Saccadic Object Recognition with an Active

An active vision system for saccadic camera gaze shifts and explorative scene analysis as a new integral approach to image understanding is proposed. The model includes several subsystems: preattentive peripheral feature detection, multi resolution foveal image identiication based on a hypercolumnar representation and object recognition by means of two memories for foveal identiications and xation positions. An egocentric interest map and a selective masking mechanism are used to integrate bottom{up and top{ down information ow. In this paper we propose an animate image understanding system for two{dimensional scene analysis and object recognition by means of camera gaze shifts ((g. 1). The function of the system is designed by a hierarchy of physiological and anatomical motivated structures (i.e. fovea, hypercolumns). The scene is supposed to be a mainly static external memory buuer. An implicit selective attention mechanism modulates perception by controlling the information ow and distributing processing resources to scene locations. A space{variant area magniication factor of a scale{sized fovea model 3] provides a small central area of high resolution coinciding the camera gaze direction and a large low resolution periphery. Loci of peripheral spatiotemporal features discontinu-ities are detected by a preattentive feature analysis. By foveating these spatial positions (bottom{up saccades) an attentive identiication process memorizes characteristic patterns. We use a pattern recognition subsystem evaluating local orientations on diierent scales. It supplies a correlation histogram based on a large set of stored foveal patterns 2] and is invariant to restricted translations and distortions. For reasons of simplicity the location of the attentive processing window is always chosen identically to the foveal area (overt attention). Thus the system is forced to shift the camera gaze for an object recognition task. The cognitive representation of an object is constructed of a set of foveal views. An interactive su-z Edited version, originally published in: B. Neu-mann (ed.), Proc. of the ECAI 92, pp. 803{805, Wiley and Sons (1992) pervised learning procedure guided by the peripheral preattentive feature detection module selects conspicuous views of an object. Their foveal patterns are stored in the pattern recognition system, while the object recognition only stores the foveal pattern identii-cations and their object centered spatial coordinates. By transsaccadic information integration a cognitive object recognition process states hypotheses about the scene and tries to verify them by claiming saccades to scene locations speciic for an object (top{down sac-cades). A selective masking of the preattentive pathway enhances the performance of the …