Discontinuity-Adaptive shape from focus using a non-convex prior

Shape from focus (SFF) is a widely used technique for determining the 3D structure of textured microscopic objects. However, SFF output depends critically on the number of observations used and the focus measure operator adopted. In this paper, we propose a new SFF method that can provide rich structure information given limited number of observations. We observe that depth is non-linearly related to the observations and pose the shape estimation as a minimization problem within a Maximum A Posteriori (MAP) - Markov Random Field (MRF) framework. We incorporate a discontinuity-adaptive MRF prior for the underlying structure. The resulting cost function is non-convex in nature which we minimize using Graduated non-convexity algorithm. When tested on synthetic as well as real objects, the results obtained are quite impressive. © 2009 Springer Berlin Heidelberg.