Font Size: a A A

Superresolution from degraded image sequence using spatial tessellations and wavelets

Posted on:2004-09-03Degree:Ph.DType:Dissertation
University:The Pennsylvania State UniversityCandidate:Lertrattanapanich, SurapongFull Text:PDF
GTID:1468390011962501Subject:Engineering
Abstract/Summary:
During the past decade, the interest in seeking enhancement of spatial resolution leading to a high-resolution (HR) image from a sequence of degraded (undersampled, blurred, and noisy) low-resolution (LR) images has been notable. HR images are required in many areas including remote sensing, military surveillance and medical diagnosis. However, it is usually not possible at the outset to achieve the desired resolution because of technology and cost constraints. In applications like astronomical imaging, the reduced size and weight of a camera in a spacecraft or satellite affect its quality. The need to perform a trade-off between size, weight, and quality necessitates the design of superresolution algorithms to obtain the desired HR image.; The superresolution system can be divided into three main parts: preprocessing part, the core superresolution algorithm, and postprocessing part. In the preprocessing phase, bad frame elimination and sampling structure singularity prevention are per formed prior to the implementation of the superresolution algorithm. This research focuses on two superresolution algorithms: one is wavelet based and the other is Delaunay triangulation based. A recent wavelet based superresolution algorithm (wavelet superresolution) is refined so that it can handle the versatile projective camera motion model. Key tools for coarse analysis developed include a method to select suitable starting scale and the maximum scale needed to prevent, respectively, the undesired effects caused by bad shift and an underdetermined system of equations. Various types of finitely supported wavelet families including Haar, Daubechies, Coifiets, Symlets, and B-spline are used and extensively tested. The B-spline wavelet family gives the most satisfactory result for wavelet superresolution algorithm from both computation and quality of approximation standpoints.; The other superresolution algorithm developed here is the Delaunay triangulation based high resolution (DTHR) algorithm. The DTHR algorithm is accompanied by a site-insertion algorithm (which is a local update feature of Delaunay triangulation) for updating the initial HR image with the availability of more LR frames till the desired image quality is attained. The surface used in this research to approximate each triangle patch in the Delaunay triangulation is modeled by bivariate polynomials subject to the imposition of smoothness constraints. Surface approximation with B-splines is also possible. The DTHR algorithm is suitable for real-time image sequence processing because of its local computational capability, the fast expected (averaged) time construction of Delaunay triangulation, local update feature, and feasibility of hardware implementation.; In the postprocessing phase, multiframe noise filtering and deblurring (following blur estimation) schemes are developed. A simple but effective multiframe noise filtering scheme based on image averaging is proposed. An iterative blind deconvolution (IBD) algorithm is selected to implement multiframe blind deconvolution. The overall superresolution system developed incorporating the core superresolution algorithm, multiframe noise filtering, and multiframe blind deconvolution is tested on a real video sequence supplied by the Air Force Research Laboratory (AFRL).
Keywords/Search Tags:Image, Superresolution, Sequence, Wavelet, Blind deconvolution, Delaunay triangulation, Multiframe noise filtering
Related items