<XML><RECORDS><RECORD><REFERENCE_TYPE>31</REFERENCE_TYPE><REFNUM>8032</REFNUM><AUTHORS><AUTHOR>Balasuriya,L.S.</AUTHOR><AUTHOR>Siebert,J.P.</AUTHOR></AUTHORS><YEAR>2005</YEAR><TITLE>Space-Variant Vision using an Irregularly Tessellated Artificial Retina</TITLE><PLACE_PUBLISHED>Workshop on Biologically-Inspired Models and Hardware for Human-like Intelligent Functions, Montreal</PLACE_PUBLISHED><PUBLISHER>N/A</PUBLISHER><LABEL>Balasuriya:2005:8032</LABEL><ABSTRACT>Space-variant processing reduces the combinatorial explosion of information involved in vision. Instead of evaluating the whole field-of-view at full resolution, system resources are biased towards more salient areas in the scene. Conventional machine vision techniques such as quadtree decomposition change the allocation of sampling resources to adapt to content in an image. However, biological systems use a fixed space-variant retina to control the allocation of resources to the scene. The processing machinery used to extract and process visual information is fixed and the animal’s whole retina is rotated towards different salient regions in the scene. Most artificial retinae used in machine vision are based on a retino-cortical transform that projects locations in the field-of-view to an associated ‘cortical space’ [1, 2]. These are based on the projection of the eccentricity of a location in the field-of-view and do not consider the locations of real receptive fields in the retina. Therefore retinae based on these transforms tend to over-sample visual information in the foveal region when sampling images [1] or have distorted foveal regions [2]. Some researchers have processed the foveal region using a separately generated receptive field mosaic [3, 4, 5]. But these caused a discontinuity in the internal representation and sampling of visual information. The authors have yet to discover an analytic mapping that can describe the gradual change in the receptive field topography of the retina from the central uniform density foveal region and the space-variant periphery. In this work the authors used a self-organisation methodology [6] to generate a retinal receptive field tessellation that seamlessly progresses from a central uniform to a space-variant peripheral mosaic. The structure of the retina locally resembles an irregular hexagonal lattice with continuity in the receptive field density. The support region of the receptive fields in the artificial retina is inversely proportional to the self-organised retina’s local receptive field density. This results in space-variant receptive fields, fine in the high density fovea and increasingly coarse in the space-variant periphery. A complete vision system was constructed using a multi-resolution pyramid of concentric irregular artificial retinae with Laplacian of Gaussian receptive fields [7]. Interest points based on scale-space extrema [8] at scene corners [9] were computed based on the space-variant information extracted by the retina. The authors calculated a local feature vector based on neighbouring receptive field response gradients similar to [10] at the exact spatial and frequency location of the detected interest points. The system is able to process an image, fixating upon areas which it finds interesting and generating interest points right across its field-of-view. Interest points generated in the foveal region are used for reasoning while those in the periphery are useful for attention i.e. deciding the next fixation point. The feature vectors extracted by the system may be used for higher level vision tasks such as object recognition, attention and visual search. The authors believe that the presented work is a useful tool for the future investigation of foveated space-variant vision and are currently extending the system to exhibit task based attention behaviour.</ABSTRACT></RECORD></RECORDS></XML>