Interactive Extraction Of Region Objects From High-Resolution Remote Sensing Images

With the development of remote sensing to high spatial, high spectral and high temporal resolution, the amount of remote sensing data is rapidly expanding. However, the data interpretation technology has not made a breakthrough.Although many scholars have done a lot of research on automatic interpretation using computer, most of the results still remain in the experimental stage. Due to the complexity of the problem, the automatic interpretation can not meet the production requirements in a short time. Hence, practical application is mainly depending on manual interpretation. This approach requires a lot of manpower and material resources. The efficiency is very low, making it a bottle neck for the interpretation of remote sensing data. Based on the above considerations, interactive extraction method is currently a viable alternative. The method can make full use of the recognition ability of human and the computing ability of the computer. Under the premise of ensuring the accuracy, it can significantly improve the production efficiency.This thesis aims for the interactive object extraction from high-resolution remote sensing images. The main contents of this thesis include:1) Exploring the natural object extraction method. In this thesis, an interactive extraction method based on fully connected conditional random field is proposed for objects with abundant spectral, texture and geometric information, such as forest, water area, bare land and farmland. The model of foreground object is estimated using the samples obtained by operator on the over-segmented results. Then, the global features of the images are described using the fully connected conditional random field. The high-dimensional Gaussian filter is used in the mean-field estimation framework Method to achieve rapid inference of the model. At last, the contour is optimized.2) Exploring the right-angle building extraction method. Most of the buildings have right-angles, and a model-driven approach is proposed. The algorithm first uses the line direction histogram to detect the main direction of the building, and then rotates the building to this direction. Then the image is divided into a series of small rectangles. The GraphCut model is established with small rectangle as the node and the rectangle shape constraint is integrated. At last, the optimal solution of the model is obtained by mincut/maxflow.3) Exploring the general building extraction method with the elevation information. The spectrum-based method is prone to error when the boundary is fuzzy and the background is similar with the foreground. If the elevation information was integrated, the robustness of the algorithm can be improved. In this thesis, LiDAR is used as the elevation source. Combined with elements of interior and exterior orientation of image, scanning line method and Z-Buffer method are applied to calculate the height value of each pixel. In reality, non-right-angle buildings are also common. Unlike natural objects, buildings often have obvious line and corner features. An interactive extraction method based on triangulation over-segmentation is proposed and the height information is integrated. The triangles are used to make full use of the corner and line features in the image. Then the triangle is used as the basic unit to build the GraphCut model and the star shape constraint is integrated. Then the model is solved and the contour is optimized.