Non-Rigid Point Set Registration Based on Variational Bayes Hierarchical Probability Model

被引：0
|
作者：
He Q.-Q. [1 ,2 ]
Lin G. [1 ,2 ]
Zhou J. [1 ,2 ]
Yang Y. [1 ,2 ]
机构：
[1] School of Information Science and Technology, Yunnan Normal University, Kunming
[2] The Laboratory of Pattern Recognition and Artificial Intelligence, Yunnan Normal University, Kunming
来源：
Jisuanji Xuebao/Chinese Journal of Computers | 2021年 / 44卷 / 09期
基金：
中国国家自然科学基金;
关键词：
Adaptive global-local constraint strategy; Bayesian linear regression; Non-rigid point set registration; Tree-structured mean-field; Two-stage prior annealing schedule; Variational Bayes hierarchical probability model;
D O I：
10.11897/SP.J.1016.2021.01866
中图分类号：
学科分类号：
摘要：
Non-rigid point set registration is fundamental research in computer vision and pattern recognition. The current popular non-rigid point set registration algorithms cannot accurately estimate correspondences between two point sets with lots of outliers, noises, occlusion, rotation, and large deformation. In this paper, we alternately perform correspondence estimation and transformation updating to gradually recover correspondences. In the correspondence estimation, we first construct a variational Bayes hierarchical probability model (VBHPM) based on a finite student-t distribution latent mixture model (TLMM) and divide it into the correspondence estimation component and outlier aggregation component, which are respectively used to estimate correspondences and to cluster outliers. Meanwhile, the Bayesian linear regression is used to resist noises. Besides, we add the prior distribution of Dirichlet to dynamically adjust the mixture proportion, and assign a smaller mixture proportion to the points with occlusion to maintain the structural integrity of the point set. In the transformation updating, we iteratively update model parameters based on the variational Bayes(VB) framework and propose the tree-structured mean-field to maintain dependencies among parameters, to obtain a tighter variational lower bound. In addition, an adaptive global-local constraint strategy is proposed to maintain the structural stability of point set, resist deformation and rotation and realize the local-global constraint process. Finally, we adopt the two-stage prior annealing schedule where the prior distribution of Gamma is used to dynamically adjust the precision in the annealing to implement the coarse-to-fine registration. In the experiment, the performance of VBHPM is tested, and the registration results in point set and image registration demonstrate that VBHPM can achieve accurate registration results and higher precision compared with thirteen state-of-the-art algorithms. © 2021, Science Press. All right reserved.
引用
收藏
页码：1866 / 1887
页数：21
相关论文
共 31 条
  • [1] Zhao M, An B, Wu Y, Et al., RFVTM: A recovery and filtering vertex trichotomy matching for remote sensing image registration, IEEE Transactions on Geoscience and Remote Sensing, 55, 1, pp. 375-391, (2016)
  • [2] Schaerer J, Fassi A, Riboldi M, Et al., Multi-dimensional respiratory motion tracking from markerless optical surface imaging based on deformable mesh registration, Physics in Medicine & Biology, 57, 2, (2011)
  • [3] Nguyen T M, Wu Q M J., Multiple kernel point set registration, IEEE Transactions on Medical Imaging, 35, 6, pp. 1381-1394, (2015)
  • [4] Ma Jiayi, Zhao Ji, Tian Jinwen, Et al., Robust point matching via vector field consensus, IEEE Transactions on Image Processing, 23, 4, pp. 1706-1721, (2014)
  • [5] Zhang S, Yang Y, Yang K, Et al., Point set registration with global-local correspondence and transformation estimation, Proceedings of the IEEE International Conference on Computer Vision, pp. 2669-2677, (2017)
  • [6] Qu H B, Wang J Q, Li B, Et al., Probabilistic model for robust affine and non-rigid point set matching, IEEE Transactions on Pattern Analysis and Machine Intelligence, 39, 2, pp. 371-384, (2016)
  • [7] Jian B, Vemuri B C., Robust point set registration using Gaussian mixture models, IEEE Transactions on Pattern Analysis and Machine Intelligence, 33, 8, pp. 1633-1645, (2010)
  • [8] Ma Jiayi, Qiu Weichao, Zhao Ji, Et al., Robust l2e estimation of transformation for non-rigid registration, IEEE Transactions on Signal Processing, 63, 5, pp. 1115-1129, (2015)
  • [9] Tsin Y, Kanade T., A correlation-based approach to the robust point set registration, Proceedings of the European Conference on Computer Vision, pp. 558-569, (2004)
  • [10] Ma Jiayi, Jiang Xingyu, Jiang Junjun, Et al., LMR: Learning a two-class classifier for mismatch removal, IEEE Transactions on Image Processing, 28, 8, pp. 4045-4059, (2019)
1234