Automatic target recognition in remote sensing images is of great importance in both military and civil application. Although object recognition technique has been developed for years, it remains a challenging problem due to background clutter, target pose variability, and partial occlusion. In this thesis, our work mainly focuses on recognition of several typical targets in remote sensing images, which include region of interest extraction, target detection, and fusion target recognition. The main contributions of the thesis are as following: 1. A new region of interest approach is proposed based on visual saliency. The method is built on human attention mechanism by which people quickly select regions of an image that contain salient objects without any prior information. We settle the visual saliency computation problem within the framework of scale-space theory, and the phase information provides a more promising approach to solve the problems caused by the non-negligible dark parts of target. We improve the saliency map by spatial competition algorithm, and a local adaptive thresholding segmentation algorithm is employed to segment saliency map into small regions, which has high accuracy and is insensitive to target type and size. Experimental results show that the proposed method is effective. 2. An efficient algorithm that recognizes typical objects in satellite airplane scenes is proposed. This algorithm systematically fuses low-level image saliency and high-level object semantics in a hierarchy, which somehow bridges the gap between the low- and high-level vision computations in the traditional satellite object recognizers. A coarse-to-fine strategy is used for both the airplane and oil tank recognitions. That is, we first rely on image saliency and object symmetry detection to roughly localize the positions of object targets, and then refine the localizations by shape matching. Extensive experiments show that this strategy not only significantly improves the recognition accuracy but also greatly reduces the computational complexity. 3. We propose a new fusion target recognition framework based on visual saliency. This algorithm has been used in tank detection by fusing forward looking infrared and optical images as well as oil tank detection by fusing Synthetic Aperture Radar and optical images. The proposed method offers a more flexible framework, it can be used in multi-source data and multi-class target, and can be applied at different ...
修改评论