Abstract:
One goal of machine vision is to recover the three-dimensional structure of a scenefrom its two-dimensional projections. Stereo matching is a common method for extractingsuch depth information from stereo image pairs. Detecting depth information is not easybecause of the various image variations between the right and left image views such as noise,occlusion, sampling effect, and intensity distortion. So, a suitable method to measure the similarity of points are necessary. Also, to solve the problem of local minima,we propose a new solution to stereo matching by using parallel multi-resolution in an area based method. In contrast to other related work using coarse-to-fine, multi-resolutiontechniques, the matching at each resolution is performed simul?.aneously to avoid the local minima problem. The energy cost of each pixel at the finest resolution and its projection on the coarser resolutions are calculated. Then, the total energy costis calculated by a weighted average of the energy cost at all resolutions. The weights are decreased continuously from the finestresolution to coarsest resolution. An approach to similarity measurement, called intensity-adjustment method, is employed in the energy cost function to solvethe problem of intensity distortion between images. Finally, the Gibbs Sampler with Simulated Annealing is used as the relaxation method to get the optimal disparity map.The results show an increase in the accuracy of the disparity map. Both synthetic and real image test cases, including cases with ground truth, have been used, and good resultshave been obtained.