Multiview imaging technologies consist of multiple cameras which are usually highly related. In some network settings, it is possible to reduce the operational quality of some cameras yet still achieve high-quality image recovery. Employing low-resolution cameras can greatly decrease the acquisition costs and complexities. The idea of Compressive Sensing (CS) is introduced to accomplish the role of low-quality cameras by operating at a diminished sampling rate. CS imposes a prior distribution on the unknown variables, and allows sparse signal recovery from sub-Nyquist measurements. In this thesis, we investigate the applications of Compressive Sensing via Belief Propagation (CS-BP) theory for low-quality cameras. In more detail, we take advantage of the side information from neighboring views, in improving the performance of BP-based multiview image recovery. The main issue in the original CS-BP is that all unknown variables have the same prior distribution, which is not true in many cases, especially in transformed data. In this thesis, we investigate the applications of multiview technology along with methods on the generalization of the CS-BP. To further improve the CS-BP, we explore the role of larger coefficients of the signal in assigning the pdf sampling step-size. As large coefficients are dominant in step-size determination, the greater the large components are, the less accurate the small components detection is. Thus, we propose methods which deal with DC and other large coefficients to attenuate their influence on the sampling step-size. The proposed method greatly improves the accuracy of signal recovery, as the sampling step-size is maintained at a reasonably small value. In addition, we evaluate the number of large coefficients that are to be eliminated from BP iterations, by introducing an adaptive technique which determines the optimum number of coefficients according to the involving costs and complexities. Application of compressive sensing in multiview technology is relatively a new idea and the experimental results show that the generalized CS-BP can greatly outperform the original CS-BP technique.
展开▼
