Research Of Image Registration Based On Deep Learning

Image registration refers to the process of finding corresponding pixels across two or more images.It is a basic problem in computer vision and is a significant prerequisite for many other tasks.In the field of robotics and autopilot,it is an essential premise for relative pose estimation,simultaneous localization and mapping(SLAM),and structure from motion(SfM).As for remote sensing,it is usually applied in transformation estimation,which is subsequently used for image fusion and change detection after the images are aligned.Therefore,it is necessary to work in the image registration problem.Classic image registration algorithms are mainly based on handcrafted features,including popular SIFT and ORB,which are robust to rotation,scale as well as illumination changes to certain degree.However,these algorithms perform poorly in complicated cases such as viewpoint changes,repeated textures and occlusion.With deep learning adopted to image processing,learning-based features are used for image registration,which are more potential to address these issues.Accordingly,the learning-based approaches are mainly explored in this thesis and corresponding methods are proposed aiming at the drawbacks.Here are the contributions:(a)One image registration algorithm based on Multi-Scale Feature Combination as well as Medium-Layer Discrimination Constraints are proposed for the sake of the contradiction between feature discrimination and invariance.(b)Multi-Orientation Feature Aggregation is adopted to improve rotation-invariance of local features which are limited to the asymmetry of the kernels in deep convolution neural networks.(c)Knowledge distillation is introduced and the Rotated Kernel Fusion is developed for improvement of the student network,which further enhances the learning capability of the student and boosts the rotationinvariance of local features.(d)Reparameterization is applied in the inference stage for the improved student network in(c),with which the performance is enhanced without abating efficiency.