深空探测多视光学影像超分辨率重建方法研究

In order to deeply dig up the information of deep space exploration images, super-resolution technology is applied to multi-view images which contain subpixel information to improve the quality of the deep space exploration data. This permits the images to play a greater role in subsequent scientific researches and engineering applications, and also can provide reference in quality improvement of deep space exploration images in the future. However, the deformation of multi-view images greatly reflects the surface fluctuation of the target celestial body, so it can’t be modeled by common global and block translation or affine transformation. And one of the degradation factors of deep space exploration image is the error of IMC (Image Motion Compensation), which leading to the limitation of Gaussian blur kernel. In addition, deep space exploration images are mostly textureless and monotonous. This makes high accuracy motion estimation and kernel estimation of super-resolution more difficult. For this reason, the major research contents of this research are motion estimation and kernel estimation in super-resolution. And this is different from most super-resolution researches that focus on reconstruction algorithms. In this research, flexible non-parametric motion estimation is employed to align multi-view images with high sub-pixel accuracy. And based on the characteristics of deep space exploration images, a fuzzy boundary detection based arbitrary blur kernel estimation method is studied. By employing advanced reconstruction method, the reconstructed high resolution image, motion field and blur kernel are updated alternately, and finally an image with higher resolution is generated. The high resolution image is evaluated with objective indices, and also used for target identification and extraction.

为提高现有深空探测光学影像质量，深入发掘影像信息，使其在后续科学研究和工程应用中发挥更大作用，并为未来深空探测影像质量改善提供参考，课题以多视影像为数据研究深空探测影像超分辨率重建方法。由于深空探测多视角观测影像之间的形变主要受探测目标体表面地形起伏影响，以及影像降晰主要受像移补偿误差影响，难以用目前常用的全局和局部参数运动模型以及高斯形式的降晰核描述，因此课题主要针对超分辨率重建中的运动估计和降晰核估计展开研究。研究思路是：采用不受模型限制、灵活的非参数运动估计方法，实现多视影像的高精度亚像素运动估计；针对深空探测影像纹理匮乏、边缘模糊的特点，研究基于模糊边缘提取的非参数降晰核估计方法；应用超分辨率重建算法的先进研究成果，通过重建影像、运动估计、降晰核估计的交替更新，实现深空探测影像的超分辨率重建。最后评估重建影像质量改善情况及其对目标识别和提取结果的影响。

本基金的主要研究内容为深空探测多视光学影像超分辨率重建方法，即利用多视角立体相机同时获取的同一区域的不同角度影像，或者相对短时间内探测器通过多次覆盖获取的同一区域不完全相同的影像数据之间的互补信息，提高影像的空间分辨率。课题通过对“非参数高精度亚像素运动估计”“非参数降晰核估计”等关键技术的研究，实现了课题的预期目标，通过超分辨率重建，提高了影像的空间分辨率，改善了影像的视觉效果，增强了影像中目标的可识别性。本课题的研究成果使得深空探测光学影像在后续行星地形地貌研究、行星表面过程研究、行星表面特征提取等科学应用，以及着陆点选取、探测规划制订、行驶路线导航等工程应用中可发挥更好的应用效果。