Research On Algorithm Of Large Remote Sensing Image Mosaics

With the development of satellite image acquisition technology,image mosaic is a hot topic in the field of remote sensing in recent years.Satellite remote sensing data can be applied to industrial monitoring,geological survey,coverage statistics,ocean monitoring,forest fire monitoring and so on.While using the remote sensing data,it is often necessary to deal with many remote sensing images in order to obtain a larger range of ground images,so as to explore the regional macro rules.Firstly,introduce the process of remote sensing image mosaic,distortion correction and interpolation algorithm.Three kinds of mosaic techniques are explained in detail.Through the comparison and analysis of the experiment,it is found that the Pyramid mosaic could integrate the characteristic information of multiple frequencies and has a better mosaic effect.Secondly,aiming at the problem of low precision of SIFT,an improved algorithm based on gradient weighted is proposed.The multi-scale space is obtained by Gauss differential Pyramid of the image.Feature points are extracted in the scale space and matched with the feature points.Error matching points are deleted by RANSAC algorithm.Calculate and correct the transformation matrix to make the minimal error sum of all the feature points.Then,get a mosaic image by fusion.The experimental results show that SIFT algorithm based on gradient weighted could improve the matching accuracy of interesting region and the accuracy is sub pixel level.Thirdly,aiming at the problem of low precision and slow speed of SURF algorithm,an improved algorithm based on target feature is proposed.This algorithm adopts region segmentation algorithm based on dynamic programming to generate target feature image,which could retain the ROI feature of the original image.Establish the integral image,and then build its multi-scale place by Gaussian second order partial derivatives to approximate the box filter.Use Hessian matrix to choose interesting points from multi-scale place.Extract feature points from interesting points by topological structure algorithm.Calculate the Haar wavelet response of feature point to determine the direction of the feature point and generate a 128-bit binary descriptor to match feature points.The perspective transformation matrix is obtained by direct linear transformation algorithm of the least-square solutions to complete the joining together of two images.The experimental results show that the algorithm based on target feature is faster and more accurate than SURF algorithms and could be applied to several large remote sensing images fast fusion.