Building Detection From High-resolution Remote Sensing Images

With the rapid development of the economy,society's demand for geospatial space is growing.As an important part of geospatial space,building contour extraction in remote sensing images has become a hot topic of research.In high-resolution remote sensing images,in addition to buildings,image information such as roads,parks,and lawns are easily confused with the edges of buildings,affecting the accuracy of building extraction.Moreover,the fact that the shape of the building is diversified and the edge of the building is occluded due to incomplete occlusion also makes it difficult to study the contour extraction of the building.In response to these problems,this paper designs a Binary Uniformity Description(BUD)that can distinguish building features from other natural scenes and non-buildings.A multi-level processing framework of “building area detection--building contour extraction” is proposed to explore a multi-level,fast and accurate new method for building detection.The binary uniformity description proposed in this paper can better adapt to the extraction of building features,and has faster computational efficiency and better classification prediction performance than the representative local feature method SURF.Experiments show that the average accuracy of multi-level contour extraction algorithm detection is more than 85%,which has a dual value of theoretical research and practical application.The main work of this article is as follows:(1)Survey commonly used building detection methods,including feature extraction algorithm,image segmentation algorithm,classification prediction algorithm.Through the investigation and analysis of various building detection methods,the extraction process of building contour can be divided into three important steps.This paper mainly introduces the relevant theoretical methods of these three important steps,and their application scenarios and the corresponding advantages and disadvantages are summarized.(2)In order to accurately extract the characteristics of buildings,this paper combines the image properties of buildings in terms of texture,similarity and color,and proposes a method for characterizing binary uniformity description.The descriptor is inspired by the ORB feature descriptor.Four sets of test values are constructed based on the feature information of texture,similarity and color,and finally combined into a binary descriptor occupying only 128 bytes of space.Compared with the traditional SURF feature descriptor,the BUD descriptor is simpler,takes less resources,and has higher computational efficiency.Especially,it is designed for building feature information,and is more suitable for feature extraction as a remote sensing image building.(3)In order to extract the contours of buildings in high-resolution remote sensing images with high efficiency,this paper designs a building contour extraction method induced by key point category attributes.The BUD descriptor is used to predict the category attribute of the key point to detect from the top down,the building area is obtained in the high space down-scale remote sensing image,and the final building contour result is extracted from the building area.By classifying the multi-layered processing twice,on one hand,the efficiency of the operation is improved,and on the other hand,the misclassification caused by the single scale is reduced.(4)In order to verify the correctness of the theory related to the method and the feasibility of the proposed scheme,a series of experiments were carried out.For the binary uniformity description,two test sets are constructed in this paper.One set of test sets is a feature sample set of two types of intersections of non-building and building in down-scale.The other set of test sets is the feature sample set of three types of intersections of non-buildings,building outlines,and interiors of scale images in original-scale.The results of predictive performance analysis using BUD descriptors for two sets of test sets show that the average correct rate is over 80%.At the original scale,the comparison between the BUD descriptor and the SURF descriptor using the accuracy-recall rate curve shows that the classification performance of the BUD descriptor is more advantageous.For the multi-level contour extraction algorithm,this paper presents the extraction results of the building contour,and carries out subjective and objective comprehensive evaluation through the intersection over union and the visual effect.The result proves that the method has high practical application value.