Image Annotation Research Based On Neighborhood Update Under Topic Classification And Weak-label Lifting Scheme With Respect To Neighborhood Diversity

Automatic image annotation(AIA)technologies provide fast and low cost text labels for massive images,its applications in aerospace,military target detection,urban traffic surveillance,biomedical image processing and public security monitoring are convenient and important.However,the complicated relationship between visual contents and scenes will lead to an uncertain mapping between visual features and labels,resulted in the inaccuracy of annotation.In this paper,we mainly focus on the problems of accurate label prediction and weak label weighting.Our contributions include:1.To establish an accurate nearest neighbor subset,a two-step voting neighborhood updating method is proposed based on the feature neighborhoods-topic shrinkage-topic neighborhoods transition.Initially,by the Naive Bayesian classifying the first round KNN neighborhoods,an unlabeled image is determined to seize a most suitable topic and reduce the semantic gap.Comparisons of global lower level feature vector and the mid-level multi-descriptors vector on annotated results,mid-level multi-descriptors vector demonstrate its validness.Secondly,in the selected topic image set of untagged image,the second round KNN classifier is used to update a feature neighborhood for prediction with new ranked label distribution for label prediction.In the condition of test images selected from different topics,compared with Semantic Co-occurrence Model based on global lower level feature vector on annotation results,experimental results prove that our method makes the whole average precision rate of predicted labels increased by 5.37%,average recall rate raised by 27.39%,eventually making the harmonic mean of labels to enhance 17.24%.2.In order to solve the problem of inaccurate labeling results caused by label imbalance,a label-weight lifting scheme with respect to neighborhood diversity is proposed.Firstly,each tag image set is generated by related images under the same tag,and then visual similarity subsets are created by K nearest neighbor classifiers in the same image tag sub to recruit expended similarity subsets from multiple tags.Secondly,in order to improve the important label-weight and predict labels,we assign the label weight according to the feature distance of the nearest neighbor image.Finally,according to label distribution frequencies of the training set,label co-occurrence probability is calculated to lift the weight of predict tags.Compared with the 2PKNN algorithm,the proposed method in the AP,AR,F1,N+ values were increased by 0.70%,1.29%,0.98%,2.Experiment shows that the proposed method can effectively improve the results of automatic image annotation.