A Methodology For Semi-automatic Building Extraction From Very High Resolution Remote Sensing Imagery

Building is one of the most important man-made objects in very high resolution(VHR) remote sensing imagery, which is closely related to human's life. The efficiency of traditional building extraction method which has higher requirements for operators is too low. Although automatic building extraction method improves the efficiency of building extraction in a certain way, it can not meet the need in common use. With few manual interactions, semi-automatic building extraction method can greatly improve the accuracy and efficiency of building extraction both. So developing semi-automatic building extraction method is feasible. In order to extract building with a better effect, we conduct a series of valuable research and exploration on semi-automatic building extraction method based on the previous works. And the research and exploration are as bellow.First, we introduce the commonly used image processing methods in VHR imagery. Bilateral filtering is an effective method to preserve building's boundary and smooth building's internal details and background details when processing the VHR imagery. Then Mathematical morphology and edge fitting methods are used to reduce the edge burrs and get building result with a regular shape.Then, we propose a semi-automatic building extraction algorithm, which is based on superpixel segmentation, region merge and segmentation based on graphcut, and we apply this algorithm to extract rectangular buildings. Superpixel segmentation is to segment the VHR imagery to several superpixels, which are treated as basic units in the subsequent processing to improve the efficiency of computing. Combining the human interaction, we get the initial foreground area by using the maximum similarity region merging method. Then we get the foreground and background's priori information with harris corner distance salient map. At last, graphcut method is applied to segment the VHR imagery and get the building extraction result, which is using Gaussian mixture model to construct the energy function and max flow/min cut algorithm to minimize the energy function.Then, on the basis of the general building extraction algorithm, the optimization and improvement are carried out to extract buildings with complex shapes. With a certain amount of human interaction, we apply the saliency detecting method to obtain more accurate priori information of building objects. At last, the graphcut method based on geodesic multi star convexity constraint are used to solve the problem of Shrinking bias and color overlap of foreground and background, which greatly improves the efficiency of buildings extraction with complex shapes.Finally, we analysis and confirm that the efficiency of SLIC superpixel function is the lowest in the whole building extraction process. We apply GPU devices to superpixel process in a parallel way, which highly improves the efficiency of building extraction method. We also apply the improved semi-automatic building extraction method to the real production system. The result proves that our semi-automatic building extraction method gets good effect.