Novelty Scene Detection Using Scan Path Topology and Energy Signature in Scaled Saliency Map

We propose a biologically motivated novelty scene detection model that can indicate novelty of the scenes having affine transformed field of views in dynamic environment. Novelty detection can play an important role as an intrinsic motivation, which makes an autonomous mental development model try to obtain new information and adapt to changing environment. The proposed model has been developed based on understanding the functions of the brain organ such as the hippocampus for novelty detection in the limbic system as well as the visual path way from the retina to the visual cortex through the lateral geniculate neucleus(LGN). In the proposed model, we represent an input scene using a topology of a visual scan path obtained from a scaled saliency map generated by a visual attention model. In order to indicate novelty of a current input scene, the topology of a current input scene is compared with that of a previously experienced scene. In addition, the energy signature with scale information in the corresponding visual scan path is also considered to decide novelty of the input scene. The computer experimental results show that the proposed model successfully indicates novelty for natural color input scenes in dynamic visual environment. Keywords— Novelty scene detection, scaled saliency map, hippocampus, bottom-up and top-down attention model, adaptive resonance theory network