Semantic Automatic Annotation Of Incomplete Image Based On Low-rank Sparse Decomposition And Label Correlation

With the increasingly mature of social network and multimedia technology,more and more netizens upload their images to the network space,and the network image resources become more and more abundant.In the face of so many images,how to allocate appropriate semantic tags for each image is an urgent problem to be solved.In particular,due to the existence of synonyms and the different focus of each netizen,the semantic labels of images uploaded manually are often incomplete.It is because of this incompleteness that the image retrieval results based on semantic tags are not good,so how to effectively complete the image semantic tags has become an important research topic.Low-rank sparse decomposition is an effective method to reduce noise.In order to improve the accuracy of the completeness of semantic labels in images,this paper studies the semantic automatic labeling method of incomplete images which integrates lowrank sparse decomposition and tag correlation.The main contents include the following three aspects:(1)A low-rank sparse decomposition model is presented.This model mainly generates a low-rank sparse mapping matrix representing the relationship between the visual features and the label features of the incomplete image,and obtains the missing candidate labels by the product of the low-rank sparse mapping matrix and the visual features,so as to complete the labels.In order to make the mapping matrix of the model more accurate and to prevent overfitting,the neighbor image set of the incomplete image is further used to establish the low-rank sparse decomposition model.Firstly,the neighbor imageset is found by combining the visual feature and label feature of the incomplete image,and then a low-rank sparse mapping matrix is fitted on the neighbor imageset by the model.Because the mapping matrix represents the common relation between the visual features and the label features of all the neighboring images,the candidate labels of the incomplete images are obtained by using the mapping matrix and the visual features of the incomplete images.Finally,experiments on Corel5 k and Flickr30 Concepts dataset verify the validity of the model.(2)An automatic semantic annotation method for incomplete images is proposed which integrates low-rank sparse decomposition and tag correlation.In order to make the obtained low-rank sparse mapping matrix more accurate,a label correlation vector is incorporated into the low-rank sparse decomposition model.Firstly,a method of tag co-occurrence frequency is used to calculate the correlation probability between the candidate tag and the incomplete image.Then all the correlation probabilities are combined into a correlation vector and incorporated into the low-rank sparse decomposition model.Finally,experiments on Corel5 k and Flickr30 Concepts dataset verify the effectiveness of the method.(3)A semantic automatic annotation system of incomplete image is realized.According to the above research results,on the Matlab platform,a semantic automatic annotation system of incomplete image is designed and implemented which integrates low-rank sparse decomposition and tag correlation.