| Abstract | According to a recent study, image thresholding can be categorized into six groups of methods that are based on histogram shape, clustering, entropy, attribute, spatial, and local information. In this
paper, we describe two algorithms for image binarization that are based on attribute similarity relying on spatial measures. The rationale of the method is to binarize an image in such a way that it best reproduces the spatial variation of the
original image across several scales. Two different measures that characterize image spatial variation have been selected to pursue that objective: semivariance and lacunarity. Semivariance measures the spatial variation of a variable at a given
scale. Lacunarity is a measure of translational invariance, at a given scale, and is often refer to as a measure of 'gappiness'. In both approaches, the threshold is selected so that the scale-dependant measure in the bi-level image best approximate,
in the least square sense, the ones of the original image. Both methods are illustrated with remote sensing images of high spatial resolution. The results are compared with some other popular thresholding techniques. |