A Visual Attention Based Object Detection Model beyond Top-Down and Bottom-up Mechanism

Abstract

Traditional saliency-based attention theory supposed that bottom-up and top-down factors combine to direct attentional behavior. This dichotomy fails to explain a growing number of cases in which neither bottom-up nor top-down can account for strong selection biases. Thus, the top-down versus bottom-up dichotomy is an inadequate taxonomy of attentional control. In our previous study, we presented a general computational framework for detecting task-oriented salient objects in images beyond top-down and bottom-up mechanism. It possesses three parts: selection history, current goal and physical salience. Selection history is integrated with current goal and physical salience to compose an integrative framework. In this extended version, our ameliorated model is applied to face and car detection and simulates taskdependent reasonable eye trajectories (visual scan paths). Experimental results demonstrate that selection history (reward) is shown to influence saccade trajectories. Our findings support the idea that attention and gaze can be directed voluntarily to regions of interest (by selection history and current goal) and can be captured by local features of an object that stand out from the background (by physical salience). IntroductionVisual search plays a key role in our everyday activities; the visual system pays attention to the salient objects for efficient search. Visual saliency plays important roles in natural vision in that saliency can direct eye movements, deploy attention, and facilitate tasks like object detection and scene understanding. Attention determines which among multiple competing stimuli are represented in the brain. . Saliency models have been developed for eye fixation prediction and salient object detection. The former focuses on identifying a few fixation locations on natural images, which is important for understanding human attention. The latter, also called salient object segmentation, is used to accurately detect where the salient object should be, which is useful for many high-level vision tasks [24] [6]. Recently, the theoretical dichotomy of attentional control between top-down and bottom-up is challenged. The dichotomy fails to explain a growing number of cases in which neither bottom-up nor top-down can account for strong selection biases [4]. Thus, the top-down versus bottom-up dichotomy is an inadequate taxonomy of attentional control. A wealth of research has demonstrated that reward information provides a third source of input to the attention system. When a large reward is received, attention is strongly primed to select the rewarded target [12] [10].Awh et al. [4] proposed selection history (including two classes of 'history' effects, i.e., selection and reward history) as a third category of control by explicitly distinguishing current goals from selection history effects. A 'priority map' which they still highlighted integrates three distinct categories of selection bias: the observer's current selection goals, selection history, and physical