Learning to attend: Visually detecting objects using primitive properties.

Our perception is finite limited by both sensory capacity and the ability to interpret stimulation. Therefore, in order that an environment may be best understood there is a need to be selective in how resources, both computational and sensory, are deployed; this selective mechanism is often termed attention. In this thesis a model of visual attention is described which may be specialised to detect targets of interest. Specifically, we explore how a target's signal may be differentiated from other stimulation in a visual input through developing knowledge of its expected appearance with respect to a basic set of features. A key question which we address is: how the separate sources of information provided by the processing of each feature type may be best be combined to provide a map of target saliency across a scene. This saliency index is then used to guide the deployment of attention. Results show an improvement in targeting efficiency of nearly 50% when compared to a similar model of attention which does not exploit target specific information. Further, we also report favourable results when comparing the performance of our approach to other detection strategies reported within the computer vision literature