Automatic vs computer-assisted 3D reconstruction

Abstract

Images from light or electron microscopy of serial sections must be aligned and corrected for distortion before use in three-dimensional reconstruction. As computer rendering supplants physical modelling, algorithms are desired for automatic alignment and rectification, requiring rotation and rubber-sheeting or warping. A new image can be constructed from each digitized original with polynomial equations relating the pixel coordinates in the new image to those in the original. Interpolation between the pixel values is used to prevent aliasing of edges. Calculating the equation coefficients requires matching points in the serial sections, plus some external or a priori information about the cutting distortion. The difficulties in finding corresponding points suggest that human interaction may be needed to locate the proper matching points, while the computer carries out the more mundane but important tasks of pixel interpolation and rendering.Two approaches to automatic alignment have direct counterparts in the related problem of matching points in stereo pair images.