Automatic Image Annotation Based on Relevance Feedback

摘要

An image automatic annotation algorithm based on relevance feedback is proposed. Firstly, images are segmented into regions, and then the regions can generate blobs according to image features using clustering. Given a training set of images with annotations, we can compute the probability of a word given the image regions so as to automatically generate keywords for un-annotated image. Considering correlations among different semantics concepts, we employ condition probability to present two types of connections among different semantics concepts, and use the user's feedback information to adjust the probabilities of the keywords in annotation. The test results with Ground Truth Database illustrate the effect and efficiency of this algorithm.