Computed alignment of dissimilar images for three-dimensional reconstructions

Three-dimensional reconstructions from serial section images require the accurate registration of those images. Image correlation is the most powerful computed alignment method and its performance on identical images, or parts thereof, has been thoroughly studied. Correlation alignments of complex, dissimilar images can fail, however, with a likelihood proportional to the magnitude of the differences. We report that alignments can be computed more reliably and more accurately (higher-valued correlation coefficients) by the combined use of lowpass-filtered product transforms (from which the correlation functions are formed), autocorrelation correction of rotational misalignment, and covariance correction of translation misalignment. A simple rule is proposed for the lowpass filter cutoff radius depending on measures of the images' differences. These methods are demonstrated with a reconstruction of a capillary loop in the median eminence of the hypothalamus.