Image Annotation Research Based On Subject Embedded BoW Model And Tag Frequency Decomposition

With the development of the Internet and the maturity of social networks,image information is propagated on an unprecedented scale.Automatic image annotation is one of the important techniques to efficiently manage and retrieve massive data.In this thesis,we focus on the semantics-embedded annotation methods in the Bag of Word model(BoW),the tag enrichment of a training set and selection of the nearest neighboring image with semantic neighborhood.The main works can be summarized in three parts:To solve the problem of poor subject discrimination of visual words in BoW model,a subject embedded Bow(SEBoW)are proposed.Taking the dominant subject of an image as the coset information of visual words,the compressed sensing texture feature of the image blocks are organized in a hierarchical tree with a subject-sub-subject structure.Therefore the similarity of the subject is embedded in the BoW model.For no subject or chaotic subject in the training set,we use tag frequency decomposition vector to reconstruction the dataset.Experimental results demonstrate that compared with the BoW in the paper that proposed PLSA-Words,SEBoW model in the ap,ar,F1 values is increased by 9.8%,7.9%,9.2%.The tag enrichment module in the FastTag algorithm has a shortcoming.It uses the boolean feature to represent tag result in it can not express the semantic overlap between the tags.An improved FastTag image annotation based on the tag frequency decomposition vector is proposed to solve this shortcoming.We use the tag frequency decomposition vector instead of the boolean feature therefore the overlap ratio between the label semantics is quantified reasonably.A new joint loss convex function optimization method is proposed,used to train the tag enrichment classifier and the tag prediction classifier.The experimental results show that the improved FastTag algorithm is 1% higher than the average precision of the FastTag algorithm,and the number of tagged labels is increased by one.Aiming at the problem that the nearest neighbor images based on the visual modal selection are not necessarily semantic nearest neighbor images,the nearest neighbor selection of image with feature and tag is proposed.The tag frequency decomposition vector is used as the information of the image text modality,combined with the image feature to search neighbor images.Therefore the nearest neighbor images contain similar labels.And then we use the nearest neighbor images as the training set,and train the specific FastTag algorithm for each image to improve the annotation effect.The comparison experiment shows that compared with the FastTag algorithm,the proposed algorithm in the ap,ar,F1,N+ values is increased by 2.7%,1%,1.8%,1.