基于全局优化的图像修复及其在GPU上实现

针对目前贪婪修复算法可能存在修复效果视觉不一致以及全局优化修复算法中未考虑结构信息的情况,提出一种新的基于全局优化的快速图像修复算法.通过定义出改进的能量函数,把图像修复问题转化为全局优化问题,并采用基于非局部均值的状态标签裁减算法,大幅度减少图中每个节点可能的状态标签数,从而大幅度降低优化算法复杂度；同时,利用图形处理器（GPU）进行加速,进一步提高了运算速度.与其他贪婪合成和最优化修复方法相比,该方法速度更快且较好地保持了纹理和结构的整体一致性.

Global optimization based image inpainting and its implementation on GPU

Since greedy image inpainting algorithms might cause visual inconsistency, and global optimization based inpainting algorithms didn’t consider structure information, a new fast image inpainting algorithm based on global optimization is proposed. The image inpainting was formulated as a global optimization problem by defining an improved energy function, and a non-local mean based label pruning algorithm was adopted to cut down the number of labels for each node so as to reduce the complexity of the optimization algorithm significantly. Graphic processing unit (GPU) was used to further improve the computing speed. Compared with greedy synthesis and global optimization based image inpainting methods, the proposed approach not only avoids the inter-block discontinuity but also achieves better efficiency.