Minimal Weighted Local Variance as Edge Detector for Active Contour Models

Performing segmentation of narrow, elongated structures with low contrast boundaries is a challenging problem. Boundaries of these structures are difficult to be located when noise exists or intensity of objects and background is varying. Using the active contour methods, this paper proposes a new vector field for detection of such structures. In this paper, unlike other work, object boundaries are not defined by intensity gradient but statistics obtained from a set of filters applied on an image. The direction and magnitude of edges are estimated such that the minimal weighted local variance condition is satisfied. This can effectively prevent contour leakage and discontinuity by linking disconnected boundaries with coherent orientation. It is experimentally shown that our method is robust to intensity variation in the image, and very suitable to deal with images with narrow structures and blurry edges, such as blood vessels.