A novel inpainting framework for virtual view synthesis

Multi-view imaging has stimulated significant research to enhance the user experience of free viewpoint video, allowing interactive navigation between views and the freedom to select a desired view to watch. This usually involves transmitting both textural and depth information captured from different viewpoints to the receiver, to enable the synthesis of an arbitrary view. In rendering these virtual views, perceptual holes can appear due to certain regions, hidden in the original view by a closer object, becoming visible in the virtual view. To provide a high quality experience these holes must be filled in a visually plausible way, in a process known as inpainting. This is challenging because the missing information is generally unknown and the hole-regions can be large. Recently depth-based inpainting techniques have been proposed to address this challenge and while these generally perform better than non-depth assisted methods, they are not very robust and can produce perceptual artefacts. This thesis presents a new inpainting framework that innovatively exploits depth and textural self-similarity characteristics to construct subjectively enhanced virtual viewpoints. The framework makes three significant contributions to the field: i) the exploitation of view information to jointly inpaint textural and depth hole regions; ii) the introduction of the novel concept of self-similarity characterisation which is combined with relevant depth information; and iii) an advanced self-similarity characterising scheme that automatically determines key spatial transform parameters for effective and flexible inpainting. The presented inpainting framework has been critically analysed and shown to provide superior performance both perceptually and numerically compared to existing techniques, especially in terms of lower visual artefacts. It provides a flexible robust framework to develop new inpainting strategies for the next generation of interactive multi-view technologies.