A Geodesic Landmark Shooting Algorithm for Template Matching and Its Applications
A Geodesic Landmark Shooting Algorithm for Template Matching and Its Applications
复制标题
DOI：
10.1137/15m104373x
复制
发表时间：
2017-01-01
影响因子：
2.1
通讯作者：
Lee, Long
中科院分区：
文献类型：
--
作者：
Camassa, Roberto;Kuang, Dongyang;Lee, Long
本文提出了一种有效的地标射击模板匹配算法及其应用。该算法的新颖之处包括使用一个常数矩阵来更新测地线射击的搜索方向，而不是传统的前后向积分的方法来更新梯度或牛顿优化，并使用一个非光滑的圆锥核的粒子系统，加速匹配的收敛。为了研究沿着翘曲算法计算的输出量的使用，例如哈密顿度量和动量场，我们引入了一种多尺度分解方法，该方法分离了与变形相关的动量和哈密顿度量的尺度/分量。我们在数值上探索使用分解的哈密顿度量和动量向量作为神经网络的聚类/分类分析的输入特征向量的潜力。我们的数值实验结果令人鼓舞。
We present an efficient landmark shooting algorithm for template matching and its applications. The novelties of the algorithm include the use of a constant matrix to update the search direction of the geodesic shooting, instead of the traditional methods of forward-backward integration for updating the gradient or Newton's optimization, and the use of a nonsmooth conic kernel for the particle system that accelerates the convergence of matching. To investigate the usage of the output quantities computed along the warping algorithm, such as the Hamiltonian metric and the momentum field, we introduce a multiscale decomposition method that separates the scales/components of the momentum and the Hamiltonian metric associated with the deformation. We numerically explore the potential of using the decomposed Hamiltonian metric and momentum vectors as input feature vectors into neural networks for clustering/classification analysis. The results of our numerical experiments are encouraging.