Research On The Invariant Feature Points Extraction And Expression Of Remote Sensing Images

Under the condition of tilt photography,Large attitude change,multi baseline photography,and unmanned small platform to earth remote sensing imaging.A large number of remote sensing images have the characteristics of serious geometric distortion,small range and high overlap.In practical applications,people often use these images which beyond the scope of human retinal,and even the high-resolution panoramic images of a wider range.But it's difficult to meet these needs using the ordinary digital camera,even if the focal length of the camera can be adjusted,the range of field is also quite limited.This requires image stitching technology to stitch these small images together to form a wide range and wide angle image.In the process of image fusion,image matching is an important step,the common image matching is based on features,which requires higher extraction of feature points for remote sensing images with multi view and illumination changes;And now the acquisition of images is almost all-weather,and it is unavoidable to be affected by the change of light,which requires the matching algorithm is much more robust to illumination changes.In view of the present situation,this paper has carried out the invariant feature extraction and expression of remote sensing images algorithm based on features.The main works are as follows:(1)The background and significance of the research are expounded,the current research status of point feature extraction and descriptor generation are introduced,analyzed and summarized,and the main problems in image registration are systematically summarized.(2)Several common algorithms of feature point extraction Harris,Shi-Tomasi,SUSAN and SIFT are introduced in this paper.The comparison experiments of efficiency,accuracy and adaptability are made,and the results are analyzed and the advantages and disadvantage of each algorithm are obtained.(3)In view of the fact that the Shi-Tomasi algorithm is not robust to the scale change of the image,the Laplace scale space is constructed to extract corner points to make it have rotation invariance.However,due to the accumulation of corner points extracted,it is not good for finding matching points later.For this reason,a method of grouping feature points is proposed,which is divided into the same group with the same structure,and then the corner points in each group are weighted according to the response value to get more accurate corner points.Experimental results show that the improved algorithm has greatly improved the scale invariance and reduced the corner clustering phenomenon.Finally,three sub-pixel corner location methods are compared,and the vector point product method is proved to be the best.(4)In view of the fact that SURF algorithm doesn't have the advantage of illumination invariance,a method based on luminance ranking is proposed to construct descriptors.Experimental results show that the improved algorithm has better performance under linear and nonlinear illumination changes.Then the improved Shi-Tomasi algorithm is described with the descriptor constructed by the luminance sorting.It is found that the stability is better than the SURF algorithm,but the number of the correct matching points is less.