Research And Implementation Of Building Change Detection Method In UAV Remote Sensing Images Based On Image Registration And Semantic Segmentation

As the main place of human activity,buildings are one of the most likely to change and most in need of updating objects in the urban geographic information database.Only by timely and accurately grasping the distribution and changes of buildings in an area can we scientifically and effectively plan,manage,and supervise the area.In recent years,the progress of UAV remote sensing technology and the rise of artificial intelligence have provided reliable technical support and guarantee for the use of UAV remote sensing images to detect changes in urban buildings.However,for high-resolution UAV images,traditional change detection algorithms are seriously disturbed by external factors,while deep learning methods lack appropriate corresponding training samples and have strict requirements for model structure and training parameters.Based on the above considerations,this thesis combines traditional change detection technology with semantic segmentation technology to conduct research on building change detection methods in UAV remote sensing images.The main work content of this thesis is as follows:(1)Aiming at the high requirements for image registration accuracy in the change detection task of UAV remote sensing images and the poor stability of traditional registration algorithms,this thesis proposes a UAV remote sensing image registration algorithm based on improved SIFT and multiple constraints.Firstly,the number of effective feature points is increased by fusing SIFT and Scharr-ORB algorithms,and then mismatches is eliminated by using the strategy of fusing multiple coarse matching and fine matching,so as to complete the high-precision registration of the image.The experimental results show that the overall accuracy of the algorithm and its performance in the extraction,coarse matching and fine matching stages of feature points are improved.At the same time,experiments also verify that the improvement of image registration accuracy can effectively improve the detection accuracy of subsequent change detection.(2)For the rich detailed information of UAV remote sensing images in complex urban scenes,this thesis proposes an improved UPer Net network model.By using Swin Transformer as the backbone network of the model and using the comprehensive loss function in the training process,the semantic information at different levels is effectively used to enhance the segmentation ability of buildings.Aiming at the problem of training samples,this thesis manually labels and produces 10800 UAV remote sensing image building samples for model training.The experimental results show that the model has achieved different degrees of improvement in various accuracy indicators,and can be well applied to subsequent building change detection methods.(3)Combining the advantages of traditional change detection algorithms and semantic segmentation methods,this thesis proposes a method for noise removal in non-building areas.The building segmentation results of the improved UPer Net network model are used to remove a large number of irrelevant noises located in nonbuilding areas in the pre-detection results of IR-MAD algorithm,and then the postprocessing methods are used to further optimize the detection results.The experimental results show that this method can greatly reduce the false detection rate of change detection,improve the accuracy of detection,and obtain relatively complete detection results.(4)Based on the analysis of practical application requirements,the building change detection method studied in this thesis is packaged and connected in series,and a set of process-oriented,visualized and automated UAV remote sensing image building change detection system is designed and implemented,which can basically meet the actual needs of application scenarios such as daily inspection of buildings.There are 45 figures,11 tables,and 91 references in this thesis.