Research On Image Semantic Knowledge Extraction And Application Of UGC Combined Image And Text

With the rapid development of the Internet and the dramatically increasing of the netizens number,the social patterns of people have gradually changed.They migrated from offline to online,while people live and work offline,they also carefully manage their second life online.Netizens with the same interests gathered together to form virtual communities,in the virtual communities,most of the content is generated by users,which is called User Generated Content(UGC).UGC is an important source of massive data on the Internet.It contains rich application value and is the premise of the development of the big data eraHowever,UGC itself has some shortcomings.With the continuous development of social media and the continuous improvement of Internet technology,a large number of UGC emerges.Those massive UGC will magnify the shortcomings of UGC itself,which will make it extremely difficult to extract the knowledge users need from UGC.Especially at present,the expression mode of UGC has changed from single text mode into text and image mode.If only the research on text is used for UGC content which combines graphics and text,the available information of text can not fully express the original meaning,and the measured effect of content is not as good as the real effect.Then,how to use image semantic features to solve the lack of text information in UGC is an urgent problemThis paper takes UGC content which combines image and text in virtual community as the research object.According to the theory of image semantics understanding,the semantic knowledge is gradually extracted from visual layer(bottom layer),object layer(middle layer)and conceptual layer(top layer)of image,meanwhile,the combination mode is gradually considered from the state of how to consider image information,to the state of image information as auxiliary information,finally to the state of image and text information as primary information.At the same time,it solves the problems of water-quality content and noisy information which exist in the current virtual community.This paper further expands the existing research on image information extraction and image-text integration,it explores and innovates the quantification methods of different levels of image semantic knowledge,which is paving the way for furture research on multimedia knowledge extraction and quantification.The specific research work is as follows(1)Based on the theory of image semantic understanding,this paper calculates and quantifies the feature semantics of image visual layer(bottom layer),and studies the influence of adding image into UGC content of image-text combination together with other features,at the same time,it solves the problem of how to consider image information in the content of image-text combination.Test on the real dataset of virtual community shows the quantification method performs well.The feature selected result can be used for image selection in the following sections.(2)Based on the feature semantics of visual layer(bottom layer)in image semantic understanding theory,the feature semantics of image object layer(middle layer)is quantified as subjective grading and objective proportion,and the extracted object layer information is used as auxiliary information to solve the problem of data sparsity when text information is taken as the main information.In the real virtual community dataset,the recommendation method with image information is more accurate than that without image information.The feature extracted method can be used for model built in the following section.(3)Based on the feature semantics of visual layer(bottom layer)and object layer(middle layer)in the theory of image semantic understanding,the mapping relationship between the feature semantics of these two layers and image conceptual layer(top layer)is established.Meanwhile,the relationship between the feature semantics of image conceptual layer(top layer)and the probability distribution of text topics is learned and found.The matching degree between image topics and text topics can be obtained through deep topic model.The matching degree is used to solve the problem of inconsistency between graphic and text topics in UGC content.In the real virtual community dataset,the quality prediction method with image-text matching is more accurate than that without adding.The semantics reflection method can be the solver of knowledge extraction when image and text combined deeply.