Adaptive Image Ranking With User Perception Support

With the rapid popularization of digital cameras and surveillance cameras, images and videos have become a common way people record information after text and voice.Usually, people manually screen and classify these images and videos for further search and use. But with the information boom, manual annotation can hardly follow data growth. Thus, computer-generated image ranking is used to help find and index images.In this thesis, we focus on adaptive image ranking with user perception support, which allows computers to imitate image rankings that conform to human perception. Such a ranking method can describe and index images and help user to easily find what he/she needs in mass data.We start from the conventional framework of image ranking. Based on the framework, we study three modules: collecting ranking list, extracting image features and learning ranking functions. By analysing deficiencies of existing methods, we propose new solutions to the problem. The main contributions of this thesis include:1. An adaptive ranking list and its corresponding data collection method are proposed.Such a ranking list can change the number of bins used and is adaptive to different user sensitivities. The hierarchical classification method collect user rankings through repeatedly asking users to divide a set of images into better and worse, and finally organizes all the images in the form of a classification tree. Such a method is used to collect an adaptive ranking list, which is consistent, consumes less time and avoids redundancy.2. Hierarchical-SVM is used to learn an adaptive ranking list. Such methods not only replicate user's original ranking, but also match the user's sensitivity in ranking.Associate-predict model is used to rank pictures of a new person. The ranking function of the associated person is used to rank pictures of the new person.3. Multiple semi-supervised methods are used to extract image features. Image features obtained in this way contain implicit semantic descriptions and are closer to how human describes images compared with simple low-level features.4. Image embedding is directly learnt from user annotated relative similarities. Such methods solve the problem that image distances in a visual feature space do not conform to human perceived image similarities. The iterative sampling-labeling-learning process can greatly decrease human efforts while learning the embedding.