A Remote Sensing Image Registration Algorithm Based On Multiple Constraints And A Variational Bayesian Framework

Aiming at complicated remote sensing image registration problem, we proposed a multi-constraint registration algorithm based on variational inference. Gaussian mixture model is used to construct the entire Bayesian network, which facilitates the estimation of heteroscedasticity noise in mapping function regression, and distribution of outliers in scene point density estimation. We use a variational parameter proxy to approximate complex posterior distribution. In addition, while maintaining the stability of global structure, similarity of local structure is restricted. In order to limit the cost of image conversion, we added constraints on spatial curvature for image. Accuracy and smoothness of the spatial transformation can be maintained by minimizing the curvature in horizontal and vertical directions. Experimental results show that compared with other point set matching algorithms, this method can achieve better results in robustness and matching accuracy.