Research On Key Techniques Of Robust Remote Sensing Image Feature Matching

Image feature matching is a fundamental and crucial issue in photogrammetry and remote sensing.Its goal is to extract reliable corresponding features from two or more images with overlapping regions.Image matching technology not only has very important applications in photogrammetry and remote sensing,but also widely used in the field of computer vision,pattern recognition and artificial intelligence,robot vision,and medical image analysis.It has always been a hot research topic in various fields.However,image matching,especially remote sensing image matching,is still an ill-posed problem.There are many uncertainties.In general,a remote sensing image pair may contain scale,rotation,radiation brightness,blurring,or even temporal differences.These huge differences in geometry and radiation pose a daunting challenge to the current remote sensing image matching algorithms,which result in a substantial reduction in matching performance and make it difficult to meet the requirements of ever-changing practical application.Therefore,it is very important to study the more efficient,universal and robust image matching algorithms.Analysis of the traditional image feature matching methods shows that to achieve universal robust image matching,three difficult problems need to be solved:(1)the robustness to various geometric and radiation differences;(2)more efficient and robust outlier detection models;(3)non-rigid deformation image matching problem.Based on these three aspects,this paper proposes a series of new algorithms to solve such problems,and applies them to multi-modal image matching,large distortion image registration,and point set matching with high outlier rates,respectively.A large number of experimental results have verified the validity and robustness of the proposed methods.The main contents of this paper are as follows:Firstly,traditional feature matching algorithms such as scale-invariant feature transform(SIFT)are very sensitive to nonlinear radiation differences;to solve this issue,the paper deeply studies structure information for feature matching.Traditional methods usually use intensity information or gradient information for feature detection and feature description.Regardless of intensity or gradient information,they are all sensitive to the nonlinear radiation differences.Therefore,the proposed method uses a phase congruency(PC)map to replace the gray-scale image for feature detection;meanwhile,a maximum index map of log-Gabor convolution sequence is used instead of the gradient map to describe feature points.The proposed method not only largely improves the stability of feature detection operators,but also overcomes the limitation that gradient information is sensitive to the nonlinear radiation differences.Experimental results on multimodal remote sensing image datasets show that the performance of the proposed method is not only far superior to the classical feature matching method,but also more robust and flexible than current state-of-the-art multimodal image matching methods.Secondly,a robust feature matching method based on support-line voting and affine-invariant ratios is proposed.In this method,a support line voting strategy is introduced to convert the point matching problem into a line matching problem,which increases the matching constraints.A support line descriptor based on an adaptive histogram is constructed to match support lines.Affine-invariant ratios are also presented to further refine the matching results and to find out as many high-precision matching point pairs as possible.In addition,a gridwise affine transformation model is established,which can effectively reduce the ghosting phenomenon in image registration.The proposed method is not only suitable for rigid images,but also suitable for non-rigid images or images with large distortions.It has widespread application value in aerial panorama photogrammetry and oblique photogrammetry.Thirdly,a lq(0<q<1)estimator is introduced for outlier detection.To effectively optimize the nonconvex and nonsmooth function,the extended Lagrange function is used to modify the equation and the alternating direction method of multipliers(ADMM)is used to simplify the problem.This paper also gives a non-random sampling method to improve the efficiency of the algorithm.The experimental results show that the lq(0<q<1)estimator can reliably handle up to 80%of outliers,and the proposed method has two orders of magnitude higher efficiency than the classical RANSAC methods.Then,aiming at the problem that lq(0<q<1)estimator is sensitive to parameters,an improved lg(0<q<1)estimator based on Geman-McClure weight function with scale factor is proposed.This method introduces a coarse-to-fine iteratively reweighted least squares(IRLS)strategy,which greatly reduces the possibility of its convergence to the local minimum.The results show that the weighted lq estimator is very stable and robust.It can handle up to 90%of outliers,which greatly increases its practical value.In addition,the weighted lq estimator is also extended to the space resection and absolute orientation algorithms.Finally,some application examples are given to verify the correctness and practicability of the proposed methods,including:(1)mapping textures of optical images to depth maps based on optical-to-depth matching;coloring maps based on optical-to-map matching.(2)applying the support line voting and affine-invariant ratios to the urban oblique images for sparse 3D reconstruction;using the proposed gridwise affine transformation model for close-range image or distorted image registration.(2)integrating the proposed weighted lq estimator with simultaneous localization and mapping(SLAM)technique for accurate indoor 3D reconstruction;introducing weighted lq estimator for 3D point cloud registration.