Landmark correspondences can be used for various tasks in image processing such as image alignment, reconstruction of panoramic photographs, object recognition and simultaneous localization and mapping for mobile robots. The computer vision community knows several techniques for extracting and pairwise associating such landmarks using distinctive invariant local image features. Two very successful methods are the Scale Invariant Feature Transform (SIFT)1 and Multi-Scale Oriented Patches (MOPS).2
We implemented these methods in the Java programming language3 for seamless use in ImageJ.4 We use it for fully automatic registration of gigantic serial section Transmission Electron Microscopy (TEM) mosaics. Using automatically detected landmark correspondences, the registration of large image mosaics simplifies to globally minimizing the displacement of corresponding points.
We present here an introduction to automatic landmark correspondence detection and demonstrate our implementation for ImageJ. We demonstrate the application of the plug-in on diverse image data.

In order to study anatomy of organisms with high-resolution there is an increasing demand to image large specimen in three dimensions (3D). Confocal microscopy is able to produce high-resolution 3D images, but these are limited by its relatively small field of view compared to the size of large biological specimens. To overcome this drawback, motorized stages moving the sample are used to create a tiled scan of the whole specimen. The physical coordinates provided by the microscope stage are not precise enough to allow reconstruction (”Stitching”) of the whole image from individual image stacks.
We developed an algorithm, as well as an ImageJ plug-in, based on the Fourier Shift Theorem that computes all possible translations (x, y, z) between two 3D images at once, yielding the best overlap in terms of the cross correlation measure. Apart from the obvious gain in computation time it has the advantage that it cannot be trapped in local minima as it simply computes all possible solutions. Computing the overlap between two adjacent image stacks is fast (12 seconds for two 512x512x89 images on a Intel ® Core2Duo with 2.2GHz) making it suitable for real time use, i.e. computing the output image during acquisition of the individual image stacks.
To compensate the possible shading- and brightness differences we apply a smooth linear intensity transition between the overlapping stacks. Additionally we extended the to generic 3D registration using gradient based rotation detection on top of the phase correlation method. We demonstrate the performance of our 3D stitching plug-in on several tiled confocal images and show an example of its application for 3D registration.