Tag Refinement Algorithms For Internet Social Images

With the development of Internet Web2.0technique and the emergence of thesocial image-sharing websites (i.e., Flickr), these social image-sharing websites allowpeople to upload, share and utilize user-defined tags to manage images. Tag-basedimage annotation and retrieval become the focus of research. However, the wrong,incomplete, inaccurate and fuzzy tags exist in social images and reduce the tag qualitybecause the users utilize the control-free tags. This limits the development of tag-based image annotation and retrieval techniques. Hence, the research of improvingtag quality becomes a hot topic in the field of image retrieval.Towards the problem of tag refinement, it is divided into four key sub-problemsbased on the types of tag noise and multiple refinement algorithms are proposed tosolve the corresponding sub-problems:1. For the problem of noisy tag cleansing, this paper proposes a multi-layer clus-tering algorithm for tag refinement. In the refinement algorithm, a clustering frame-work is developed to utilize the bi-modality information on tag correlation and imagesimilarity. Then the whole dataset could be divided into different sub-dataset. A prob-abilistic model between the tag and image group is proposed and utilized to clean tagswith tag statistics and relevance. Compared to the-nearest-neighbor algorithms, theproposed algorithm is suitable for the applications with large-scale images due tothe multi-layer clustering integrated with bi-modality information. Combined withthe tag relevance, the proposed algorithm enhances the relevance of low-frequencycontent-relevant tags and improves the refinement gain of wrong tag detection.2. For the problem of missing tags, this paper propose a regularized optimizationalgorithm for tag refinement. The non-negative matrix factorization (NMF) algorithmis utilized to discover the relationship of missing tags. The holistic visual diverseis proposed and then utilized to enhance the relation between images and tags. Theregularization term is utilized to penalize the complexity of the optimization model fortag completion. In contrast to the traditional algorithms, the proposed algorithm couldintegrate the tag relevance and the image visual diversity. Meanwhile, the algorithmenhances the whole characterizing of visual image and improves the accuracy of tag completion. Moreover, the proposed algorithm is clear and fast.3. For the problem of abstract tag, this paper proposes a tag refinement algorithmbased on the ontology and neighbor voting. A semantic ontology is proposed tocharacterize the semantic relationships between tags. Then this ontology is used todetect the abstract tag candidate. The tag context and image context are developedto discover the specific tag. The nearest-neighbor voting algorithm is used to detectthe specific tag in which the specific tag can improve the description ability of tagsfor image content. The abstract tag problem and its solution are presented for thefirst time in this paper. The experiments illustrate that the algorithm could detect theabstract tags through the ontology and discover the specific tags by neighbor voting,where the recall of image retrieval is also improved.4. For the problem of tag-to-region assignment, this paper proposes a diversedensity based multiple instance learning algorithm. In the algorithm, two strategiesare proposed to improve the computation of diverse density and speed up the compu-tation process. The automatic detection of boundary threshold is proposed to improvethe performance of multiple instance learning algorithm. The algorithm can assignthe tags to regions and it is possible for tag-based image object retrieval. In con-trast to traditional algorithms, this proposed algorithm applies the multiple instancelearning algorithm to the tag-to-region problem and describes the relationships moreaccurately between image regions and tags. Our revised multiple instance learningalgorithm computes the global optimization solution faster and has a simpler com-putation procedure based on two extended strategies. The automatic detection ofboundary threshold could improve the accuracy of tag-to-region, which is better thanthe fixed threshold method.