A TV flow based local scale estimate and its application to texture discrimination

This paper presents a local region based scale measure, which exploits properties of a certain type of nonlinear diffusion, the so-called total variation (TV) flow. During the signal evolution by means of TV flow, pixels change their value with a speed that is inversely proportional to the size of the region they belong to. From this evolution speed, one can derive a local scale estimate based on regions instead of derivative filters. The main motivation for such a scale measure is its application to texture discrimination, in particular the construction of an alternative to Gabor filters. When the scale estimate is combined with the components of the structure tensor, which provides orientation information, it yields a texture feature space of only four dimensions. Like Gabor features, this sparse feature space discriminates textures by means of their orientation and scale, yet the representation of orientation and scale is less redundant. The quality of the feature space containing the new scale measure is evaluated in texture segmentation experiments by comparing results to those achieved with Gabor filters. It turns out that one can gain a total speedup of a factor 2 without loosing any quality concerning the discrimination of textures.