On the generalization of symmetry relations in visual pattern classification

Abstract-Human observers were trained to criterion in classifying compound Gabor signals that were pairwise mirror-images of each other (left and right versions). They were then tested with 18 thresholded versions of the learning set. For the initial learning stage, the probability of confusing mirror-image patterns was significantly higher than that of confusing nonsymmetric patterns. This difference was greatly reduced for the test data, thus indicating a high degree of generalization of the concept of symmetry. The generalization to dark-only and light-only test signals, as well as to dark-and-light test signals, was found to be equally good. The behavioural data were analyzed in terms of various types of signal representation (pixelwise, Laplacian pyramid, curvature pyramid, ON/OFF, local maxima of Laplacian and curvature operators) and a minimum-distance rule. Results were suggestive of the existence of a structure-oriented ‘blob-code’. Whether such a code could be used in conjunction with simple classifiers or should be transformed into a propositional scheme of representation operated upon by a rule-based classification process remained an open question.