Detecting Building Roposals Based On Polygonal Superpixels

Building extraction based on remote sensing images is regarded as an important research topic of photogrammetry and remote sensing,however,the traditional algorithms consumes a lot of manpower and time.The research results can provide efficient means for eliminating redundant information to reduce the computation space and time overhead of extraction algorithm,in which new idea of timely acquiring and updating geographic information has been put forward.In this thesis,a novel building detection method based on polygonal superpixels is proposed according to the latest research results of polygonal superpixels and proposal detection in the field of computer vision.The main achievements and contents are as follows:First of all,unlike the previous building detection methods,this thesis introduces the concept of object proposal detection in the field of computer vision applied to the detection of buildings in remote sensing images.The method aims to pick out some proposals efficiently,which highly likely include buildings,that can reduce the area where buildings exists and provide some good basic data for subsequent complex building identification algorithms.Secondly,different from the existing proposal detection methods based on the superpixel merging strategy,this thesis first divides the image region into a large number of small convex polygons according to the geometry of buildings,and then merges the small polygons to obtain a few large polygons.Each large polygon is used as a candidate building area,which means proposal building.Then,constructing the graph of image to represent the polygons and their adjacency relationship.For purpose of describing the similarity of neighboring polygons,this thesis sets the weights of the corresponding edges between neighboring polygons according to the features such as color and texture.Minimum spanning tree algorithm is used to merge similar nodes and the geometric features are introduced into the growth termination condition.Finally,this thesis has implemented the above method and has carried out the experiments and evaluations for the weight variables of our method,in addition,compared with the state of the art.The result turns out that our algorithm performs a better detection rates in building detection,and average processing speed of a single image with a size of 1000x750 is 0.987 second.The best proposals of an image can fit the ground truth of building objects tightly.