随机并行梯度下降图像匹配方法

基于灰度的图像匹配方法具有匹配精度高等优点, 但在实时性方面还存在不足。本文提出一种基于随机并行梯度下降优化算法的快速图像匹配方法,该方法同时对匹配图像间所有的变形参数施加相互统计独立的随机扰动, 然后计算扰动前后待匹配图像间性能评价函数的改变量, 并用这一改变量对变形参数进行更新, 进行迭代运算, 最终实现快速的最优变形参数估计及图像匹配。将SPGD匹配方法用于对仿真和真实的图像序列中的目标进行跟踪,均证明了该匹配方法能进行正常匹配, 匹配精度和最小二乘方法相当, 但速度要较最小二乘匹配方法快。

On the Image Matching Method based on Stochastic Paralelle Gradient Descent

Grey-level-based image matching method has advantage of high precision, however, there are still insufficient in terms of real-time. A fast image matching method is presents in this paper, based on an optimization algorithm of stochastic parallel gradient descent (SPGD). This method simultaneously imposes statistically independent random disturbance on all deformation parameters of images to be matched, and then the variation of performance evaluation function between images to be matched before and after disturbance is calculated. Based on this variation, the deformation parameters are updated and used in iterative calculation. Finally, the fast optimization deformation parameter estimation and image matching are realized. SPGD matching method was used in target tracking for both simulated and real image sequences. Results show that this matching method is able to implement normal matching and its matching precision is equal to that of least squares method, but with faster speed.