A Research On Extraction Of Rural Buildings From UAV Images Based On Deep Learning

In the current transformation phase of urban and rural planning,rural areas are one of the key targets.The maps of rural buildings are important auxiliary decision-making data.In the past,most of the drawing methods were manual on-site measurement or manual processing using remote sensing images,which required a lot of manpower and material resources and was inefficient.Therefore,it is very important to find a way to quickly and accurately extract rural buildings.The rapid development of aerial remote sensing platforms for multi-rotor UAVs allows researchers to quickly and flexibly obtain high-resolution images in a small area at a relatively low cost.However,it is a difficult problem to automatically and accurately extract rural buildings from drone images.Traditional remote sensing image processing and building extraction algorithms have gradually entered the development bottleneck.With the improvement of image resolution,the details of buildings in the image are more abundant,but the same background information is also increasing.How to extract effective building information from complex scene information is an important research direction of remote sensing technology today.How to extract buildings from images quickly,intelligently,and automatically is an important problem that the remote sensing field research solves.In recent years,convolutional neural networks have achieved rapid development,and their applications in the field of remote sensing have achieved remarkable results.However,according to the current research situation,the extraction of rural buildings from UAV images still has shortcomings such as unsuitability of the network and poor segmentation accuracy.In this paper,the principle of regional convolutional neural network R-CNN is studied,and on the Mask R-CNN deep learning method based on this algorithm,the study of rural building instance segmentation is carried out.By improving the feature extraction network,it can be used in the absence of the human-machine aerial image data is centralized to realize the extraction of rural buildings.The main work of the thesis is as follows:(1)Introduced the basic structure and basic principles of convolutional neural networks,and constructed a drone image data set in rural areas.The principle of R-CNN is studied and tested on the data set to ensure the feasibility of the deep learning method to classify buildings from the background.Research the principle of Mask R-CNN,analyze its network structure,extract the network Res Net34/50/101 for different features,and conduct training and testing on the data set constructed in this article.(2)On the basis of the existing deep learning methods,combined with the characteristics of rural buildings in the drone image that the roof debris is piled and blocked,the building materials are different,and the background is complex,the feature extraction the network is improved.The Res2 Net network with multi-scale residual units replaces the original feature extraction network of Mask R-CNN,and it is trained and tested.It is found that the improved model of Res2Net50/101 has an accuracy of 93.87% and 92.60%,which are high The original model of Res Net34/50/101 was developed to prove that the feature extraction network of multi-scale residual unit is an effective improvement.(3)Modify the output branch of the model to output the outline of each building and vectorize the output result.The outline of the building is simplified by the classic Douglas-Puck algorithm in the GIS field.Get the revised building outline.