SAR Image Registration Based On SAR-SIFT And Deep Learning

Image registration is the process of matching and superimposing images obtained at different times,different sensors,different viewing angles and different imaging conditions.It is the basis of other image processing techniques.The work of this paper is mainly from the feature-based registration method,and put forward several different image registration methods.Experimental analysis shows the correctness of these methods.The main work of this paper is as follows:1.Based on SAR-SIFT and MI(Manual Information)image registration method,this method improves the SAR(Synthetic Aperture Radar)image registration accuracy.Based on the SIFT method,the SAR-Harris scale space is established by using the weighted average filtering algorithm,and the feature points are extracted in the SAR-Harris scale space.The RANSAC(Random Sample Consensus))Matching method match the image,and then use the mutual information optimization strategy to optimize the image registration parameters to achieve accurate matching.2.SAR-SIFT image registration method based on MRF(Markov random field)image segmentation.The method of image segmentation based on image segmentation is a method to improve the image registration efficiency.The method divides the two images separately,and uses the significance detection to select the image segment after the image segmentation,and the image block is mapped to the original image.The SAR-SIFT method is used to process the image blocks,and the feature points are obtained for the matching process of the images.3.Using the existing feature extraction methods such as SIFT and SAR-SIFT to deal with the reference map and the part of the area to be registered obtain the correct match point pair and make the point of the neighborhood information as the point of the expression.The vector of the neighborhood is used as the training sample,and the training sample is input to the deep learning network to learn and obtain the network parameters.Then the SIFT and SAR-SIFT methods are used to extract the feature points quickly,and the extracted feature points are composed of data sets and input into the trained deep learning network.The output result of this network is the matching point.Re-screening point pairs can get the final correct match point,and then can be obtained in the final registration parameters.This paper has done a lot of simulation experiments to verify the effectiveness and accuracy of the above method.