The Research On Multi-modal Image Registration And Fusion Based On Local Features

With the rapid development of imaging equipment,it has provided us with richer and more reliable contents for understanding things,such as multi-modal medical images(PET,CT,etc.)and multi-modal images in remote sensing.Image fusion plays an important role in displaying the description information of the same object by different imaging devices.At the same time,image registration is a necessary pre-processing for image fusion and other filed in image procession.However,existing literature in the field of image fusion,it is assumed that images have been processed through image registration.And the most widely used in the field of image registration,scale-invariant feature transformation method(SIFT)based on local features,has been improved(Dense-SIFT)for image fusion.Therefore,this paper explores the relationship and applying between multi-modal image registration and fusion algorithm based on local feature.In the aspect of feature description,the defects of the most widely used gradient amplitude in image registration in the description of multi-modal image feature points are analyzed,and another kind of gradient information(gradient occurrence)in the existing literature is studied.By analyzing the characteristics of two kinds of gradient information,although gradient amplitude and gradient occurrence describe the gradient information of image local region form different perspectives,they are complementary.Therefore,two new feature point description strategies containing two kinds of gradient information are proposed in feature description.One is to reduce the influence of gradient changes in multi-modal images by normalizing the gradient amplitudes.Another is to change the judgment conditions of gradient occurrence according to threshold of gradient amplitudes so as to obtain more abundant and accurate gradient information.Besides,RANSAC is utilized to improve the final accuracy,so that the reliability of matching can be enhanced while the accuracy is no worse than before.In addition,for the nearest neighbor distance ratio matching strategy(NNDR),which is the most commonly used in feature matching,it has been proven by experimental result that NNDR only includes the nearest feature point but no second nearest neighbor which has fundamental effect.So this paper proposes an improved method,which significantly improves the correct matching pairs of feature points.For the similarity measurement work,this paper introduces the l_p-norm distance and m_p-dissimilarity distance.Cause the m_p-dissimilarity can make up for the fault of l_p-norm that the l_p-norm distance ignores the connection between data,a similarity measurement method combining l_p-norm distance and m_p-dissimilarity distance is explored in our work.In this paper,precision and recall vs 1-precision are employed to measure the effect of image registration algorithm.Experimental results prove that the proposed method can achieve ideal results.In addition,Dense-SIFT which improved by SIFT is introduced.Dense-SIFT has been shown that feature points can be directly used for image fusion which is significant for image fusion.So this paper studies the fusing effect of Dense-SIFT in multi-modal image fusion and improved feature point descriptors are employed in image fusing.Last,our experiment has proven this new feature point descriptor is meaningful to image fusion.