Research On Feature Extraction And Stereo Matching Algorithm Of Optical Remote Sensing Image

Stereo matching of optical remote sensing image is a process of imaging the same scene or ground object in different positions or from different perspectives by using the principle of epipolar line constraint,and obtaining the stereo information of the target by calculating the bias of the corresponding points in the image from different perspectives.Stereo matching of remote sensing image has become one of the most essential research areas in recent years as it can provide an important principle for the restoration of target height information and the stereo reconstruction of the remote sensing images.Based on the imaging principle of high-resolution optical remote sensing images,several stereo matching methods are explored in different situations in order to obtain a dense disparity map which can be directly used as stereo reconstruction.In the dissertation,the stereo imaging process of high-resolution remote sensing satellite is firstly studied,and the image is rectified by the epipolar line principle,that is,the image is transformed geometrically through the relationship of coplanar between the epipolar line and the epipolar points,so that the disparity only exists in one direction after transformation of the left and right images.The epipolar line constraint lays a foundation for stereo matching methods in several different situations as a preprocessing method.Aiming at unsupervised stereo matching without training samples,the local features and regions are combined to achieve dense matching on the premise of ensuring matching accuracy.Firstly,improved LSD algorithm is performed to extract the edge line features in the high-resolution remote sensing image,in which the edge of the target can be well shown.Then the extracted line segments are matched by a certain similarity measure function,and finally the matching based on region is guided and constrained by the line feature matching results,and using the adaptive size of matching window,we can obtain a dense disparity map that reflects the edges well.In order to further reduce the complexity of stereo matching algorithm and reduce the time of stereo matching,a small number of training samples are used and the sparse representation principle is applied to the stereo matching process.Firstly,the positive and negative sample pairs are constructed according to the truth value of disparity map,and the dictionary updating model for stereo matching is given.Then,the updated learning dictionary is used to sparsely represent the left and right two images,and the stereo matching cost volume is calculated by sparse coefficient.At last,semi-global cost aggregation and leftright consistency checking are used to improve the matching accuracy.In addition,in order to further improve the accuracy of stereo matching and robustness of the algorithm,and to meanwhile take the advantages of the large sum of data,the deep neural network method is introduced to realize the stereo matching of remote sensing images.First,deep residual network is used to extract features from images,then pyramid pooling module is used to integrate features from different scales,and finally,multiple stacked hourglass networks composed of 3D CNN are used to aggregate the costs,so as to obtain dense disparity map of stereo matching.Experimental results of disparity map show that compared with the current popular stereo matching algorithms,the optimized algorithm in the dissertation has made improvements in matching accuracy,speed and robustness correspondingly.