Image Registration with Global and Local Luminance Alignment
(Leo J. Jia)

We develop a unified and robust approach that uses tensor voting to address the problem of global and local intensity alignment for image registration. Our method globally and locally adjusts intensities of two overlapping images, without assuming a complex camera model or any simplified assumptions other than the monotonic constraint. Our iterative scheme converges quickly, thanks to the robust estimation of the replacement functions by tensor voting. Compared with other techniques, tensor voting is novel as it provides a fundamentally different approach to performing intensity alignment and effective since an optimal function under the monotonic constraint is obtained. In the whole process, only a rough focal length for the first reference image is required. We have applied our voting methodology to a variety of applications: image mosaicking of static scenes, image mosaicking in the presence of occluding objects, intensity compensation, and correcting saturated images.