Leveraging collective knowledge in image folksonomies for enhanced image consumption

Abstract

Current multimedia applications rely on the availability of user-supplied tags in order to retrieve user-contributed images from an image folksonomy. However, the presence of weakly annotated images and non-relevant tags in an image folksonomy hampers the effectiveness of tag-based image retrieval. In this dissertation, we propose novel methods for image tag recommendation in order to mitigate the presence of weakly annotated images in an image folksonomy. In addition, we propose novel methods for image tag relevance estimation in order to mitigate the presence of non-relevant tags in an image folksonomy.First, we propose a maximum a posteriori-based image tag recommendation method, making use of an image folksonomy that consists of images that are visually related to the seed image used. We evaluate the effectiveness of the proposed method for image tag recommendation by means of the average number of true and false positive tags, using the publicly available MIRFLICKR-25000 image set. Our experimental results show that the use of an image folksonomy for image tag recommendation allows taking advantage of a rich and unrestricted concept vocabulary, compared to the use of conventional approaches for image tag recommendation that make use of a training database with a limited number of images and concepts.Second, we propose a sparse representation-based image tag recommendation method, again taking advantage of an image folksonomy. Compared to other sparse representation-based methods for image tag recommendation, the proposed method estimates the relevance of a tag with respect to the content of a seed image by taking into account both the local and global distribution of sparse coefficients. We evaluate the effectiveness of the proposed method for image tag recommendation by means of two experiments, taking into account two different imbalanced tag distributions. Our experimental results show that the proposed method allows for more robust estimation of the r...