Super-resolution for synthetic zooming

Abstract

Optical zooming is an important feature of imaging systems. Inthis paper, we investigate a low-cost signal processingalternative to optical zoomingsynthetic zooming bysuper-resolution (SR) techniques. Synthetic zooming is achieved byregistering a sequence of low-resolution (LR) images acquired atvarying focal lengths and reconstructing the SR image at a largerfocal length or increased spatial resolution. Under theassumptions of constant scene depth and zooming speed, we arguethat the motion trajectories of all physical points are related toeach other by a unique vanishing point and present a robusttechnique for estimating its 3D coordinate. Such a line-geometry-based registration is the foundation of SR forsynthetic zooming. We address the issue of data inconsistencyarising from the varying focal length of optical lens during thezooming process. To overcome the difficulty of data inconsistency,we propose a two-stage Delaunay-triangulation-based interpolationfor fusing the LR image data. We also present a PDE-basednonlinear deblurring to accommodate the blindness and variation ofsensor point spread functions. Simulation results with real-worldimages have verified the effectiveness of the proposed SRtechniques for synthetic zooming.