英文摘要 | Multi-sensor remote sensing image fusion can reduce the uncertainty of single sensor images, and make each other complementary. Specifically, fusion of SAR and SPOT images, the most often used images in remote sensing, can integrate the optical and microwave spectrum of a ground object, and thus provide more comprehensive information for object analysis. Accordingly, as the key step for fusion, SAR and SPOT image registration is important. However, due to the very different imaging condition, it is very challenging for SAR and SPOT image registration. Up to now, few methods can register SAR and SPOT images very well. Meanwhile, traditional evaluation on registration methods is qualitative and subjective. Focusing on these problems, this dissertation mainly addresses the following issues: 1) Based on the quantitative descriptions of registration curves, we propose an objective assessment method on remote sensing image registration. All researches are focusing on developing a registration algorithm, whose registration curves are sharp and smooth. Instead of visual comparison, we use the ratio of half band-width to half amplitude to measure the sharpness of the peak, and curvature variation to measure the smoothness of the curves. Our evaluation is objective, quantitative, and furthermore, only two images are needed. Experiments results demonstrate its validity. 2) If the SAR image is not preprocessed, we notice that mutual information (MI) based method can register SAR and SPOT image well. Nevertheless, it is time consuming and not robust enough. Combined with feature based approach, the traditional MI based registration is improved in the following ways: ① We note that for MI based SAR image registration, different data have different extent of influence on the registration function, and the areas we call MI salient regions, contribute more significantly than the other data. Hence, to decrease the huge computational cost, with the MI salient regions, we propose an accelerated MI based SAR and SPOT image registration method. Experimental results prove the significance of MI salient regions, and they also show that the accelerated approach is satisfactory. ② To improve the robustness of MI based registration, we try to combine MI with orientation information and contrast measure respectively, and then present two robust methods: MI based registration with orientation information (MIOI) and MI with contrast measure (MIC). Based on the evaluation with registration curves, we compare MI, MIOI and MIC quantitatively. Experimental results indicate that both MIOI and MIC are more robust than traditional MI, and moreover, MIOI is more robust than MIC, with more computational cost. |
修改评论