Line Recognition.

Euclid defines a line as having length but no breadth. Lines in the context of machine vision are the perceptually limiting cases of narrow contrasting strips. They are the principal features in engineering drawings and also occur in many natural scenes. Line recognition in machine vision is often treated as if it were an edge detection problem, i.e., lines are found by looking for their edges. Substantial breadth is, however, an a priori assumption in edge detection. The response of linear edge operators to lines is shown here to produce systematic errors in the measured position of the edges, while being inherently noisy when compared to optimized recognition of lines as distinct features. Optimum line detection and localization filters based on orthogonal matched and Wiener filters are developed. We show that the line filters and the Boie-Cox edge filters are closely related and how they are integrated into a combined line and edge recognition system.