Research On Tag Propagating And Tag Refining In Nearest-neighbor-based Image Annotation

The explosive growth of image resources in the Internet has spawned the efficient image retrieval technology.Automatic image annotation(AIA)is a key issue in image retrieval,assigning keywords to an image based on its mapping relationship between visual features and semantic labels.However,to annotate an image in line with its semantics is still one of the difficulties in AIA.In this paper,we study the image annotation algorithm based on nearest neighbor model,and focus on tag propagation and tag refining problems.Main works are as follows.Aiming at the problem that the latent relationship between visual features and labels cannot be fully excavated in traditional nearest-neighbor-based methods,a reverse label reconstruction method based on Multi-view Non-negative Matrix Factorization(MultiNMF)is proposed.MultiNMF can relate visual features and labels of the image by consistent clustering.Firstly,because of the consistent clustering characteristics of MultiNMF on visual features and semantic feature,the label propagation process of the nearest neighbors is conducted by the label propagation method based on MultiNMF.Secondly,considering the ability of measuring the importance of labels in image datasets of Term Frequency-Inverse Document Frequency(TF-IDF),the TF-IDF is proposed to correct the labeling probability of each labeled word in the test images to improve the annotation results.Finally,on the foundation of the reversibility of the similarity relation between the nearest neighbors,nearest neighbors are selected by the method based on forward and backward selection.When compared to 2PKNN algorithm without metric learning,the experimental results in Corel 5K database show that the AP of the annotation method based on MultiNMF was decreased by 1.3% and the N+ is 19 less,while the AR and F were raised by 3.2% and 0.8% respectively.In order to improve the accuracy of low-frequency labels in an image dataset,an annotation method is proposed.The asymmetric relationship of semantic symbiosis can improve the labeling results of low-frequency tags.Firstly,because the similarity relationship between tags can be measured effectively by the semantic asymmetry of label semantics,a tag refining method of random walk model based on semantic asymmetry of label semantics is proposed to modify the probability of labels achieved by the nearest-neighbor-based method,improving the annotation results of low-frequency labels.Secondly,considering the fact that constraint relationships of visual features and semantic labels can guide the completion of image labels,an algorithm based on multiple constraints is introduced to enrich the original semantic labels of the images in dataset.When compared to 2PKNN algorithm without metric learning,the experimental results in Corel 5K database show that the annotation performance of the low-frequency labels in this method were superior to the traditional labeling methods,where the AP,AR and F were increased by 1.2%,2.7% and 1.9% respectively,the N+ was 5 more.At the same time,the AP,AR and F of all the labels in this method were increased by 1.5%,3.3% and 2.3% respectively,the N+ was 8 more.