Research On Automatic Matching Method Of Low Altitude Oblique Stereo Image With Big Inclination

Three-dimensional modeling technology has been developed rapidly with the construction of "digital earth" and "smart city" in recent years.Low-altitude oblique photogrammetry can adapt three-dimensional modeling with taking photos in different viewpoints and obtaining the top and profile information of features at the same time.Applying this technology in the process of three-dimensional modeling can save the cost and also greatly improve the efficiency of 3D modeling,has more advantages than the traditional low-altitude photogrammetry.However,due to multiple directions of shooting objects,the low-altitude oblique image contains more complex three-dimensional scene information as well as more information,data volume and data redundancy,with greater radiation distortion and geometric deformation,image objects in the occlusion of each other,the shadow of objects,parallax fracture and other phenomena are more common,which increase the difficulty of oblique image matching greatly.In this paper,we study the automatic matching method of stereo image with low altitude and big inclination,combined with the deep learning method,one of the hotspots of computer vision research currently,propose an image matching method based on deep learning.This paper uses deep learning method to classify the scenes of images,and matches the image based on the classification results.The main contents of the research work are as follows:(1)This paper proposed an image matching method based on deep learning,proposed a recursive recognition model of deep convolution neural network.Firstly,select image samples to train the convolution neural network and then input the image to be classified.The model firstly recognizes the input image initially and then distinguishes the classification results in the unit grid automatically,identifies and classifies the images accurately according to the initial classification results,and ultimately identifies and positioning the scene object accurately;(2)Based on the accurate scene classification of image,extracts image features and matches images.This paper implemented the image matching algorithm which is based on SIFT and Harris-Laplace on the programming platform VS2008,using C++.Firstly,constructing multi-scale space of image,detect Harris points of interest,adopt LoG operator to filtrate interest points to obtain feature points and use SIFT feature descriptor to describe feature points.In the period of searching match points,the search scope is performed in the same scene type as the type which the feature points belong to.So,this method can reduce the matching search amount and improve the matching accuracy and efficiency;(3)This paper used scene classification method and image matching method based on deep learning to carry out experiments,analyzed the experimental results.The experimental results show that the deep learning scene classification method to high resolution remote sensing images can obtain high precision classification results.Based on this result,the image matching algorithm of this paper can effectively increase the number of matching points and improve the matching precision and efficiency.