gms | German Medical Science

Artificial Vision — The 2nd Bonn Dialogue. The International Symposium on Visual Prosthesis

Retina Implant Foundation

19.09.2009, Bonn

Learning of image encoding

Meeting Abstract

Suche in Medline nach

  • author Rolf Eckmiller - Department of Computer Science Division of Neuroinformatics, University Bonn, Germany

Artificial Vision – The 2nd Bonn Dialogue. The International Symposium on Visual Prosthesis. Bonn, 19.-19.09.2009. Düsseldorf: German Medical Science GMS Publishing House; 2009. Doc09ri06

doi: 10.3205/09ri06, urn:nbn:de:0183-09ri060

Veröffentlicht: 30. November 2009

© 2009 Eckmiller.
Dieser Artikel ist ein Open Access-Artikel und steht unter den Creative Commons Lizenzbedingungen (http://creativecommons.org/licenses/by-nc-nd/3.0/deed.de). Er darf vervielfältigt, verbreitet und öffentlich zugänglich gemacht werden, vorausgesetzt dass Autor und Quelle genannt werden.


Gliederung

Text

Purpose: Simulation study of a novel retina implant with perceptual feedback, oculomotor feedback for simulated miniature eye movements (SM), and with optional feedback from spontaneous nerve impulses (NI) of the stimulated regions.

Methods: Simulations were performed with a novel retina encoder (RE-3) with a filter module (FM) consisting of an input array of 20x20 pixels for presentation of a pattern P1 and 100 spatio-temporal (ST) filters for generation of selective stimulation signals for 100 electrodes at the retinal output, a novel inverter module (IM-3) to mimic parts of the central visual system and to map the FMRef-output onto a simulated percept P2, and a dialog module (DM). DM simulated the perceptual feedback from a human user. SMs were generated on demand for movements of P1 by one pixel in a given direction to mimic typical eye movements during fixation. Spontaneous NI could be considered to avoid stimulus pattern disturbances: a) stimulation signals with NI-suppression capability and b) neural feedback from neural tissue to RE-3 via bi-directional electrode arrays for stimulation and recording.

Results: (a) FM was specified as a regular distribution of three ST filter types to mimic receptive field properties of primate retinal ganglion cells. (b) Both P1 and FM filter array were described as multi-dimensional vector matrices, which had several advantages, including easy changes of pixel numbers or filter numbers as well as FM-output calculation as matrix products. (c) The software for IM-3 was designed to process the FM-output matrices with an efficient algorithm and to invert the partly ambiguous FM mapping with the help of SMs. (d) Due to the matrix structure, inversion of the FM-output by IM-3 could be processed in a single matrix run through all coefficients.

Conclusions: Since ‘Gestalt’ perception in humans requires active vision as a continuous interaction between sensory and oculomotor processes, a combination of neural-, oculomotor, and perceptual feedback may be important to optimize the function of bionic visual prostheses.

This lecture is available as video recording (Attachment 1 [Attach. 1]).