Automatic Detection Of Buildings In Remote Sensing Images Based On Multi-scale Features

With the rapid development of society,the level of computer hardware and software technology has been greatly improved.With the vigorous development of the field of computer vision,traditional remote sensing and photogrammetry technology have also entered a new era of development with the improvement of computer technology.This discipline is also increasingly playing a role in all aspects of production and life.In urban areas,buildings account very much of all feature types,which is one of the most important feature types.Building detection has a wide range of applications in the fields of urban planning,urban geographic information system platform construction,map updating,illegal building monitoring,smart city construction,and military reconnaissance.It's easy to visually interpret buildings from remote sensing images,but it is very difficult to achieve automatic computer detection.Computer vision is mainly aimed at solving the problem of natural scenes,and the field of remote sensing is mainly to study high-resolution remote sensing images with more complex background information.Due to the huge amount of remote sensing images,the semantic information is very rich,traditional building detection and extraction The algorithm is only applicable to a specific type of building,and it can't perform satisfactorily on a large range of data sets,nor can it meet the requirements of today's rapid,intelligent and automated extraction.Therefore,this paper proposes an algorithm that uses deep learning multi-cascade target detection framework for multi-scale building detection,and uses edge detection ideas to guide the segmentation of building instances to obtain the accurate contour of the building.On the building dataset made by ourselves,the residual network Res Net-50 is used as the backbone feature extraction network,and the multi-cascade detection convolutional neural network Cascade R-CNN is used to learn and train the target features to obtain the building target detection.The template parameters are then tested on the test set to obtain the preliminary detection results,that is,the rectangular labeling frame containing the building and the confidence score of each frame.Among them,a feature pyramid network module is added to the reference network to build a feature pyramid and merge multiple layers of feature information,so that the output features can better represent the information of various dimensions of the output image,thereby realizing the detection of buildings of different scales and improving The robustness of the algorithm.In addition,drawing on Mask R-CNN and edge detection ideas,a mask branch is added after the ROI Align module,and an edge protocol header is added on this basis.By using traditional edge detection operators in both prediction masks and ground truth values,Correct the edges of the mask to make it more fit to the edge of the actual building.This module can improve the accuracy of segmentation edges of building instances without basically increasing the size of the model.By using the training model for building detection and extraction on the test set,the classification,precise edges and confidence of the building in each image are obtained.In this paper,a comparison experiment of control variables is carried out with a variety of target detection and instance segmentation methods,and the accuracy of detection and extraction results is evaluated by multiple evaluation indicators.The algorithm in this paper can achieve end-to-end building extraction.The highresolution remote sensing image data set produced in this paper is used for test experiments and compared with many classic algorithms.Experimental results prove that the algorithm in this paper has better detection effect and extraction accuracy than other methods,and has certain advantages and practicality.