A Dual-Layer Model of High-Level Perception

The human visual system is capable of identifying objects because it uses several kinds of information: low-level features and high-level schemata are combined to extract data and identify objects. [1,3] Here, we present a dual-layer model of high-level perception, which learns both basic features and high-level schemata, combining this information into a hierarchy. [2] We have developed a two-level system to recognise faces from images. Our model is built from two layers of Support Vector Machine (SVM) classi ers. The rst layer comprises four feature-detection experts, which in-dependently identify and label components within the image. A feature map is constructed, which describes the structure of the object. A second layer is used to classify the high-level feature-map representation. A connectionist autoassociator identies likely feature maps from partial feature maps, and so predicts the location of missing features. [4] Based on a test set of 405 images, our model achieved a success rate of 80%. The dual-layer model adds the use of a feature map and autoassociato