Super-resolution stereo light microscopy for real-time tasks using normal flow and geometric constraints

Prior knowledge about the observed scene represents the key to recovering frequencies beyond the passband of an imaging system (super-resolution). With regard to microscopy mainly two super-resolution mechanisms have been reported: 1.) analytic continuation of the frequency spectrum and 2.) constrained image deconvolution. This paper describes an alternative super-resolution mechanism. Prior knowledge is introduced on a higher, more symbolic level of visual inference. We exemplify our concept based on the 3D reconstruction of a micro-pipette moving in close proximity to immobile target objects. Information about the shape and the mobility of the pipette is incorporated in order to localize the pipette tip at sub-Rayleigh distances to the target. The algorithm was tested in a micro-robot environment. A machine vision module using stereo light microscopy automatically controlled the manipulation of microscopic objects, e.g. latex beads or diamond mono-crystals. In the theoretical part of the paper we prove that knowledge of the form `the pipette has moved between two consecutive frames of the movie' must result in a twofold increase of resolution. We used the normal flow of an image to decode positional measures from motion. In practice, super-resolution factors between 3 to 5 were obtained. The additional gain originates from the geometric constraints which were included to reconstruct the pipette position.