A Faithful DoG is All you Need

Abstract

Compared to edge-based models, region-based active contour models (ACM) have demonstrated superior performance in a number of areas, including noise tolerance, back- ground complexity and inhomogeneity correction, initialization resilience, and speed of curve evolution. However, combining both of their credentials with suitable and relevant parameters exhibits promising potential in enhancing segmentation performance. Therefore, this work reports an effective fusion of optimized Difference of Gaussian (DoG) edge estimation, with the region scalable fitting ( RSF) m odel t o c apitalize o n t heir a ttributes. A locally computed edge entropy image is also used as a weight to the energy functional to infuse local edge information in the energy functional. With the integration of relevant edge and region based feature descriptors, the proposed model thereby, outperforms the established ACMs in terms of iteration time, noise tolerance, initial contour convergence, inhomogeneity suppression and segmentation accuracy.