Study On Structure Features Based Registration Of Multi-sensor Images

Registration of multi-sensor images is an popular and unresolved problem in the field of computer vision and videometrics. Algorithms for registration between visible, infrared and radar images are studied.Area-based method, used widely and sucessfully in registration between homo-sensor images, faces big problems in multi-sensor images registration, while the feature-based method becomes a hot research field. The feature-based method is well studied, and a novel registration idea is brought forward in chapter five.The kernel content includes four parts:1. Detection of point features and registration between two point sets are studied.A novel method based on invariable description for the registration between two point sets is put forward-RBID (Registration Based on Invariant Description). According to different registration models, two algorithms are presented, which are based on similarity invariant and affine invariant. The character of RBID is to describe the relation of the poins with invariable description firstly, then the difference between the invariable descriptions is totalized, under a hypothesis. In the process, the meaningless calculation of coordinate transformation is removed. Compared with former methods, the computational complexity is depressed one order of magnitude, and the probability of success and loss behaves well under the noise of point coordinates and the number of disturbed points.2. Detection and processing of edges are studied.A systemic method to process edges, including the definition of affinity between edges, the rules to patch edges and the rules to cut spurious edges is given, and testified by real images. At last the method is applied sucessfully in runway detection.3. Line-segment description of edges and registration between edges are studied.The method to get the optimal line-segment of edges after processd is modified, and based on the optimal description of line-segment, a method to extract the closed contour is given. And a registration method based on the line-segment support-RBLS (Registration Based on Line-segment Support) is put forward. RBLS can handle registration between two edge contours with arbitary angle rotation and scale change. Finally the false matched contours are deleted under the registration model of two images, so the registration between images is achieved.4. A novel idea for registration based on the structure support-RBSS (Registration Based on Structure Support) is presented.Traditional registration methods are based on area or feature information. The gray information is irrelative and the percent of corresponding features extracted from multi-sensor images is small, which confines further development of traditional methods. RBSS using the structure, composed of edges on reference image while area information on realtime image. The connection between them is set up through support measurement. Based on this idea, algorithms about two kinds of structure and three different ways to compute support measurement is brought forward and verified by a lot of real images. The experiment result almostly reaches the requirement of engineering application, as in several cases the correct percentage is more than 95%.Though the algorithms are performed on the registration between multi-sensor images, they also can be used on the registration between homo-sensor images.