A Hybrid Approach to Segmenting Hair in Dermoscopic Images Using a Universal Kernel

Hair occlusion often causes automated melanoma diagnostic systems to fail. We present a new method to segment hair in dermoscopic images. First, all possible dark and light hairs are amplified without prejudice with a universal matched filtering kernel. We then process the filter response with a novel tracing algorithm to get a raw hair mask. This raw mask is skeletonized to contain only the centerlines of all the possible hairs. Then the centerlines are verified by applying a model checker on the response and the original images. If a centerline indeed corresponds to a hair, the hair is reconstructed; otherwise it is rejected. The result is a clean hair mask which can be used to disocclude hair. Application on real dermoscopic images yields good results for thick hair of varying colours. The algorithm also performs well on skin images with a mixture of both dark and light hair.