Purchase this article with an account.
C. Ratliff, P. Sterling, V. Balasubramanian; Negative Contrasts Predominate in Natural Images . Invest. Ophthalmol. Vis. Sci. 2005;46(13):4685.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: OFF ganglion cells have smaller dendritic fields and denser sampling arrays than the corresponding ON cells. We tested whether this corresponds to a naturally asymmetric distribution of negative and positive contrasts. Methods: Natural images (100 from van Hateren's library) were convolved with divisively normalized, center–surround filters at a range of spatial scales. The resulting contrasts were rectified into positive and negative channels; then the contrasts for each channel were quantized into ten equally probable response levels. We calculated for small arrays of rectified filters the mutual information that they convey about their intensity inputs. This calculation was repeated for various filter sizes and spacings. Results: For natural scenes, narrower filters were more likely to report contrasts near zero, and broader filters to report a more uniform distribution of contrasts. Regardless of filter size, negative contrasts outnumbered positive contrasts by about 40%. As controls we convolved these filters with two kinds of artificial images: (i) gaussian distribution of intensities with 1/f2 power spectra identical to natural images ('pink noise'); (ii) natural distribution of intensities with 1/f2 power spectra ('natural pink noise'). For pink noise narrow and broad filters reported the same distribution of contrasts with equal proportions of negative and positive contrasts. For natural pink noise, narrow and broad filters reported the same distribution of contrasts, but negative contrasts outnumbered positive contrasts by about 50%. The difference between natural images and natural pink noise indicates that natural scenes contain spatial correlations beyond those that are scale–invariant. These extra correlations account for about half of the information due to local correlations in natural scenes. Conclusions: Natural images comprise asymmetric distributions of contrast, with dark predominating over bright. The magnitude of this asymmetry roughly matches the asymmetric structure of OFF and ON ganglion cell mosaics. The power spectra and intensity distributions of natural images alone explain the predominance of dark over bright, but not the differences observed between the narrow and broad filters.
This PDF is available to Subscribers Only