Figure/ground separation using stochastic pyramid relinking

摘要

This paper presents an algorithm to perform figure/ground separation, which is based on the global optimization of a functional representing the local fit error of an assumed model describing the variation of luminance over the local regions in the image. The optimization is carried out using the simulated annealing algorithm with state changes not being limited to single pixel sites but to whole regions described by the linkage structure of pyramid data. The algorithm is noteworthy in that it performs well even in low signal-to-noise ratios and requires no pre-assigned parameters such as weighting/smoothing parameters or thresholds (except for the parameters controlling the simulated annealing such as the cooling schedule parameter). A number of computer simulations on synthetic data are presented which demonstrate the performance of the algorithm.