Attentional Scanning

Eric O. Postma1, H. Jaap van den Herik1, and Patrick T.W. Hudson1 Abstract. A model for attentional scanning is constructed in the form of a gating network which consists of gating lattices. A gating lattice is a sparsely-connected neural network. The process of covert attention is interpreted as a biological solution to the problem of translation-invariant pattern processing. We arrive at the nal result by a sequence of pattern translations channelled through the gating network. Simulation studies and theoretical considerations reveal that the gating lattice gives rise to a trade o between gating quality and gating exibility. The gating network is shown to be capable of translation-invariant processing of object patterns that are part of a natural image. 1 BACKGROUND Visual systems succeed remarkably well in extracting invariant properties of incoming patterns. An example of such a property is object identity. Visual priming (i.e., the facilitated speed and enhanced accuracy of identi cation due to prior object exposure) has been shown to be independent of the position of prime and object [1]. Our aim is to realise a model of covert attention capable of translation-invariant processing of object patterns. Inspired by the biology of the visual system, we introduce a hierarchical structure in which an object pattern, placed at an arbitrary position within a retinal pattern, is routed to a centralized pattern recognizer. Our approach requires to deal with the segmentation problem and the channelling problem. The segmentation problem, treating the question which part of the visual data form objects (see, e.g., [2], is not included in this contribution. Instead we focus on the channelling problem, being the problem of how an object pattern, once segmented, can be translated into a format appropriate for identi cation (cf. [3]). Our model is based on a gating network achieving translationinvariant pattern processing through a sequence of pattern translations. The gating network incorporates a solution to the channelling problem. This contribution is organized as follows. Section 2 shows how our interpretation of the process of covert attention leads to a solution of the channelling problem. In Section 3, we describe the building block of the gating network: the gating lattice, a sparsely-connected neural network capable of performing threefold pattern translations. The results of simulation studies with gating lattices are presented and discussed in Section 4. Then, in Section 5, the gating network is described 1 Computer Science Department, University of Limburg, P.O. Box 616, 6200 MD, Maastricht, The Netherlands and shown to be capable of translation-invariant pattern processing. Finally, Section 6 concludes that our model provides an arti cial realisation of the covert-attention process for active computer vision. 2 COVERT ATTENTION Human observers can make an active spatial selection of visual data in two ways. The most obvious one is through gaze control [4]. Directing the eyes towards an object provides a rough control over what part of the visual environment is taken as input. The second way of selecting visual data is through the process of covert attention. This process allows for a ne-grained spatial selection of visual data (e.g., [5]). To a certain extent, its action may be likened to a searchlight illuminating a contiguous part of the visual scene (i.e., the attended part) [6]. The e ectiveness of the attentional searchlight can be veri ed through Figure 1 (after [7]). Keeping the eyes xed on the \+"-sign in the centre, individual letters can be selected for identi cation at will. (\attention selects for identi cation".)