Automatic TEM image alignment by trifocal geometry

Here we propose a novel method for automatic, markerless, feature-based alignment of TEM images suitable for electron tomography. The proposed method, termed trifocal alignment, is more accurate than the previous markerless methods. The key components developed are: (1) a reliable multi-resolution algorithm for matching feature points between images; (2) a robust, maximum-likelihood-based estimator for determining the geometry of three views-the trifocal constraint-required for validating the correctness of the matches; and (3) a robust, large-scale optimization framework to compute the alignment parameters from hundreds of thousands of feature point measurements from a few hundred images. The ability to utilize such a large number of measurements successfully compensates for point localization errors. The method was experimentally confirmed with electron tomography tilt series of biological and material sciences samples, consisting of from 40 to 150 images. The results show that, with this feature-based alignment approach, a level of accuracy comparable with fiducial marker alignment can be achieved.