Registration Of Multi-temporal SAR Images Based On Point Features

The registration of multi-temporal remote sensing images is a process to geometrically align two or more images.It is the basis of image processing such as image fusion and change detection,and impacts the accuracy of those follow-up image processing.As an active microwave coherent imaging system,synthetic aperture radar(SAR)emits microwave,which can penetrate the cloudy and rainy atmosphere.Compared to the optical imaging system which has limitation in imaging time and weather conditions,SAR can work all-time and all-weather,which is the reason for its wide use in many domains.The thesis firstly addresses the SAR image stable feature point detection,and then studies the registration of multi-temporal SAR images method based on point features.The main work and contribution can be concluded as follows:Firstly,we analyze the methods of corner detection of SAR images.To solve the mismatch problem of feature points caused by the change of the area in the multi-temporal SAR image,a stable feature point detection method is proposed.Firstly,we propose a corner detection method,named SAR-FAST,which can be simply applied and has high detection precision and is insensitive to speckle noise of SAR image.As an improvement of the optical FAST detection algorithm,SAR-FAST maintains its high detection accuracy and high computational efficiency.To reduce the influence of speckle noise and improve the detection efficiency,the Rolling Guidance filtering algorithm is used to preprocess the image,meanwhile the detection radius and the detection window are enlarged.Finally,the stability information is added to the corner to obtain the stable feature point.Secondly,the descriptors are used to obtain the corners' local feature information,and then by matching corner,the transformation parameters between the registration images can be obtained as the rough registration result.Four existing descriptors are chosen to match the feature points:the classical SIFT descriptor,the DSP-SIFT,which has better performance than SIFT;the SAR-SIFT,an improvement of SIFT for SAR mages;and binary descriptor LATCH,which can be calculated simply and has low calculation consumption and is insensitive speckle.Besides,we apply the idea of DSP-SIFT to LATCH to obtain DSP-LATCH descriptor.The matching experimental results of the five descriptors above show that DSP-LATCH can get more matching corners and has better match performance.Finally,the SAR image precise registration base on the rough registration is applied.We use three methods for precise registration:the classical mutual information method,the point set registration,and the triangulation method.We use the Powell algorithm to find the optimal registration position of mutual information between two images.The point set registration update the transformation parameters according to the correspondence matrix between two point sets.The triangulation method uses Delaunay triangulation to obtain triangulation between the feature points,and can solve the contradiction between local geometric deformation and global geometric deformation.The experiments prove the effectiveness of the three algorithms above in the process of precise registration.