Resolving edge-line ambiguities using probabilistic relaxation

A Bayesian model is presented that captures the uncertainties introduced when feature detection and classification are attempted using oriented quadrature filter pairs. This model commences by constructing probability density functions for the quadrature filter responses. These densities are used to compute initial probabilities for edge and line labels. The ambiguities inherent in classifying edge and line features are captured in terms of the relative phase of the filter responses. The probabilistic representation of the quadrature filter bank is then refined using a novel dictionary-based relaxation scheme to obtain unambiguous and consistent feature contours. Experiments on noisy aerial infrared images illustrate the power and robustness of the technique