Image segmentation: Generic modeling, detection, and estimation of discontinuities in image surfaces

Image segmentation has long been an important module in image understanding. A variety of segmentation algorithms have been developed. However, their performance is still far from satisfactory. By inspecting carefully previous algorithms, it was found that their problems actually originate from the edge model. Their edge models are basically 1-D. These models cannot describe shading. Without accounting for shading, huge biases occur in the detected edge parameters. Because of these biases, the edge linking process is complicated and ad hoc.;In this dissertation, a 2-D model is built to describe the image surface around an edge. Based on this model, a family of edge estimation operators for delta, step, and crease edges is developed from first principles. The operators are based on a common estimation process, which is not sensitive to shading. The performance of these new operators appears to be improved dramatically over any comparable operators.;The statistical performance of the new edge operators is analyzed. The covariance of edgel parameters determines edgel detection from false alarm rate and determines uncertainty in the position and orientation of a detected edgel.;With the help of the new family of edge operators and the model of statistical performance, a new algorithm for linking edgels is presented. This algorithm generates hypotheses of tuples of edgels that fall on a common curve. It uses a constant curvature local model for a curve. Results from this primitive algorithm demonstrate effective extended edges that are informative even in very complex images.