Study On Building Extraction Method Based On Image Segmentation And Deep Learning Network

As one of the important technology of acquiring spatial geographic information,UAV low-altitude remote sensing technology's fast and accurate acquiring of building information shows great significance to basically construct and update the geographic information.Although a lot of researches and great progress has been made in this area,the automatic extraction of buildings using a single data type(2D remote sensing image or light detection and ranging 3D point cloud)is still not enough to accurately describe the building outline for automatic mapping.This paper proposes an effective method for extracting buildings from drone images by combining super pixel segmentation and semantic recognition.This method uses an improved simple linear iterative clustering(SLIC)algorithm and Multiscale Siamese Convolutional Network(MSCN).The framework of building extraction is jointly constructed,and experiments show that the method has a good demonstration effect on largescale application of building extraction in UAV images.The main research contents and results of this study on building extraction methods are as follows:(1)In terms of image segmentation,an improved image segmentation algorithm based on the SLIC algorithm(6D-SLIC)is proposed.This method adds extra height information for the super pixel segmentation model.First,the pixels are clustered according to the color similarity of the pixels to generate super pixels in a two-dimensional image plane space.Then,the height information of the 3D point cloud is used to cluster the pixels.Class,which effectively solves the problem of over-segmentation of color similar regions and improves the segmentation accuracy of edge fuzzy regions.It is suitable for boundary detection of buildings with similar radiation characteristics and buildings in image backgrounds.(2)In terms of vegetation detection,a gamma-transformed green leaf index(GGLI)is proposed to detect super pixels containing vegetation for further processing.GGLI extracts vegetation by enhancing vegetation intensity and using adaptive thresholds,which improves the Robustness and efficiency of building detection.(3)In terms of building detection,the proposed MSCN(including feature learning networks and binary decision networks)is used to automatically learn multi-scale hierarchical feature representations and detect building objects under various complex backgrounds.The MSCN building detection model can extract non-linear high-level semantic features that are not easily affected by image gray changes,and has higher robustness.This research also proposes to use Douglas-Peucker algorithm and the regular algorithm of iterative optimization.