| One of the reasons for the power in human vision system is its ability to combine information from different visual processes. Fusion of visual information of multiple visual sensors can resolve some of the unsolved complex computer vision problems. The thesis investigates the problem of information fusion in visual reconstruction, and visual recognition. In reconstruction fusion, the reconstruction of a single sensor information is improved by the information of other sensors, and in recognition fusion, the recognition of patterns is contributed by the information from other multiple sensors.; Regularization techniques have been developed to reconstruct visual data consisting of discontinuities. Rotational symmetric continuity stabilizers have been utilized as continuity stabilizers for visual reconstruction. These stabilizers often result in oversmooth solution, and it has been difficult to make a compromise between the discontinuity detection and smooth reconstruction. To resolve this problem, we introduce directional stabilizers for visual reconstruction. 1-D stabilizers are fitted in orthogonal directions, and bound together to form directional stabilizers in multidimensional space. With directional stabilizers, the energy penalties to appear as a step discontinuity, or a crease discontinuity can be specified considering the neighborhood of the discontinuity. This facilitates the extension of regularization approach to information fusion in visual reconstruction.; A class of neural networks, which self-organizes to recognize visual discontinuity patterns is considered. This includes both neocognitron and MARA. These networks are hierarchical. The cells at the lower level of the hierarchy self-organize to recognize simple features and the cells at the higher level self-organize to recognize complex patterns which are combinations of low level features. We propose a neuromorphic model for fusion of activities in such networks processing the information relating to the same visual pattern. The fusion process increases the recognition ability and the reliability of the networks. |
