Image Tag Completion Based on Low-Rank Sparse Decomposition and Optimization

Due to the randomness of user semantic annotation,a large number of uploaded network images result in incompleteness of image tags,which greatly reduces the efficiency of image retrieval.Low-rank sparseness is an effective method to reduce data noise.To improve the completeness of image semantic labeling,this paper proposes a method of image tag completion based on low-rank sparse decomposition and optimization(LRSDO).Firstly,the features and semantics of an incomplete image are combined to search its nearest neighbor image sets.Secondly,the mapping relationship between features and semantics of the image is obtained by low-rank sparse model,and then its candidate labels are predicted.Finally,in order to achieve more accurate image label completeness,candidate labels are optimized by using an individual-oriented label co-occurrence frequency method.Experiments on benchmark dataset Corel5K and real dataset Flickr30Concepts show that our method has better performance in image tag completion.