Remote Sensing Image Registration Method Based On Robust Self-Learning Descriptor And Local Neighborhood Information

Image registration is a process of matching the image of the same scene with different time,different viewpoints,and different sensors by using image information to transform coor-dinate.It is the basis of many image processing problems such as image Mosaic,image fusion,transform detection,3d terrain reconstruction and so on and it has wide application and research valueWith the rapid development of computer science and technology,image registration tech-nology is also developing.According to the different characteristics of images,registration methods are divided into three categories:registration method based on the gray level infor-mation,registration method based on the transform domain and registration method based on the feature based methods.Due to the updating and iteration of sensors and other factors,the imaging mechanism of remote sensing image changes,making the spatial resolution of image constantly improve.In addition,the complex noise of remote sensing image also makes the image registration technology face many challenges.Based on the research of the existing remote sensing image registration technology and the scale-invariant feature trans-form(SIFT)algorithm,the following three remote sensing image registration methods are proposed:(1)Aiming at the problem that the significant nonlinear gray changes of remote sensing im-age lead to wrong matching point pairs or less matching point pairs,a remote sensing image registration method based on local similarity constraint and local regional structure is pro-posed.In this method,SIFT algorithm is used to extract and match feature points,and local similarity constraint strategy is used to carry out secondary matching of matching point pairs to get high-precision matching point pairs.The experimental results show that using the lo-cal similarity constraint is robust to remote sensing images with large nonlinear differences Then the feature points extracted from SIFT are rematched with the local region structure The experimental results show that the local structure information can effectively increase the number of matched feature points(2)Aiming at the problem that only using gradient extraction of image features lead to ab-sence of neighborhood information,a feature matching method based on robust self-learning image descriptor is proposed.Combining with the idea of genetic programming,the method makes full use of the neighborhood information of feature points,extracts feature vectors of feature points,and gets the image descriptor with high fitness value by optimizing.The trained image descriptors are applied to the test data set and the matching point pairs with larger distance of feature vectors are deleted.Experimental results show that the proposed method for feature extraction and feature matching are effective(3)Aiming at the problem that the feature descriptors are interfered by a lot of noise in remote sensing images,and the feature points are mismatched,a feature extraction method based on improved first-order statistics is proposed.Based on the self-learning image descriptor,the terminal set is changed to 25th percentile,75th percentile,mid and stdev.Because the noise contained in the remote sensing image changes the pixel value which should change uniformly and continuously,the improved first-order statistics can reduce the influence of noise on the feature.Experimental results show that the algorithm can effectively improve the performance of remote sensing image registration.