A VLSI Implementation of Visual Based Attentive Processing

1 Introduction Visual attention describes the way that biological vision systems prioritize information to balance the need for rapid processing of complex scenes and the limited amount of processing resources [7][62][69][90]. The attention system enables the sequential selection of parallel subsets of data such that the most complex processing does not have to include all data from the visual field. (The visual field is defined as the area in the physical world that is imaged onto the retinae at any point in time.) Two forms of attentional shifts enable prioritization of information by using two very different mechanisms. The most obvious form of attentional shifts is overt attention. Overt attentional shifts use the movement of the eyes to control which objects in the visual field are imaged onto the highest density of photoreceptors in the retina known as the fovea. Overt attentional shifts are controlled by the saccadic [6][78] and smooth pursuit [67] oculomotor systems, which coordinate fast point-to-point eye movements and smooth tracking movements. The second form of attentional shifts is covert attention. Covert attentional shifts occur regardless of eye movements and are evidenced by the ability of the primate visual system to attend to objects in the periphery of the visual field [69]. The covert attentional shifts enable prioritization of information through neural mechanisms that route specific signals to the higher processing centers in the brain [89]. Attentive processing in biological systems has inspired the development of artificial active-vision systems that use a sequential search method to process subsets of data in the imaging array. A number of engineers have applied models of the primate saccadic and smooth pursuit systems to robotics in order to create autonomous machines that perform specialized tasks such as target tracking, path planning, or pattern recognition [13][15][16][26]. Most of these systems use general-purpose serial-processing hardware that differs greatly from the highly-collective, parallel architectures of biological systems. Traditionally, the designers of active-vision systems have not made attempts to understand the underlying processing mechanisms in biological systems even though many of the performance goals for the two types of systems are the same. When an application requires real-time visual processing, the computational capabilities within engineered systems are severely limited because of the serial processing architecture (in most cases) and other competing constraints such as size, power, and cost. The goal of understanding how attentive processing is performed within the primate visual system is shared by …