The problems of digital image registration and geometric correction can be subdivided into two parts: 1) determination of the warping function which will transform the geometry of the scene to the desired geometric coordinate system; and 2) processing of the digital image intensity samples, given the warping function, to produce image samples on the desired coordinate grid. The latter process, called "resampling", is a subset of the problems of image reconstruction, i.e., determination of the continuous (analog) image from a set of samples of the image, and is the subject of this paper.
This paper defines the process of image resampling in more detail in terms of general imager system models, the requirements of digital image geometric manipulation and constraints of available digital processing systems. The problem is then formulated as a constrained linear estimation problem with suitable image models and optimization criteria. The resulting reconstruction filters are compared to more heuristic approaches, such as nearest neighbor, bilinear interpolation, Lagrange interpolation, and cubic convolution (cubic and quartic spline interpolators). Finally, the various resampling techniques are compared against theoretical image models, synthetically generated imagery, and actual ERTS MSS data. Nearest neighbor, bilinear, and Lagrange interpolation resamplers are shown to give significantly poorer reconstruction accuracy than TRW Cubic Convolution and the optimal constrained linear estimator.
Date of this Version