摘要
三次B样条函数拟合小形变需要大量控制点,且非刚性配准的迭代算法和归一化互信息计算量巨大,使得非刚性配准缓慢.为了提高配准速度,提出基于B样条函数的二级并行算法,其中对归一化互信息使用数据并行算法;对梯度下降流使用任务并行算法,并将数据并行算法嵌入到任务并行算法中.为减少计算量,提出图像多层次局部熵提取自由形变场活动控制点的算法,使活动控制点仅分布于待配准的目标之上,并使用B样条系数的快速算法进一步减少计算量;对由于控制点分布优化造成的各线程块并行计算量不平衡的问题,使用类似于Greedy算法的计算平衡算法使各线程块的计算量均衡.实验结果表明,使用B样条系数快速算法可以减少约50%的B样条系数计算量;与串行算法相比,使用二级并行算法以及控制点分布优化算法可以达到60～80倍的加速效果;比现有的数据并行配准算法可提速约6倍.
The non-rigid registration is slow due to large number of control points and the high cost of iterative strategy and the normalized mutual information（NMI）.A parallel algorithm with a B-spline coefficient optimization is proposed to accelerate such registration.In this approach,the data parallel algorithm computes NMI and the task parallel algorithm,in which the data parallel algorithm is embedded,computes the gradient descent flow.Control points are restrained to be distributed on the targets according to the image local entropy for further reduction on computational cost.A balanced algorithm is presented to solve the computational imbalance problem caused by the uneven distribution of control points.Experiments showed that the use of the B-spline coefficient optimization can reduce about 50% coefficient computation.The introduced parallel algorithm can accelerate the non-rigid registration about 60~80 times compared to the serial version,about 6 times compared to the existing data parallel approaches.
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2012年第4期485-493,共9页
Journal of Computer-Aided Design & Computer Graphics
基金
中国科学院知识创新工程重要方向(KGCX-YW-909-1)
关键词
并行算法
归一化互信息
梯度下降流
局部熵
parallel algorithm
normalized mutual information
gradient descent flow
local entropy
