Image understanding requires segregation of the visual scene into perceptual objects. Due to the projection of the three-dimensional world on two-dimensional sensor surfaces, objects closer to the observer occlude those which are more distant. At any given occlusion border, it is important to decide which side is the foreground (figure) and which is the background, a decision which is influenced both by global and local image contents. In this report, we focus on local cues. We randomly select small image patches located on figure-ground borders in complex natural scenes. Spectral anisotropy features are extracted from the patches and used to train a non-linear Support Vector Machine. Using data from two large image databases (LabelMe and BSDS300), the classifier achieves an accuracy near 70% per local patch on the task of deciding which side of an occlusion is the foreground. Although in many cases global influences are important for figure-ground segregation, we suggest that the low computational cost of local computation can make it a useful strategy for figure-ground segregation.