Research And Application Of Feature Learning-based Remote Sensing Image Registration Method

Remote sensing image registration is crucial to the accurate extraction and analysis of land cover information.Accurate image registration is especially important for re mote sensing data that contains different types of ground objects and multi-platform sensors.Traditional registration methods mainly rely on the consistency of image features and the precise location of homonymous points and have poor performance in complex surface areas..The development of deep learning provides new solutions for image registration,but the existing deep learning-based registration methods have low accuracy and cannot be applied to remote sensing images with large displacements.To address these problems,this thesis proposes improvement schemes from the aspects of edge point removal,mismatch filtering and deep convolutional features,and conducts experimental verification with UAV images and satellite images.The main contents are as follows:(1)A UAV image registration algorithm based on edge feature point removal and mismatch filtering is proposed.Edge features are extracted by Canny algorithm and noise in the edges is removed by wavelet transform.Edge points in the feature points are filtered by edge images to obtain a stable and uniformly distributed feature point set.Adaptive threshold calculation and PROSAC mismatch filtering are combined to achieve filtering of feature matching sets without manual setting of thresholds.(2)A satellite image registration algorithm based on deep convolutional features is proposed.The idea is to first extract two feature maps from the warp image and the reference image by a siamese network,and then use a deep feature extraction network to predict the displacement parameters of the four corners of the warp image block relative to the reference image,skipping the steps of feature matching and description.Attention mechanism and residual structure are introduced in the network to enhance the feature extraction ability of the network.In addition,a new remote sensing registration training data generation method is proposed to make the generated registration images more realistic.(3)An online remote sensing resource sharing and registration processing system is built using FLASK framework and Alibaba Cloud OSS for image data storage.It supports users to upload images,realize resource sharing and resource attention functions.And the above two registration algorithms are deployed to the background server to realize online image registration function.The proposed methods are tested on UAV images and Sentinel-2 satellite images,and compared with SIFT and other deep learning registration methods.The experimental results show that the proposed methods are superior to all comparison methods in terms of accuracy,robustness and efficiency,providing new ideas and techniques for remote sensing image registration.