Vicky Cai¹, Jonathan Victor²; ¹Stuyvesant High School, New York, USA, ²Weill Cornell Medical College, New York, USA

Typical of sensory systems, many aspects of human vision are tuned to the characteristics of its natural inputs. These characteristics include global statistics (power spectrum) and local statistics (multipoint correlations). While the basis for the global statistics of natural scenes is largely understood, the origin of the distinctive pattern of local image statistics is not. Medical radiographs constitute a useful contrast: they have a steeper spectral slope (-2.6 to -3, vs. -2 for natural images), a different pattern of local statistics, and analysis of radiographs has important consequences. Moreover, natural images and radiographs are formed by different physical processes: occlusion is typical in natural images; transparency is typical in radiographs. To address the origin of the statistical properties of these image classes, we built a series of generative models. Starting with the “dead leaves model” for natural images (Ruderman 1997), we replaced occlusion by transparency, generalized it to 3D, and considered several object size distributions. We then calculated global and local image statistics for each model across a range of scales. As expected, (Metheany et al., 2008), the shift from occlusion to transparency accounted for spectral slopes. For local image statistics, we considered the means of multipoint correlations (overall characteristics of each image class) and their standard deviations (characteristics that distinguish images within a class). All models accounted for the prominence of pairwise statistics, compared to three-point and four-point statistics, seen in natural and medical images. Occlusion vs. transparency and size distribution had a large effect on the mean values of multipoint correlations. They had a smaller effect on their standard deviations and did not account for the differences between three-point and four-point statistics seen in natural images and radiographs. Overall, simple generative models can explain some, but not all, of the statistical characteristics of natural images and radiographs.

Acknowledgements: NIH EY07977, Fred Plum Fellowship in Systems Neurology and Neuroscience