积分图像的快速GPU计算  

Fast integral image computation on GPU

作　　者：王志国[1] 王贵锦[1] 施陈博[1] 苗权[1] 林行刚[1] 

机构地区：[1]清华大学电子工程系,北京100084

出　　处：《计算机应用研究》2011年第10期3913-3916,共4页Application Research of Computers

摘　　要：提出了一种在GPU上计算积分图像的方法。积分图像可通过对输入图像的行实行前缀加法后再对列实行前缀加法构建。前缀加法是指对于一个数组,求取起始位置至每一个下标位置的数组元素的和的操作。提出了分段前缀加法原理,当将其运用到GPU图像积分时有如下优点:减少了线程间的数据依赖;降低了内存访问开销;提高了GPU线程的工作效率。提出的算法相对以前算法在速度上提高了约两倍。该算法可运用到使用积分图像的图像处理算法的GPU加速中。This paper described an algorithm about parallelly constructing integral image on GPU.Integral image can be built through prefix sum on rows of input image followed by prefix sum on columns.Prefix sum is an operation on an input array in which each element in the result array is the sum of the elements in the input array from the beginning to its index.This paper introduced algorithm of segmented prefix sum,and while applied to integrating image on GPU,there were seveval advantages as follows：data dependency between GPU threads was reduced;expenditure of memory access was cut off,work efficiency of GPU threads was promoted.Proposed algorithm is about 2 times faster than its counterparts,and it can be applied to GPU acceleration of those image processing algorithms in which integral image is used.

关 键 词：积分图像 前缀加法 计算统一设备架构 并行计算 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]