Robust global image registration based on a hybrid algorithm combining Fourier and spatial domain techniques

A variety of image registration techniques have been investigated for applications such as image analysis, fusion, compression, enhancement, and creating mosaics. In particular, robust registration is a key component of successful multi-frame processing aimed at super-resolution or high dynamic range imaging of space objects. Image registration techniques are broadly categorized as global (area) or feature-based, and can also be classified as being performed in either the Fourier or spatial (image) domain. Spatial domain methods are typically used for applications requiring accurate estimation of sub-pixel motion, such as multi-frame super-resolution based on dealiasing. However, these techniques often rely on the availability of a priori information (good initial guess), and are therefore limited in terms of dynamic range of the global relative motions between camera and scene. A Gaussian pyramid approach is one standard method to extend the region of convergence of spatial domain techniques. On the other hand, Fourier domain-based correlation techniques such as the log-polar FFT (L-P FFT) method provide fast and reasonably accurate estimates of global shifts, rotation, and uniform scale change, and tend to perform well over a large range of frame-to-frame motion magnitudes. In this paper we begin to explore possible hybrid algorithms for robust global registration based on combining the L-P FFT and spatial domain techniques. Initial results are presented for a hybrid algorithm using the output of the L-P FFT method as an initial guess for a spatial domain technique. In addition, we explore the potential benefits of normalized gradient correlation (NGC) in performing the coarse L-P FFT registration. The use of NGC, as opposed to phase-only correlation, has recently been explored for improving performance of the L-P FFT method in terms of robustness and dynamic range of scale factor estimates. The results presented here are based on both simulated image sequences and image sets captured in the laboratory using a CMOS machine vision camera mounted on computer-controlled micro-stepping motion stages. We have also begun to investigate optical flow algorithms for application to image enhancement in scenarios where relative motions cannot be adequately described by affine transformations. Initial results from our optical flow algorithm studies are presented based on well-controlled image sequences captured in the laboratory.