Statistical analysis and selection of visual fixations

By using a variable resolution sampling of the visual field, the human visual system (HVS) has evolved to an efficient imaging system that allows for a wide field of view without the accompanying data glut. A direct consequence of this multi-resolution perception of the visual scene is the active nature in which the human eyes interact with the visual environment to gather information efficiently. The HVS has conceivably evolved multiple mechanisms for controlling gaze. The interplay of high-level cognitive and low-level image features influences eye movements in many intricate ways and makes the problem of selecting gaze a formidable task. Since mechanisms that require relatively little image interpretation are likely to be relevant in artificial vision, the goal of this dissertation is to develop a bottom-up, image-based theory of human eye movements. Analysis of stimuli at the observer’s point of gaze can provide an understanding of strategies used by observers in visual tasks. In this dissertation, bottom-up, low-level image features that influence gaze are revealed by directly measuring the statistics of image patches at the point of gaze of observers in visual search (involving the search for a target) and visual surveillance (a free-viewing scenario). I demonstrate that structural cues used by observers in visual search tasks can indeed be revealed by analyzing the stimulus at the locations of the observer’s eye fixations. In particular, using the classification image analysis technique, I demonstrate that observers use idiosyncratic, target-dependent structural cues to deploy their fixations as they searched for simple geometric targets embedded at very low signal-to-noise ratios in noise stimuli that had the spectral characteristics of natural images. I show that even in the absence of an explicit visual task, observers do not cast their gaze about randomly, but instead select image patches that have higher luminance, contrast, and a stronger center-surround influence than patches that are selected randomly. A foveated framework is presented to analyze the statistics of image patches at the ‘next’ fixation point at the spatial resolution at which the observer perceived the patch. Examples of selecting fixations in novel scenes using these features are presented.Electrical and Computer Engineerin