Research On Remote Sensing Image Matching Technology Based On Convolutional Neural Network

Satellite remote sensing detection technology has become one of the important means of ground detection in modern countries,and remote sensing image registration is a key step in detected image processing.Registration is the process of transforming two pictures of different scenes into the same coordinate system through the best coordinate transformation model.Most of the features extracted by traditional registration algorithms are based on artificially defined features,and these features only have low-level semantic information.In recent years,researches on convolutional neural networks have attracted increasing attention.Convolutional neural networks can extract high-level semantic features of images and perform well in a variety of image processing fields.In this paper,the ability of feature extraction by convolutional neural networks is applied to the field of remote sensing image registration,it mainly studies the registration technology based on convolutional neural network.The main research contents of this paper are as follows:(1)We investigates the significance and current research status of remote sensing image registration,and studies the principles of traditional SIFT and SURF registration algorithms and their advantages and disadvantages.The basic structures and working methods of convolutional neural networks are introduced,then this paper studies several new and commonly used convolutional neural networks.(2)In order to solve the problem that the feature extracted by the traditional algorithm is low-level and does not make good use of the deep information of the image,an algorithm based on the feature of convolutional neural network is proposed.This paper improves the original Dense Net algorithm and deepens the large-scale feature extraction module,then proposes a registration algorithm based on the improved Dense Net network.Image features are extracted through the improved Dense Net network,and feature points are coarsely matched and filtered at the deep feature layer.The characteristics of maximum pooling are used for reverse fine matching to find the precise position of the matching feature points in the input image to complete the image registration.Experiments show that the algorithm takes into account both registration accuracy and registration time,and has better overall registration performance.It also illustrates the feasibility of convolutional neural networks in the field of image registration.(3)The traditional SAR-SIFT algorithm gets the feature descriptors based on the gray value changes of the feature point neighborhood.These feature descriptors have a low accuracy rate for matching.So a fusion registration algorithm based on SAR-SIFT and Siamese network is proposed.After using the feature point extraction function of SAR-SIFT,the feature blocks of the reference image and the image to be registered are matched by the Siamese network whose main structure is Dense Net in Chapter 3,and the image registration is completed by using the coordinate points of the matched feature points correctly.Experiments show that the algorithm improves the number of correct matching points and has better registration accuracy than traditional SAR-SIFT.