Modeling Stimulus-driven Attentional Selection In Dynamicnatural Scenes

摘要

In this paper we have developed a neuromorphic model of bottom-up (BU) visual attentional selection. The output of a recently developed neuromorphic multi-channel retina model has represented the input of our model. As a first step, a saliency map has been calculated for each retinal channel which, next, has been integrated into a master saliency map. Model parameters have been optimized based on human eye movement data measured during viewing dynamic natural scenes. We have tested two different strategies for weighting the channel-specific saliency maps during integration into a master map. In the first case, channel weights have been kept constant throughout the verification measurements, whereas, in the other case, they have been updated on each frame, according to the specific properties of the visual input. Surprisingly, the constant channel weighting strategies have performed better than the continually updated ones. We have measured the model's accuracy by defining the hit ratio (concurrence) between the first few predicted locations (the most salient locations) and the measured fixation locations. Constant weighting methods have achieved ～74% hit ratio on four predictions. For a comparison, the accidental chance for this case has been less than 20%. This pure BU approach has performed surprisingly well on dynamic natural input. Some practical applications have already been made with task-dependent simplifications.