Image segmentation using maximum gradient profiles orthogonal to edges

For each pixel in an image the local direction of greatest grey level change was determined, and pixels were linked in this direction to form maximum gradient profiles (MGP). Adjacent MGPs were linked side-by-side to form hillside regions. Within each hillside region a path close to the points of steepest slope on each MGP was established, and is referred to as an edge section. The edge sections were found to be similar to those from a Canny operator. For the present application, which was concerned with detection and characterization of blood vessels on angiograms, subsequent grouping was restricted to anti-symmetric hillside regions which formed a ridge with high grey values in the centre. The ridges corresponded closely to blood vessels shown in the image. The width of each ridge and the locus of its centre line were derived from anti-symmetric MGP pairs, and showed good agreement with expert judgement of vessel width and centre line location. Advantages of the MGP grouping method include: explicit representation of the profiles of grey values orthogonal to edges; ease of implementation of higher-order grouping stages, which are natural extensions of lower level grouping; and a hierarchical data structure, which allows improved interaction of bottom-up and top-down processing.