Image Tag Ranking Based On Sparse Coding Algorithm

In today's fast internet era,there have been a rise in the popularity of numerous digital imaging devices as well as improvements in Internet technologies.Digital images can be easily created,uploaded,shared and distributed on the internet.Unlike the situation about a decade ago,web images are nowadays playing a more important role in our lives.Web image search has become an active and rather challenging research issue.Lately,the internet world is witnessing an unprecedented increase in "community contributed" multimedia content available online(e.g.Facebook,YouTube,Flickr,and Zoomr etc.).These social media sources allow users to upload their personal multimedia data and annotate the content with descriptive keywords which are called tags.With these tags,users may conveniently organize and access shared multimedia content.Tagging images aid in the better organization of image search and retrieval.Furthermore,ranking the tags of images increases the effectiveness of image search and retrieval.The recent advancement in internet speed and flexibility of interfaces provided by social media providers have significantly stimulated multimedia sharing,exchange,and propagation among users.The attractiveness of these social multimedia platforms has tremendously invited an enormous mass of users to tag and share multimedia on the internet globally.A major difficulty for web image search and retrieval is that most images are typically not annotated with proper tags,and many of them are even completely unlabeled.In addition,even for the annotated images,their associated tags could be noisy,irrelevant and often incomplete in describing the contents of the images.The main purpose of this project is to improve the level of relevancy of tags.The dataset used for the project was obtained from Flickr.However,prevailing studies has revealed that so many tags provided by Flickr users are ambiguous or imprecise thus approximately 50%of tags are not actually related to the images they are assigned to.Also,the relevance levels of the tags cannot be obtained from the initial tag list of the images,where the order is mostly according to the input arrangement and carries very little information about its relevancy.This project refines tags associated with an image according to its relevancy using sparse coding.Sparse coding is originally used in the field of signal processing.Sparse coding has recently shown significant improvement in computer vision or multimedia processing,hence the decision to try it on tag ranking due to its effective and robust nature.The sparse coding algorithm used in this project was Least Absolute Shrinkage and Selection Operator(LASSO).It was used to build a dictionary of relevant neighbors in the dataset to reorder the tags according to its relevancy to the images it has been assigned to.However,due to budget constraints,a lesser data size was used,but it still showed an increase in the performance of sparse coding.Therefore,running the algorithm suggested in this project with a bigger dataset in the future may still show highly accurate results.The Combination of sparse coding with other probabilistic approaches improves tag ranking.The sparse coding algorithm was initially used to run the data set to build a sparse representation of the tags.After which a random walk refinement was applied to further improve the results.The LASSO algorithm used in this project showed a significant improvement in the results as compared to the previously proposed probabilistic algorithms by other researchers.The application of LASSO and Random Walk Refinement together show much better results than the use of only LASSO or random walk refinement.