Intermediate view reconstruction for multiscopic 3D display

• Main Author: Karajeh, Huda Abdel-Rahim
• Published: Durham University 2012
• Subjects: 621.36
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.549195

This thesis focuses on Intermediate View Reconstruction (IVR) which generates additional images from the available stereo images. The main application of IVR is to generate the content of multiscopic 3D displays, and it can be applied to generate different viewpoints to Free-viewpoint TV (FTV). Although IVR is considered a good approach to generate additional images, there are some problems with the reconstruction process, such as detecting and handling the occlusion areas, preserving the discontinuity at edges, and reducing image artifices through formation of the texture of the intermediate image. The occlusion area is defined as the visibility of such an area in one image and its disappearance in the other one. Solving IVR problems is considered a significant challenge for researchers. In this thesis, several novel algorithms have been specifically designed to solve IVR challenges by employing them in a highly robust intermediate view reconstruction algorithm. Computer simulation and experimental results confirm the importance of occluded areas in IVR. Therefore, we propose a novel occlusion detection algorithm and another novel algorithm to Inpaint those areas. Then, these proposed algorithms are employed in a novel occlusion-aware intermediate view reconstruction that finds an intermediate image with a given disparity between two input images. This novelty is addressed by adding occlusion awareness to the reconstruction algorithm and proposing three quality improvement techniques to reduce image artifices: filling the re-sampling holes, removing ghost contours, and handling the disocclusion area. We compared the proposed algorithms to the previously well-known algorithms on each field qualitatively and quantitatively. The obtained results show that our algorithms are superior to the previous well-known algorithms. The performance of the proposed reconstruction algorithm is tested under 13 real images and 13 synthetic images. Moreover, analysis of a human-trial experiment conducted with 21 participants confirmed that the reconstructed images from our proposed algorithm have very high quality compared with the reconstructed images from the other existing algorithms.