Remote Sensing Image Registration And Building Change Detection Algorithm Based On Deep Learning

With the continuous improvement of the spatial resolution of remote sensing images,we can obtain more ground object information,but the complex ground object environment also increases the difficulty of our remote sensing algorithm research.Special,because the remote sensing image registration and image data change detection algorithm using the images taken at long intervals,terrain features in high resolution image change is more obvious,such as building demolition reconstruction,changes in vegetation,moving targets,etc.,all of these buildings for remote sensing image registration and change detection algorithm has brought the huge challenge.Aiming at the problem that existing registration algorithms have low registration accuracy in high-resolution remote sensing images with obvious changes in ground features,this paper proposes an image registration algorithm Descorner Net based on deep learning,which can extract invariable features under the situation of large changes in ground features and improve the registration accuracy.The Des Corner Net network with a Encoder-Decoder structure is a dense output,which can improve the corner positioning accuracy.Des Corner Net has two output branches: corner detection and feature description vector.Corner detection uses Focal loss and feature description vector uses hinge loss,both of which can effectively alleviate the problem of sample imbalance.In order to reduce the risk of overfitting,this paper also makes a lightweight design for the network.In order to solve the problem of high false alarm rate in corner detection,this paper applies the non-maximum suppression method(NMS)in the field of target detection to corner detection results,and effectively reduces the false alarm rate.To solve the problem of lack of high resolution remote sensing image registration data set,this paper proposes an automatic annotation method based on Bagging theory,which can make corner annotation of remote sensing image without human supervision and generate effective remote sensing image registration data set.Aiming at the problem that randomly generated homography matrix often leads to excessive distortion of images,this paper designs a scheme that can generate reasonable homography matrix,which can simulate the projection changes caused by different camera shooting angles.Through comparative experiments,the algorithm proposed in this paper is better than the traditional registration algorithms SIFT and ORB in the registration task of high-resolution remote sensing images,and the registration accuracy is even higher than the algorithms LIFT and Super Point,which are also based on deep learning.The task of building change detection in remote sensing image requires to mark out the area of building change in the image.Because the detection results of target detection and image classification are relatively rough,the pixel-level segmentation method is adopted in this paper.According to the existing evaluation criterion of segmentation,Intersection over Union(IOU)can only measure the similarity of segmentation regions,but can not measure the segmentation accuracy of building contour.In this paper,a precision criterion of building contour segmentation is proposed to measure the accuracy of change detection algorithm for building contour segmentation.At the same time,aiming at the problem of inaccurate segmentation of building contour,this paper proposes a Rebuild Block module to effectively improve the segmentation accuracy of building contour.Aiming at the problem that the algorithm is not complete in segmenting the changing region of large buildings while the segmentation of small buildings is missing,this paper uses the pooling structure of spatial pyramid to improve the similarity of the segmented region.Aiming at the problem that the building change detection model misjudges the new road as the reconstructed building,this paper uses the attention mechanism to make the algorithm model obtain the global information of the image,so as to avoid the misjudgment.Finally,compared with several deep learning-based change detection methods in the experiment,the effectiveness of the Rebuild-CD algorithm in the high-resolution remote sensing image building change detection task is proved.