Fully automatic stitching and distortion correction of transmission electron microscope images

In electron microscopy, a large field of view is commonly captured by taking several images of a sample region and then by stitching these images together. Non-linear lens distortions induced by the electromagnetic lenses of the microscope render a seamless stitching with linear transformations impossible. This problem is aggravated by large CCD cameras, as they are commonly in use nowadays. We propose a new calibration method based on ridge regression that compensates non-linear lens distortions, while ensuring that the geometry of the image is preserved. Our method estimates the distortion correction from overlapping image areas using automatically extracted correspondence points. Therefore, the estimation of the correction transform does not require any special calibration samples. We evaluate our method on simulated ground truth data as well as on real electron microscopy data. Our experiments demonstrate that the lens calibration robustly corrects large distortions with an average stitching error exceeding 10 pixels to sub-pixel accuracy within two iteration steps.