Abstract
Neurons in the early visual system are often considered “edge detectorsr”. However, there is some psychophysical evidence that supports an alternative view, in which the junctions of edges are often more salient than edges themselves. Last year we presented preliminary results (Martinez-Conde et al., VSS'02) showing that single units in the LGN and area V1 of the awake monkey respond more strongly to junctions than to edges. These results suggested that junction processing may not be a subsequent stage to edge processing, but that junctions may in fact be a more optimal stimulus than edges for early (i.e. center-surround) receptive fields. We have now recorded from 110 neurons in the LGN and area V1, confirming our preliminary conclusions. We have now also developed theoretical computational models of junction processing in the retina/LGN and area V1, which make quantifiable and principled predictions of how early neurons of the visual system should respond to junctions versus edges. By measuring the response from these models to visual scenes (both illusory and non-illusory) that contain edges and junctions we have discovered that early visual receptive fields (i.e. those from center-surround cells, simple cells, and complex cells) are shaped so as to theoretically respond more strongly to junctions than to edges. This is the first time that the salience of junctions has been quantified against the salience of edges, either physiologically or computationally. Both approaches match our qualitative perception of visual illusions in which junctions are more salient than edges. Moreover, our data are compatible with the idea of junctions being processed very early in the visual system (i.e. starting in the retina, with antagonistic center-surround receptive fields), without the implication of high-level mechanisms, such as feedback loops or intracortical circuits.