Image Browsing And Retrieval Method Based On Personalized Recommendation

Web image resource is increasing explosively in today's Internet, many researches are focusing on how to manage such huge amount of images and provide the users access to their desired images. Basic solution includes image retrieval and image browsing, and both of them are based on image annotation. This paper involves the methods about image annotation, image retrieval and image browsing.For image annotation, we proposed a novel method to improve image classification performance by leveraging the spatial context of image concepts. It consists of two layers of classification. The first layer applies automatic semantic annotation to image local patches, and the second one combines the local semantic annotation and the spatial layout of them into a single vector called Concept Layout Vector (CLV) as the feature to do the final classification. The experiments on both scene image dataset and animal image dataset show quite positive results.For image retrieval, we proposed a semantic-network-based translation method for cross language image retrieval query translation. The semantic network is computed based on large scale bilingual parallel corpus. It takes semantic concepts as nodes, and computes the edge weight using co-occurrence probability and semantic relation. By considering that different semantic concepts of one polysemous term would have different neighbor nodes in the semantic network, we built our translation model by using both single node and its neighbors in the semantic network and effectively solved the problem of polysemous term. After the query translation, the semantic network also provides semantic expansion of the query to further improve the retrieval results. The experiment comparing our method with bilingual dictionary shows the advantage of our method.For image browsing, we proposed a personalized recommendation-based image browsing method, which is different from other methods that focus on visualization. This method combines text-based image retrieval technique as the basic tool, the ontology theory as organization method of image concept, user preference modeling for personalized experience and recommendation scheme to provide the users an image browsing service with semantic coherence and free choice. Enlightened by the association process of human beings described in psychology, we designed a recommendation method that consisted of two parts, that is, collection and decision. The collection part collects candidate images from large scale image dataset based on user's choice and the modeling of the user's preference. Ontology is used to represent the semantic concepts in image and the relations between them. Then semantic expansion and TBIR are applied to collect to images from dataset. The decision part recommends images to users based on the candidate images and the user preference. A hyper-graph is used to model the candidate image set, and a random walk algorithm is applied to rate each candidate image. The images with local maximal rating are output to the user as recommendation. Based on our user testing result, the proposed personalized recommendation-based image browsing method has gained quite positive results from the measurement of users'rating, browsing depth and coherent rate.