Foveated analysis of image features at fixations

Abstract

Analysis of the statistics of image features at observers' gaze can provide insights into the mechanisms of fixation selection in humans. Using a foveated analysis framework, in which image patches were analyzed at the resolution corresponding to their eccentricity from the prior fixation, we studied the statistics of four low-level local image features: luminance, RMS contrast, and bandpass outputs of both luminance and contrast, and discovered that the image patches around human fixations had, on average, higher values of each of these features at all eccentricities than the image patches selected at random. Bandpass contrast showed the greatest difference between human and random fixations, followed by bandpass luminance, RMS contrast, and luminance. An eccentricity-based analysis showed that shorter saccades were more likely to land on patches with higher values of these features. Compared to a full-resolution analysis, foveation produced an increased difference between human and random patch ensembles for contrast and its higher-order statistics.

Distractor effect and saccade amplitudes: Further evidence on different modes of processing in free exploration of visual images

“…Furthermore, one could argue that repetitive fixations (refixations) explain our findings. Regions of higher visual complexity are usually fixated with higher frequencies, longer durations, and the preceding saccades often are relatively short (Rajashekar, van der Linde, Bovik, & Cormack, 2007). Even if refixations are similar to focal fixations, this hypothesis alone does not explain the current findings, for instance, the differences of effects for expanding and shrinking distractors (Experiment 2).…”

Influence of affective image content on subjective quality assessment

“…Mean subjective quality ratings averaged over stimulus images and participants for each of the five objective quality settings (JPEG compression level, q1-q5), shown in .30), respectively, highlighting that participants did not employ the entire subjective quality range (1)(2)(3)(4)(5)(6)(7)(8)(9)(10), but that the mean subjective quality rating does increase monotonically as the objective quality setting rises, confirming participants' generally successful discrimination between objective quality settings. This is exemplified by a significant "large" [43] correlation between objective quality setting and mean subjective quality assessment [r 0.54 (2548), ptwo-tailed < 0.001].…”

“…Fidelity-based objective quality assessment methods typically rely upon relatively unsophisticated numerical metrics, such as mean absolute error (MAE), mean square error (MSE), peak signal to noise ratio (PSNR), and linear correlation coefficient (LOC), among others [5], and are fast and easy to compute [6], but since they often correlate poorly with human responses [4,7] they are of limited utility where the ultimate receiver is the human visual system (HVS). More specifically, not all numerically equivalent image degradations are equally noticeable [3], and not all image regions enjoy equal attention [8]. Conversely, PVQMs may employ a model of, or derive inspiration from, the sensory computations performed in the early HVS [9,10] or utilize psychophysically derived knowledge of visual performance.…”

Face saliency in various human visual saliency models

“…Observers unconsciously tend to select central locations of the image in order to catch the potentially most important visual information [4]. Human vision relies extensively on the ability to make saccadic eye movements to orient the high-acuity fovea region of the eye over the targets of interest in the visual scene [14,18].…”