Unsupervised level set parameterization using multi-scale filtering

This paper presents a novel framework for unsupervised level set parameterization using multi-scale filtering. A standard multi-scale, directional filtering algorithm is used in order to capture the orientation coherence in edge regions. The latter is encoded in entropy-based image `heatmaps', which are able to weight forces guiding level set evolution. Experiments are conducted on two large benchmark databases as well as on real proteomics images. The experimental results demonstrate that the proposed framework is capable of accelerating contour convergence, whereas it obtains a segmentation quality comparable to the one obtained with empirically optimized parameterization.