Form specific adaptation and pattern recognition: an experimental and theoretical study

Abstract

Among the theories of visual pattern recognition are structural theories which propose that patterns are encoded in terras of features and their spatial relationships (relations). Structural theories are examined here with both adaptation and pattern recognition techniques. In particular, the effects of changes in the relational- structure of patterns formed from bear and disc features are examined. For this purpose a novel adaptation technique is developed and used for measurements of the contrast threshold elevation effect.
Data are presented which show that the visual system is adaptationally sensitive to the shape of local features, but adaptationally insensitive to their relative positions. An exception to the latter conclusion was the finding of adaptational sensitivity to local periodicity. It is argued, however, that this periodicity sensitivity may simply be a result of size selectivity.
Data are also presented which were obtained in a discrimination- under-uncertainty experiment in which stimuli analogous to those in the adaptation experiments were used. These data reveal sensitivity to local feature changes and insensitivity to feature relative- position changes. Because of the similarity in results from the adaptation and discrimination-under-uncertainty experiments it is argued that both techniques reveal the properties of the initial stages of pattern processing. It is also argued that these results show a fundamental difference in the way in which features and relations are processed prior to pattern recognition.
To investigate how structural theories may be applied in a pattern recognition task a relational-structure encoding model is developed and its predicted pattern recognition performance is compared with experimental data. Once equipped with the capability of performing certain discrete operations on the relational-structure representation this model-provides a good fit to the experimental data.