Extracting straight edges in images without conventional thresholding

The work described uses a different approach to the grouping of pixels according to edge orientation, with the intention of developing this concept three stages further. Firstly, by introducing a two-stage approach to the grouping of edge pixels, the thorny problem of dealing with two disjoint representations is avoided. As a result, the possibility of fragmentation (whereby a `long' edge in the image is represented by two or more generated lines) is minimised. Secondly, it is the aim of the work to establish a 1:1 correspondence between generated edge support regions and valid linear features in the image (i.e. all valid edges are successfully detected and no spurious additional edges are generated as a by-product). Thirdly, the technique seeks to determine all vertices in the image to sub-pixel accuracy