Multi-Scale Blur Estimation and Edge Type Classification for Scene Analysis

摘要

Signatures, in this work, are multi-scale representations of local gray-level information, tied to places in gray scale images where regional differences are locally maximal. The information may involve the regional differences themselves (called Gaussian differences or signed normalized gradient magnitudes, (Korn, 1988)), or, distance relations between edges (apparent width measurements), or, absence of edges in pulse edge pairs, at coarser scales. Using signatures involves the classical problem mentioned by Marr and others of relating information across scales. A novel result is that a fruitful way of doing this is to build scale paths from coarse-to-fine exploiting edge focusing and associate with pixel positions, along these paths, the three quantities Gaussian differences, apparent width and the binary information absence/presence of edges (in edge-pairs). Such a structure, if used together with proper conditional tests, serves the purpose of classifying edges with respect to profile-type, and can also be used for measuring global contrast, degree of diffuseness, deblurred line width, and qualitative labels such as diffuse versus sharp. The structure is used simultaneously for labelling tasks and quantitative measurements. Theory on apparent widths, absence/presence of edges in pulse edge pairs is developed. For measuring diffuseness and global contrast from Gaussian difference signatures a linear least squares approach is suggested. Extensive experimental results are presented. Possible applications are in image segementation, junction analysis, and depth-from-defocus. For the purpose of distinguishing between objects and illumination phenomena, such as diffuse shadow edges, classification of contours with respect to diffuseness seems useful.