A facet based road network detection procedure is described for radar imagery. The procedure includes a line detection part and a road detection and connection part. The line detection part analytically detects linear features using a facet Valley finding technique. Valleys are defined as zero crossings of the first directional derivatives of a bicubic facet model taken in a direction extremizing the second directional derivative. The road detection and connection part statistically screens the linear features on a component by component basis, and then optimally connects the screened linear features using a dynamic programming algorithm.

This thesis also includes as a preprocessing technique for noisy images, an adaptive noise removal algorithm, and suggests a practical method of estimating a local noise variance.