Researches On Automatic Image Annotation

With the diversification of technologies such as the Internet,cloud computing,artificial intelligence,and big data,the number of images on the Internet has increased dramatically.This has created a problem that has attracted much attention from users: how to efficiently retrieve images that meet the needs in these massive images.The effect of image retrieval depends heavily on image semantic annotation.However,due to the existence of "semantic gap" between underlying image features and high-level semantics,it still has a great challenge for image semantic labeling.In response to this problem,this article has carried out in-depth research on image auto tagging and has mainly done the following work:(1)Based on the image classification model of the support vector machine and the image annotation model of the fusion semantic topic,this paper proposes an image annotation method that fuses the probabilistic semantic analysis model of image classification information.The advantages of the two are complementary.First,the image classification based on the support vector machine compensates for the problem that the image feature data in the probabilistic semantic analysis model still loses important information in the process of quantization.Second,the classification model based on the support vector machine The color features of the image are extracted one by one,and the probabilistic semantic analysis model is only a general feature for extracting the underlying features of the image.After fusion,it can effectively avoid the adverse effect on the image annotation result caused by improper feature extraction methods.After the fusion of the two,the probabilistic set of candidate annotation words of the image is obtained.Experiments show that the performance of the image annotation has been significantly improved.(2)In order to further optimize and improve the quality of image annotation,this paper proposes an image annotation method based on similarity model.The idea of this method is: using image similarity algorithm to calculate the similarity between the image to be annotated and the annotated image set.Degree,statistics of the similar images of text labels and numbers,generate an image of the similarity weights,into the candidate markup word probability concentration,the optimization of the labeling results,experiments show that the labeling performance on the basis of the original has been significantly improved.(3)Based on the proposed annotation model,a micro-system for image annotation and retrieval is developed,which includes three modules: image model training,image annotation and image retrieval.Firstly,the image annotation model is trained by using the training set of the image data set,and then the annotation image is marked.Finally,the image can be retrieved by the text keyword or the image instance.This paper uses Corel5 k,Espgame and Iaprtc12 image sets to conduct simulation experiments,and uses the precision rate,recall rate and comprehensive evaluation index to evaluate the labeling results.And compared with the translation model TM,cross-media related model-CMRM,continuous space related model-CRM,PLSA-WORDS,PLSA-FUSION,RNN,ANNOR-G and FFSS and other classic annotation model.Experiments show that the method proposed in this paper has achieved good results.