An efficient refinement algorithm for multi-label image annotation with correlation model

Abstract

With the explosively rising popularity of photography devices, collections of personal digital images are growing rapidly both in number and size. There is an increasing desire to effectively index and search these images to meet user requirements. The content-based image retrieval (CBIR) system facilitates effective image indexing and retrieval according to image features. However, the semantic gap between the low-level visual features and high-level semantic concepts hinders further development. Image annotation is a solution intended to resolve the CBIR system inadequacies. However, there are two problems with the annotation. (1) It is very difficult to represent an image using only a few keywords; (2) the manual annotation process is very subjective, ambiguous, and incomplete. This paper focuses on refining image annotation to cluster the most representative keywords, as the annotations to image with a small semantic gap. We propose the Hierarchical_Twin Rings algorithm to refine the quality of annotations in order to close the well-known semantic gap problem. Moreover, we present another Centroid-based Convergence method of automatically assigning relevant multi-keywords to a user specified image which could greatly improve the retrieval accuracy and fast response requirement. The key contributions of our work areas follows: (1) Weintroduce the problem of the mining of representative image keywords as the annotation for image indexing and retrieval from a large set of image collection. (2) We use Bayesian framework to integrate the image and image annotation into a unifiedframework. (3) Our formulation allows one to refine the relevant annotations of an image and remove redundant annotations. We evaluated the performance of our algorithm by means of images collected from Flickr, the photo sharing website. Our experimental results show that the Hierarchical_Twin Rings algorithm is a realistic and effective method for multi-label image annotation.