Dynamic Segmentation of Grey-Scale Images in a Computer Model of the Mammalian Retina

Biological studies suggest that neurons in the mammalian retina accomplish a dynamic segmentation of the visual input. When activated by large, high contrast spots, retinal spike trains exhibit high frequency oscillations in the upper gamma band, between 60 to 120 Hz. Despite random phase variations over time, the oscillations recorded from regions responding to the same spot remain phase locked with zero lag whereas the phases recorded from regions activated by separate spots rapidly become uncorrelated. Here, a model of the mammalian retina is used to explore the segmentation of high contrast, gray-scale images containing several well-separated objects. Frequency spectra were computed from lumped spike trains containing 2×2 clusters of neighboring retinal output neurons. Cross-correlation functions were computed between all cell clusters exhibiting significant peaks in the upper gamma band. For each pair of oscillatory cell clusters, the cross-correlation between the lumped spike trains was used to estimate a functional connectivity, given by the peak amplitude in the upper gamma band of the associated frequency spectra. There was a good correspondence between the largest eigenvalues/eigenvectors of the resulting sparse functional connectivity matrix and the individual objects making up the original image, yielding an overall segmentation comparable to that generated by a standard watershed algorithm.