Self-adjusted active contours using multi-directional texture cues

Parameterization is an open issue in active contour research, associated with the cumbersome and time-consuming process of empirical adjustment. This work introduces a novel framework for self-adjustment of region-based active contours, based on multi-directional texture cues. The latter are mined by applying filtering transforms characterized by multi-resolution, anisotropy, localization and directionality. This process yields to entropy-based image “heatmaps”, used to weight the regularization and data fidelity terms, which guide contour evolution. Experimental evaluation is performed on a large benchmark dataset as well as on textured images. The segmentation results demonstrate that the proposed framework is capable of accelerating contour convergence, maintaining a segmentation quality which is comparable to the one obtained by empirically adjusted active contours.