Learning users' interest to assist image browsing and searching

In this thesis, we first investigate how to improve user experience of browsing large photos on mobile devices. Currently, the predominant methods for accessing large photos on small devices are down-sampling or manual browsing by zooming and scrolling. Image down-sampling or thumbnail view results in significant information loss due to the excessive resolution reduction. Manual browsing can avoid information loss but it is often time-consuming for users to catch the most crucial information of an image. Employing the concept of information asymmetry, the thesis investigates an image attention model and its extensions under different application scenarios. The image attention model can successfully solve the small screen browsing problem. First of all, different parts of an image contain different information, computer vision technology can analysis image and extract the attention objects to build the image attention model. Secondly, browsing large photo on small display can be formulated as a problem of optimized manipulating of attention objects to improve viewer's browsing experience, which brings about new functions such as automatic browsing and interactive browsing. Thirdly, the screen limitations of the mobile devices will force users to scroll and zoom while browsing images, which can be regarded as direct indications for users' attention, therefore new attention objects can be discovered by analyzing the past browsing log. Lastly, since psycho-physiological research tells us that the importance of an attention object may vary with the users' preferences and tasks, so attention model should also be adaptive to the users' interest. The thesis also investigates the problem on how to improve the usability of the current image search engines, that is, how to explore the image search results on desktop PCs and mobile devices in a more effective way. The research results show that unsupervised learning technology can help improve the browsing efficiency. The similarity-based presentation method and clustering-base method are better for desktop PC platform and mobile platform respectively. The navigation operations are re-designed to make the new interface easy to use. Based on these results, the thesis also tries to investigate the problem on how to incorporate users' interaction in content-based image retrieval relevance feedback process. The new approach employs the parameter embedding visualization, multi-classifier learning and fusion methods to provide a seamless integrated browsing, user feedback and classifier learning process. (Abstract shortened by UMI.)