Web Image Annotation using Non-negative Matrix Factorization-Based Tag Clustering
Sun-Young Cho, Jae-Seong Cha, Hyeran Byun · Jeongbo gwahaghoe nonmunji. so'peuteuweeo mich eung'yong · 2011
Recently, with the development of social multimedia tagging through Flickr and Facebook, many researches have studied by using social tag-annotated web image. However, previous methods produce the tag list with low consistency since low correlation among extracted tags. This paper proposes web image annotation through NMF-based tag clustering for the consistent tag extraction. Our method derives the ideas from assumptions and properties for weh image dataset, connects the annotation problem with NMF-based clustering. We construct the data matrix using tag-frequency vector of neighbor images and predict the tags relevant to the query by analyzing two matrices decomposed by NMF. We conduct an experiment on annotated web image dataset collected from Flickr. We show that the proposed method gives more high performance in tagging accuracy and consistency than the previous methods.