable Me and UIUC-Sport datasets show that the image annotation performance of the proposed models is improved.