The Algorithm Design Of Remote Sensing Image Stitching Based On Convolutional Neural Network

The technology of remote sensing image stitching has a wide range of applications,ranging from military reconnaissance,surface analysis,resource exploration,to personal mobile phone positioning,traffic travel,querying road conditions,and so on.All fields depend on the accuracy of remote sensing images.Life is closely related.Because the viewing angle and resolution of the remote sensing images acquired by a single imaging device are limited,it is often necessary to shoot multiple remote sensing images containing overlapping areas for stitching in order to obtain images with a wider field of view and higher resolution.The starting point of this paper is: remote sensing satellites have their specific orbits,and drones will also be restricted by their own flight capabilities and legal policies.Therefore,these high-altitude imaging devices will have image deletions in certain areas when acquiring remote sensing images..To solve this problem,for areas where remote sensing images cannot be obtained,images can be obtained from other satellites or drones that can cover this area.Finally,the splicing of remote sensing images from different imaging equipment is completed.However,this approach can also cause problems.The remote sensing images from different devices have large viewing angle differences when they are taken,so the same thing will appear largescale deformation in different images.This situation will lead to difficulties in remote sensing image registration.This research aims to solve this problem.For the preprocessing stage of remote sensing images,this experiment studies the image enhancement technology,and filters the image mosaic process with noise as a series of interference factors.For remote sensing images of the same area on the ground but from satellites in different spatial locations,this paper proposes a unified scale solution.The conversion method between the coordinate system of the remote sensing image and the rectangular coordinate system,and the method of determining the overlapping area of the remote sensing image by the coordinate position are studied.For the image registration and geometric transformation stages,this paper analyzes the mainstream algorithms currently applied to remote sensing image registration,and proposes its own methods to optimize;this experiment also studies the hot deep learning principles in recent years and combines them Applied in the field of remote sensing image splicing,and proposed a model of my own design.The training model is used to realize image splicing,and experiments have verified that the method proposed in this research has a higher accuracy in remote sensing image registration.The current registration method mainly used in remote sensing image splicing technology is based on the classic algorithm of SIFT and a series of variants.The SIFT algorithm first extracts SIFT features that are irrelevant to scale scaling,rotation,and brightness changes from the image,and then compares the key points in the two images,and determines whether they are matched points by the Euclidean distance of the key point feature vectors.However,this method also has shortcomings.When the appearance of objects in the image is greatly deformed,the correct stitching effect is often not obtained.This research mainly studies the three stages of feature extraction,feature matching,and geometric transformation in remote sensing image splicing technology,and analyzes the principles of each step.Based on the principle design,a CNN model that can be trained end-to-end is proposed.Regarding the choice of the number of layers of the model structure,this study refers to the VGG-16 structure and conducts comparative experiments on different structural models from 6 to 13 layers.The thesis analyzes from the three perspectives of training complexity,image registration timeconsuming,and registration accuracy,and finally determines a 13-layers structure as the choice.This experiment completed the training of the CNN model on the processed DIOR data set.In the test of the accuracy of remote sensing image registration,MSE,NRMSE,SSIM,MI are used as evaluation indicators,and statistics are performed on different remote sensing image data sets through the CNN model method and the traditional SIFT algorithm and the SURF algorithm.The image registration capability is evaluated.Experimental results prove that the CNN model has a better capability of remote sensing image stitching.At last,the remote sensing image stitching effect achieved by the method proposed in this paper is tested in the actual remote sensing image stitching application scenario.