Research On Image Registration Algorithm Based On Features

Image registration is a fundamental problem in image process. Image registration iswidely applied in many aspects such as remote sensed image interpretation, target recognition,image fusion, robot vision, medical image process etc.Concepts and methods of image registration theory are introduced at first. Feature basedimage registration method is mainly studied in this paper. Those features contain points, edgesand regions. SIFT is the most valuable and widely used method in feature-points-basedimage registration. SIFT descriptor has invariant characters such as rotation, illumination andscale invariance. But the computation speed of SIFT is very slow, and many unstable featurepoints are always extracted. By further analyzing we found that the process of creatingdescriptors occupied almost50%of all computing time of SIFT. We proposed using Hessianmatrix to remove the unstable features, which will decrease number of features formingdescriptors. As a result, the overall computing time will decrease and the efficiency of SIFTwill be increased. Experimental result demonstrated the effectiveness of the proposedalgorithm.For aerial images image registration process generally failed to work for rich in texture.Feature points are difficult to extract. We proposed to solve this problem by integrating ShapeContext and the modified SIFT. First, feature points of reference image are extracted with theimproved SIFT algorithm, and the edges of the test image are extracted by Canny operator.Secondly, feature points are formed by Shape Context descriptors. The descriptors arematched between two images with a cost function. At last, image registration was completedby minimizing the cost function. Experimental result demonstrated that aerial images aresuccessfully registered by the proposed method.