A biomimetic real time retinal simulation software with non separable spatiotemporal processing, chromaticity and non linear local adaptivity - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2007

A biomimetic real time retinal simulation software with non separable spatiotemporal processing, chromaticity and non linear local adaptivity

Résumé

The early visual system has probably been the most widely studied part of the human brain. As a consequence, it is now possible to confront structural observations, neurophysiological data and functional properties of this early stage in visual processing. One way to integrates these approaches is modelization and simulation. In this field, the use of analogies between retinal cells and retinal wiring with electronic circuits proved to be a particularly relevant way to understand signal processing in the early visual system and to reproduce its functionnal properties (Mead & Mahowald, 1988; Beaudot, 1994; Zaghloul & Boahen, 2004). Those models led to hardware implementations. However, those simulations require heavy technological means and remain difficult to manipulate. To the opposite, computationnal model may be implemented as software (Shah & Levine, 1996; Pelayo et al., 2004), thus benefiting from a good adjustability and versatility. However, sophisticated simulation would not be able to work real time, and simplified ones would be weakly biologically reliable. To answer this problem, we developed an analytic model of the retina based on the transfer functions of the "retinal circuits" designed for electronic simulations (Hérault & Durette, 2007). We then implemented them as software to propose a biologically reliable retina simulation that is adjustable, versatile, and works real-time on a standard PC-Computer. As a main difference with other computational models, the transfer functions of our retinal cells layers are high-pass spatiotemporal filters with non-separable variables. In other computational simulation, the non-separability beween space and time is eluded. Linking these filters by a simple architecture directly inspired from the synaptic patterns of connections in the retina leads to the generation of three different streams that match the known properties of the Magno- and Parvo-cellular pathways. Non-separability plays a key role in our approach: it is responsible for the main part of computational parallelism; it enables adaptive local inhibition in space and also in time, and gives a further insight into the so-called "Coarse-to-Fine" streaming process in the visual system. Local spatial and temporal gain-controls are replicated, at both the levels of photoreceptors and of ganglion cells, using an adaptive Michaelis-Menten law. It is responsible respectively for luminance and contrast equalization, which makes it suitable, particularly, for texture extraction. Its last feature is the capability of chromatic computation at the same rate as monochromatic computation using the bio-inspired mosaic encoding scheme for color (Alleysson et al., 2005). The software application is the first retinal simulator software built on a strong physiological background which accounts for many functional observations such as non-separable spatio-temporal behavior, color, adaptive gain control, and is optimized enough to work in real time on a PC-computer. It is modular and fully configurable so that it can be used in a wide field of applications, from vision research to real-time image processing and computer vision (Benoit et al., 2007, ERM submitted abstract).
Fichier non déposé

Dates et versions

hal-00258601 , version 1 (22-02-2008)

Identifiants

  • HAL Id : hal-00258601 , version 1

Citer

Barthélémy Durette, Alexandre Benoit, Brice Chaix de Lavarène, David Alleysson, Jeanny Hérault. A biomimetic real time retinal simulation software with non separable spatiotemporal processing, chromaticity and non linear local adaptivity. European Retina Meeting, Oct 2007, Frankfurt/Main, Germany. pp.S4-P03. ⟨hal-00258601⟩
216 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More