A Study About The Process Of Automatic Image Annotation

With the rapid development of Internet technology and intelligent communication,digital images have become the primary manifestations of network resources. Explodingof the image resources, corresponding management system is desired. In order to meet thediversity demand of users, it’s become an urgent issue to find the image convenientlyfrom the bottomless well of resources. Currently, image retrieval technology has manyproblems, such as the subjectivity and inefficiency by manual annotation, and the "semantic gap" phenomenon based on the content retrieval. Automatic image annotationis precisely an effective way to solve the above problems. In this paper we summarize theadvancement and the outstanding problems of automatic image annotation, and then putforward a series of innovative methods pointedly. The main research work is shownbelow:The stage of model generation:(1) The mean value of different features contributing to cluster is different. Thispaper puts forward a new algorithm of feature weighted- analysis of weights based onintra-class and inter-class distance. This algorithm makes up for the defect of Relief Fwhich indicates poor anti-noise and traps into local optimum easily. The experimentalresults show that our algorithm improves the accuracy of cluster greatly.(2) This paper introduces 2D clustering algorithm, namely cluster based on Depthand Density. This algorithm consists of two steps: determining the cluster number andinitial cluster centers respectively, which actually has advantage than K-means algorithmthat depends on the initial values excessively. At last, we evaluate the effectivenessthrough cluster entropy.(3) In this paper, we do the semantic cluster following the visual cluster, such as tosetting apart from these samples with the similar visual features but different semanticfeature by making full use of semantic information. Traditional generation models use alarge amount of "irrelevant" training samples to complete the probability calculation. Ourmodel is not. The model reduces the "semantic gap" effectively.The stage of semantic annotation:(4) Generally, we couldn’t accumulate the semantics of partial regions rigidly as thewhole image annotation. People are always interested in just one or two parts which coverthe main semantic information of the image. This paper introduces salient level assemantic contribution to regional.(5) In order to choice the semantic annotation, this paper not only accumulatinginformation based on the probability distribution function to obtain the threshold, but alsoconsidering the effect of confidence factor. Borrowing Apriori algorithm which belongs todata mining, we acquire association rules during semantics, and then perfect the result ofimage annotation.