Abstract
In a very recent work (Bertalmío 2014) we proposed a neural model, derived from an image processing technique for color enhancement (Bertalmío et al. 2007), that is able to predict lightness induction phenomena, to perform contrast enhancement and to improve the efficiency of the representation by flattening both the histogram and the power spectrum of the image signal. This neural model consists in a modification of the generalized Wilson-Cowan equations (Wilson and Cowan 1973) for neural activity in V1, in which the multiplicative weight associated with the sigmoid summation term is allowed to vary pixel-wise, depending on the (local) standard deviation of the signal, instead of being a global constant like in the original formulation. This modification was not suggested by any kind of neurophysiological data, but rather it was adopted because it was able to provide results showing lightness assimilation, whereas our use of the original Wilson-Cowan formulation was only able to produce lightness contrast. Nonetheless, it may appear as an ad-hoc change in the model, not easy to reconcile with the basic postulates of Wilson and Cowan's theory. In this talk we show how we can employ the original Wilson-Cowan formulation and nonetheless obtain assimilation results, as well as contrast enhancement and redundancy reduction effects. The key lies in letting the size of the summation kernel be much smaller than the extent over which the mean average estimate for the activity baseline is computed. This correlates well with data on spatial summation in the visual cortex.