Polygonal approximation of digital curves using adaptive MCMC sampling

Polygonal approximation (PA) of the digital planar curves is an important topic in computer vision community. In this paper, we address this problem in the energy-minimization framework. We present a novel stochastic search scheme, which combines a split-and-merge process and a stochastic approximation Monte Carlo (SAMC) sampling procedure for global optimization. The SAMC sampling method can effectively handle the local-trap problem suffered by many local search methods, while the split-and-merge process is used to construct a more informative proposal distribution, and thus further improves the overall sampling efficiency. Experimental results on various benchmarks show that the proposed algorithm can achieve high-quality solutions and comparable results to those of state-of-the-art methods.