| An aim of conventional low level image processing is to generate useful binary images from gray scale images. The binary images generated by enhancement and threshold edge detectors are usually unrefined outlines of the underlying 3-D scene. Such images must be further processed to isolate and identify region boundaries; which, in the case of polyhedra, consist of straight line segments. The line segments and intersections or connection points (which are referred to as vertices) collectively form a line sketch of the original scene. This line sketch contains considerable information about objects in the scene and may be used efficiently to identify and locate targets or objects.;In this thesis we devise and explore methods for extracting reliable line sketch renderings of scenes with primarily polyhedral shapes. The methods fall into three categories: finding and isolating straight segments, identifying vertices, and establishing consistencies in the evolving sketch. Straight segments are located in images using modified Hough Transforms and novel projection based schemes. The detection of the absence or presence (and type) of line junctions or vertices is accomplished using likelihood functions derived from assumptions of an underlying noise contamination model. Consistencies in the sketch are propagated and inconsistencies are resolved using various graph models and reduction and labeling algorithms.;Experiments have been conducted and the results reported in this thesis to verify the utility and judge the performance of the proposed schemes. Synthetic images with controlled amounts of noise are used to compare the different methods and to quantify the effectiveness under varying noise conditions. The results show that the proposed schemes are useful and may be used to obtain line sketches for object recognition. |
