Image Tag Refinement Under Social Media Environment

The development of digital photographic equipment and network technology makes the number of Internet images increase rapidly.The appearance of social websites such as Flickr,Zooomr provides a new way for image resource management.Users are allowed to annotate shared images manually,which is an important feature of the websites.The process is known as social tagging,and initial tags provided by users on social media are known as social tags.Social tags play an important role in the management,transmission and sharing of information.However,social tagging is completely free and unfettered,and users have different education and life backgrounds,so it causes-many problems of social tagging.Incomplete tags and low quality tags are two major problems,which limits the application of social tags in the fields of personalized recommendation,image retrieval and others.How to improve social tags and make it describe resource contents in a more comprehensive and accurate way are of great practical significance.It is also a hot issue in computer vision.In combination with machine learning and pattern recognition,we do research on image tag refinement under the social media environment.The main works in this dissertation are as follows:(1)We systematically analyze the major problems in social tags and the causes,and analyze the current research on image tag refinement.Based on the existing research,information from Internet data and dictionary database,we propose a synthesized tag relevance measurement.Compared with traditional methods,the proposed method discusses tag-tag relevance from different aspects(both content relevance and hierarchy correlation).Besides,we fully utilize image feature extraction and semantic analysis technology to explore the relevance between images.(2)In order to solve the problem that social tags are incomplete,we propose an image tag enrichment algorithm based on tag semantic and image visual.The algorithm mainly includes three steps.Firstly,a set of candidate tags of diversity and relevance are recommended based on original tags.Then the relevance between images and candidate tags are measured by analyzing image and tag information.At last,only high correlated candidate tags are reserved.Compared with other existing methods,the algorithm fully integrates the image-image,tag-tag relation.Thus the proposed algorithm can accurately recall the missing tags,and has the advantage of simple calculation and low cost.Experiments are conducted on the MIRFlickr dataset,the results demonstrate that the proposed method can solve the problem of incomplete tags effectively.(3)In order to solve the problem of low quality tags including noisy and ambiguous tags,we propose a novel image tag refinement algorithm.Firstly,the information provided by users is initialized based on random walk model.Then,a mathematical model is constructed on the basis of semantic relevance between tags and images.At last the image tag refinement problem is transformed into a mathematical model solving problem.Compared with other existing tag refinement models,the proposed model systematically considers three aspects:semantic consistency,sparseness of deviation,sparseness of the tagging results.Thus,the proposed model can guarantee the correct correspondence between images and tags,and has a better refinement result.Experiments are conducted on the MIRFlickr dataset,systematically demonstrate the advantage of preprocess,the effectiveness of proposed mathematical model,and the robustness of the proposed algorithm.