Nonlinear operator for blob texture segmentation

Texture is an important part of the visual world of animals and men and their visual systems successfully detect, discriminate and segment texture. Relatively recently progress was made concerning structures in the brain which are presumably responsible for texture processing. Von der Heydt et al. (von der Heydt et al. 1992) reported on the discovery of a texture processing neuron in areas V1 and V2 of the visual cortex of monkeys which they called grating cell. Grating cells respond vigorously to gratings of bars of appropriate orientation, position and periodicity. In contrast to other orientation selective cells, grating cells respond very weakly or not at all to single bars which do not make part of a grating. This behaviour of grating cells cannot be explained by linear filtering followed by half-wave rectification as in the case of simple cells, neither can it be explained by threestage models of the type used for complex cells. Elsewhere we proposed a model of this type of cell and demonstrated the advantages of grating cells with respect to the separation of texture and form information (Kruizinga & Petkov 1995, Petkov & Kruizinga 1997). Tanaka et al. (1991) found another type of texture processing neuron, that responds to dot-patterns. These texture cells, which we call blob-texture cells in the following, have similar characteristics as grating cells. They do not react to single dots but only to a pattern consisting of a number of dots. Neurophysiological experiments revealed a preference of the cells for a regular dot pattern in comparison to more random patterns of dots. Grating cells are not activated by these random dot patterns, though regular dot patterns cause a slight grating cell response. In this paper we propose a computational model of blob-texture cells that is capable of explaining the results of neurophysiological experiments. Our model of blob-texture cells consists of three consecutive stages. The final stage (blob-texture cells) receives its inputs from multiple units in the second stage, the so-called blob-pattern subunits, which in turn receive their input of the first stage (blob detectors). Furthermore, the model is used as an image processing operator and compared with existing texture operators like the Gabor-energy operator and the cooccurrence matrix operator. This evaluation is done by comparing the results of a texture segmentation task, in which an image containing a number of blob textures is segmented on the basis of features obtained with the three texture operators. The method is similar to the evaluation of the grating cell operator, with respect to the processing of oriented texture (Kruizinga & Petkov 1998).