Nonlinear Adaptive and Stochastic Optimization Using Intensity Based Metric for Automatic Satellite Image Registration Problem

Abstract

Satellite image registration problem is a critical area of research focuses on building automatic procedures to cater to the needs of compositing multi spectral image datasets, overlaying temporal data, and fusing multiple sensor data sets as the case may be. Remote sensing based change detection methods require precise registration of multi temporal images with a chosen similarity measure. Feature based methods are frequently employed to do these tasks with varying degree of accuracies depending upon the even distribution of detected feature points and the degree of image matching achieved. We here present an intensity based image registration procedure, wherein no preprocessing is involved to compute similarity metrics as they use image pair joint histogram directly in comparison to feature based methods where a preprocessing step of feature detection is exercised before the image correspondence is established. It is relevant to state that, the stochastic gradient descent optimisation technique is generally used for multi-dimensional root finding problem, employed here to find transform parameters to align pair of images. The minimisation problem considered here is a specific case for satellite image registration, where negative value of mutual information is chosen as cost function associated with image similarities. Stepsize computation in the optimisation update rule in every iteration is made adaptive based on the maximum allowed displacement per iteration. Sub pixel level accuracies are demonstrated in multidate and multi sensor remote sensing satellite image registration tasks which is required for further data analysis.