Building Areas Detection Technology From Remote Sensing Images Based On Key Parts Association Identification

With the continuous development of remote sensing data acquisition technology,the detection technology of building area in large field of view remote sensing image based on airborne,satellite and other platforms has a very urgent demand in the field of national economy,military and anti-terrorism,and related research has become a hot topic.Due to the lack of fixed significant features and sparse distribution of buildings,the development of relevant technologies for building detection is relatively lagging behind.Although the scene of remote sensing image is complex,the proportion of building area is relatively small,and the large of image is non target area.Some existing algorithms do not make full use of this feature,and do not take effective candidate area screening strategy,but directly detect the building in the image.In this way,not only the calculation is large,but also the detection effect is not ideal.Secondly,due to the distributed arrangement of building areas,and the number and distribution of single buildings in the building area are quite different,this makes the existing local descriptor and the algortithm based on strict spatial topological relation difficult to adapt to the detection of such distributed targets.In addition,the shape of suburban building is different,and the difference within the class is large,the existing detection algorithm has limited ability to resist the deformation within the class,and it is easy to cause missed detection.In view of the above problems,this paper proposes a remote sensing building area detection technology based on key psrts association identification,focusing on the three key technologies of building area key structure extraction,orientation multi structure pattern matching model,and orientation structure deformation adaptive model identification.First,this paper proposes a key structure extraction technology of remote sensing images building area.Through the analysis of a large number of remote sensing data,it is found that the building area in the remote sensing image is smaller than the background area.In view of this,this paper first proposes a preliminary screening strategy of large field building area based on the deep learning network mask RCNN,and quickly extracts the preliminary screening area from the image segmentation.Next,LSD line segment detection technology is applied to the extracted preliminary screening area,to quickly obtain all line segments in the preliminary screening area.Because of the problems of light and occlusion,most of the roofs of buildings do not show the standard rectangular shape,but there are almost parallel straight lines.In view of this,this paper proposes a key structure extraction technology based on the roof parallel structure,in order to reduce the missing detection caused by light and occlusion,and further improve the detection rate of the algorithm.Secondly,this paper constructs a pattern matching model of azimuth multi structure.In this paper,the strategy of basic element description of key structures in building area is proposed.This strategy can effectively improve the ability of anti class deformation of the algorithm,and then improve the overall detection rate of the algorithm.Based on the observation of the remote sensing image building area,it is found that although the elements in the building area are sparsely distributed,there is also a certain topological structure,based on this discovery,this paper constructs a set of topological configurations through strategies such as data augmentation of topological configurations and generation of typical topological configuration categories,all the basic element configurations in the set are from the real layout of buildings in the countryside,which is more in line with the layout law of buildings.Finally,the orientation multi structure pattern matching model based on topological configuration set is realized to better adapt to the detection of suburban building areas.Thirdly,a candidate region identification method based on adaptive model of azimuth structural deformation is designed.Because the buildings usually show the characteristics of sparse distribution,this paper designs a candidate area identification model with deformation adaptability.For each extracted topological configuration,the positional variables between the single primitive and the standard primitive are calculated,as well as the cumulative positional variables of all primitives in the configuration,the anti deformation criterion is proposed,that is,if and only if the positional variables of single primitives and the cumulative positional variables meet the threshold value,it can be determined as the building area.By judging whether the extracted topology meets the general layout rules of buildings in large field of view,this criterion can realize the accurate identification of topology and improve the detection and identification ability of the whole algorithm for distributed buildings.Finally,on the basis of the above three key technologies,this paper constructs a complete software system for the detection of building area in remote sensing image.It realizes the rapid and accurate detection and identification of suburban building areas in complex remote sensing images.In this paper,with the support of two kinds of optical remote sensing image data of general configuration and complex configuration,the key technologies and the overall framework of the algorithm studied in this paper are verified and analyzed by using the multi-level and multi angle comparative experimental means.