Research And Implementation Of Image Advanced Semantic Annotation Algorithm Based On The Attention Model

How to manage and retrieve huge image resources on the Internet is the focus of multimedia technology research.Interest mining and image annotation are the key ways to solve this problem.Earlier image annotation techniques were limited by the size of the data set,and the labeling performance is not high.Existing image tagging technology does not need to build a mapping model based on the data set,it only needs to find labeled images with similar image characteristics on the Internet,and use the semantic information of these labeled images to achieve the labeling of unknown images.However,the quality of data set annotations on the Internet is uneven,and these annotations are intended for all Internet users and do not have individual characteristics.Therefore,the annotation results have many differences from human psychological expectations.Personalized image annotation needs to solve several key issues:first,how to describe the evolution of user interest;then,how to personally identify the area of image annotation;and finally,how to bridge the semantic gap and implement advanced semantic annotation of images based on the relationship between images and tags.In response to the above-mentioned key scientific issues,the main work and innovations completed in this thesis are as follows:(1)Aiming at the problem of describing the evolution of user interest,a user interest mining method based on the Deformable Interests Model(DIM)is proposed.This method uses the Interests Tracking Model(ITM)to map labeled words to frequent pattern space.Considering the time factor,the user's long-term interests are mined.the user's long-term interests and contextual interests are combined through a Deformable Interests Model(DIM)to comprehensively and accurately mine user interests.On this basis,this method uses the real-time influence update mechanism to adaptively update the influence of long-term interest,short-term interest,and hot topics on user interest.Experimental results show that the method can mine user interests in real time,comprehensively and accurately.(2)Aiming at the problem of personally identifying labeled areas and generating corresponding high-level semantics,a high-level semantic generation algorithm(Local-HSNN)based on a cross-modal local neighbor analysis model is proposed.First,the algorithm determines the objects that generate high-level semantics.Then,the algorithm calculates cross-modal feature similarity.Finally,Local-HSNN learns strong ranking model of high-level semantic generation to generate image regions and semantic keywords that meet the observation task.The experimental results show that this method can accurately identify the labeled area and generate the corresponding semantic keywords accurately.(3)We design and implement an image high-level semantic annotation system based on the attention model,which crosses the"semantic gap" and realizes high-level image semantic tagging.The image advanced semantic annotation system is implemented by B/S architecture,which has functions of automatically generating personalized advanced semantic annotation,rich visualization,and comprehensive user management.It has high security and fast response speed.