基于互信息和随机优化的超光谱遥感图像配准

Registration of Remote Sensing Hyperspectral Imagery Using Mutual Information and Stochastic Optimization

精确的谱间配准是从超光谱遥感图像中获取光谱信息的基本前提之一。而谱间配准的主要困难在于宽的成像光谱范围使波长间隔远的图像缺乏相似性,超光谱图像本身海量的数据也限制了配准算法的复杂性。提出了一种结合互信息和随机优化技术的多分辨率配准方法。该方法采用互信息作为相似性测度,能很好的适应超光谱图像光谱特征的变化;二阶同步试探随机逼近(2SPSA)算法的应用,解决了互信息的多变量优化问题;通过一种具有平移和旋转不变性的小波分解实现算法的多分辨率形式,能明显加快算法的收敛速度并保证搜索结果的全局最优性。实验结果表明该算法适用于配准波长范围很宽的超光谱图像,并能达到子像素的配准精度。

Accurate inter-band registration is indispensable to exploit inherent spectral information in hyperspectral remote sensing imagery. Due to wide spectral range of hyperspeetral sensor, there is a general lack of similarity between the pairs of images from widely separated wavelengths, which makes registration of hyperspectral imagery difficult to carry out. On the other hand, huge volume of hyperspectral dataset also limits some complex registration algorithms to be adopted. A multiresolution registration method using mutual information combined with a stochastic optimization technique is presented to solve these problems. First, mutual information is used as the similarity metric in registration, which is robust against variations of spectral characteristic of images. Then, the second-order simultaneous perturbation stochastic approximation （2SPSA） method applied in mutual iniormation optimization algorithms, because it requires only five object function measurements at each iteration^independent of the problem dimension. Moreover, the multiresolution implementation of the registration algorithm based on a rotation- and translation-invariant wavelet and on a coarse to-fine updating strategy effectively reduces the searching region and ensure that the algorithm can reach the global maximum. In experiment some pairs of band images, which are misaligned by rotation and/or translation, are registered by our scheme, And the results show the algorithm is efficient for registering hyperspeetral imagery and yield sub-pixel accuracy.