Mask Matching for Linear Feature Detection.

Abstract : Mask matching is a well known procedure (1) in which the detection of specific features in an image is carried out by matching a set of templates or masks with the neighborhood of each pixel in the image. This paper describes a set of 3 X 3 binary masks that can be used to extract thin features from an image. The masks are used to assign various labels to each pixel (each label corresponding t a particular mask), and to associate with each label a set of confidence measures based on the homogeneity of the foreground and background in the mask, and the difference between them. The idea of this approach is to record a large set of useful information at the pixel level, in order to efficiently make use of it at the later stages of the linear feature detection process. Given is a probabilistic analysis of the frequency of matches and their expected robustness for specific masks and classes of masks in white noise images. These results may help indicate whether or not a given image region should be considered interesting, as regards frequency of occurrence of line-like masks, for example.