Study On The Method Of Heterologous Image Registration

Image registration is a process of aligning and superimposing images taken in the same scene under different conditions.Due to the wide application of different imaging methods,complementary or redundant information provided by different sensors can directly provide or supplement information that is not available in the visible spectrum,simultaneous interpreting scenes better than any single sensor image.Integrating the image information from different sensors can get more complex and detailed information in the image.The registration results can be further applied to computer vision,image processing and other important areas.However,due to the different physical characteristics of different sensors,registration of heterogeneous images is often complicated.On the one hand,the visual features in a sensor image may not appear in other sensors,so it is a difficult problem to find the features with high similarity in different images.On the other hand,due to the different gray-scale characteristics of different images,and the features of the two images are often not well preserved,so the extracted features should be stable.In addition,there are many different image registration methods in which feature descriptors have higher dimensions,and there are a lot of external points in the process of matching which cause mismatches,so the matching quality and algorithm robustness need to be improved.This paper focuses on two parts: feature extraction and feature matching.Firstly,this paper combines the phase consistency and the Improved SIFT algorithm to extract reliable and stable features in the heterologous image registration.In order to shorten the algorithm matching time,a descriptor based on layered region is proposed to reduce the dimension of feature vector.Then in order to solve the problem of poor registration quality in heterologous image registration,this paper uses NNDR method to match first,and proposes a JEED method to match again,and finally uses MS-SIFT method to optimize the matching point pair to improve the image registration quality.The main contents of this paper are as follows:(1)In order to select reliable and stable features in the process of heterologous image registration,this paper first uses homomorphic filtering to preprocess the image in the frequency domain to remove a part of noise and uninterested parts of the image.Then,in the frequency domain,the phase consistency is used to extract the edge and point features.Then feature points are extracted based on SIFT algorithm,which makes the number of feature points extracted more and the accuracy higher.In the part of descriptors,a kind of descriptors based on hierarchical region is proposed,which reduces the 128 dimension descriptors to 48 dimensions in SIFT algorithm,so as to reduce the complexity of the algorithm and the running time of the algorithm.(2)In order to solve the problems of low matching efficiency and poor matching quality in the process of heterologous image registration,this paper first uses the nearest neighbor distance ratio(NNDR)method to initially match the feature points in the matching part.Then the paper proposes a measurement JEED which combines the position error,scale error,direction error and Euclidean distance of each feature point in order to increase the number of matching point pairs.Finally,a mode search(MS-SIFT)method is used to optimize the matching point pair,so as to eliminate the mismatch and improve the matching quality.In addition,this paper compares the feasibility and superiority of this method in the field of heterologous image registration through verification experiments and comparative experiments.There are 25 figures,2 tables and 114 references in this paper.