Multi-orientation analysis by decomposing the structure tensor and clustering

The structure tensor yields an excellent characterization of the local dimensionality and the corresponding orientation for simple neighborhoods, i.e. neighborhoods exhibiting a single orientation. We show that we can disentangle crossing structures if the tensor scale is much larger than the gradient scale. Mapping the gradient vectors to a continuous orientation representation yields a frac12D(D+1)-dimensional feature vector per pixel. Clustering of the vectors in this new space allows identification of multiple orientations. Each cluster of gradient vectors can be analyzed separately using the structure tensor approach. Proper clustering yields an unbiased estimate of the underlying orientations