Font Size: a A A

Research On Algorithms For Planar Contour Processing

Posted on:2007-09-14Degree:DoctorType:Dissertation
Country:ChinaCandidate:B J ZhongFull Text:PDF
GTID:1118360185959776Subject:Aviation Aerospace Manufacturing Engineering
Abstract/Summary:PDF Full Text Request
In this paper some basic algorithms for planar contour processing are discussed, including the algorithms of multiscale corner detecting, multiscale shape describing, and direct corner detecting.In the first part of the paper (Chapter 2, 3 and 4), a theory of direct curvature scale-space (DCSS) and a theory of curvature scale-space (CSS) for corner detection are established. The behavior of isolated single and double corner models in the two kinds of scale-space is investigated. Based on the established theories, the DCSS or CSS image for a given contour curve is transformed into a tree organization, and then corners on the curve are detected in a multiscale sense. Numerical experiments are conducted to show that the multiscale corner detectors can perform successfully on curves with multiscale features.In comparison, the DCSS technique is cheaper in terms of computational costs, while the CSS technique is more effective in suppressing image noise. To combine both the advantages of the two techniques, a hybrid CSS/DCSS scheme is formulated. In particular, a switching criterion is developed, with which the hybrid scheme can switch from CSS to DCSS automatically at the time when noise is suppressed. It is shown that the hybrid corner detector and shape descriptor can work equally well as the CSS technique does on noisy curves, however, at much less cost.In this part, how planar curves shrink in their scale-space is also studied. It is shown that for a curve with corners, it always shrinks at the corner points firstly; and as the scale increases, all the fine features on the curve will disappear. The established theory is helpful for designing multiscale algorithms.In the second part of the paper (Chapter 5), a refinement strategy for digital curves is first proposed, and then a new measure for discrete curvature, referred to as L-curvature, is presented. A refined curve is obtained from a digital curve by connecting each pair of adjacent pixels with a straight-line segment implicitly. Since the refined curve is continuous, precision of calculations on digital images can be...
Keywords/Search Tags:image processing, computer vision, pattern recognition, scale-space, curve convolution, digital curves, curvature, corner detection
PDF Full Text Request
Related items