基于区域边界约束和图割优化的稠密匹配算法

摘要

Two main challenges of stereo matching algorithm via graph cuts global optimization are discontinuity and occlusion problems.Energy function with convex smooth term has a global optical solution but is over-smoothing at the boundary of scene object,while energy function with non-convex smooth term preserves discontinuity but just has a se-cond global optical solution via iterative optimization, and problems of occlusion, orderingness, uniqueness are not dealt with appropriately as well.Therefore, with the observation that disparity almost jumps at the color discontinuity, this paper proposed an approach of energy function presentation for dense stereo matching based on restrictions between re-gion boundary pixels, which assured not only a global optimal solution but also discontinuity-preserving, meanwhile treated with issues of occlusion, orderingness, uniqueness and greatly improved computing efficiency.%基于图割全局优化的稠密匹配算法中,平滑项表示和遮挡处理是关键问题.具有凸性平滑项的能量函数可以求得全局最优解,但所求结果在视差跳变处过于平滑;而具有非凸平滑项的能量函数虽保留了视差的非连续性,但目前只能使用循环算法求得次优解.为此,基于"视差跳变绝大部分发生在颜色的不连续处"这一设定,提出一种利用区域边界和边界像素间的约束构建能量函数的稠密匹配表示方法,使得该函数既能求得全局最优解,又能使最终结果满足平滑项的"非连续保留"性,且体现遮挡约束、顺序性约束,并显著提高计算效率,在速度和效果上取得较好平衡.