Semi-supervised Learning with Local and Global Consistency on Bi-relational Graph for Image Annotation

Abstract

We present a semi-supervised learning algorithm based on local and global consistency, working on a bi-relational graph of images and labels. By incorporating two types of dierent entities in a single graph, we exploit the label propagation to measure the relevance score between a specic label and unannotated images. The principle of the label propagation process is similar to many other semi-supervised learning methods, in which each node receives the information from its nearby points, and also retains its initial information. However, in our model, the neighbor concept is extended between dierent types of entities. As a results, the label correlation is captured, increasing the accuracy of image annotation. Moreover, in the propagation process, nodes from the same group are not treated equally, each has dierent relative reliability for the propagation process. We perform our method on two benchmark multi-label image data sets and gain encouraging experimental results compared to the existing work.