Abstract
Image registration based on feature is a commonly used approach due to its robustness in complex geometric deformation and larger gray difference. However, in practical application, due to the effect of various noises, occlusions, shadows, gray differences, and even changes of image contents, the corresponding feature point set may be contaminated, which may degrade the accuracy of the transformation model estimate based on Random Sample Consensus (RANSAC). In this work, we proposed a semi-automated method to create the image registration training data, which greatly reduced the workload of labeling and made it possible to train a deep neural network. In addition, for the model estimation based on RANSAC, we determined the process according to a probabilistic perspective and presented a formulation of RANSAC with the learned guidance of hypothesis sampling. At the same time, a deep convolutional neural network of ProbNet was built to generate a sampling probability of corresponding feature points, which were then used to guide the sampling of a minimum set of RANSAC to acquire a more accurate estimation model. To illustrate the effectiveness and advantages of the proposed method, qualitative and quantitative experiments are conducted. In the qualitative experiment, the effectiveness of the proposed method was illustrated by a checkerboard visualization of image pairs before and after being registered by the proposed method. In the quantitative experiment, other three representative and popular methods of vanilla RANSAC, LMeds-RANSAC, and ProSAC-RANSAC were compared, and seven different measures were introduced to comprehensively evaluate the performance of the proposed method. The quantitative experimental result showed that the proposed method had better performance than the other methods. Furthermore, with the integration of the model estimation of the image registration into the deep-learning framework, it was possible to jointly optimize all the processes of image registration via end-to-end learning to further improve the accuracy of image registration.