People have an extraordinary ability to almost instantly characterize objects by their lightness and color. People have a corresponding ability to characterize surfaces and objects by their visual texture properties. In their NSF-funded research project, Charles Chubb and George Sperling at UC Irvine will use new experimental paradigms and new analytic methods (involving perturbation methods and dimensional analysis) to characterize humans' ability to perceive textures. These procedures have already enabled the investigators to isolate previously unknown texture-perception mechanisms. The current project now applies these procedures to the formal systematic description and characterization of the vast range of human texture-perception mechanisms with the goal of compiling a Table of the Elementary Dimensions of Visual Sensitivity, somewhat analogous to the Periodic Table of Elements.
Understanding the elementary visual mechanisms that enable humans to accurately perceive their visual environment is of fundamental importance and ultimately is likely to have many useful applications. For example, the analysis of visual textures is carried out in the brain, so understanding these visual processes will enable a better characterization of brain injuries that interfere with these (previously unmeasured) processes. Knowing both the range and the limits of human perceptual processes will yield better understanding of what humans can and cannot perceive in images, such as medical X-rays, military camouflage, and computer displays. Indeed, simulating such human sensory abilities has been a critical element in the construction of successful robotic sensing systems.
Overview. This project addressed basic questions concerning human visual sensitivity. The research was based on the assumption that human vision contains a number of different field-capture channels, neural arrays each of which operates continuously in time (like a movie camera) to gauge the distribution across visual space of a particular visual property. This research assumed further that two visual textures spontaneously differ in appearance only if they produce different levels of activation in one or more of these field-capture channels. General questions that motivated this research include: How many field-capture channels exist in human vision? What are the visual properties that these different field-capture channels sense? Can people use top-down attention to combine information from different field-capture channels to heighten sensitivity to some features vs others? The research was especially challenging because no fully adequate methods had been previously developed to answer questions of this sort. Thus, a primary goal of the work was to develop the necessary methods. The three questions above are very broad. Tha actual research was more narrowly focused. The field-capture channels sensitive to grayscale scrambles. One branch sought to analyze the field-capture channels that enable people to discriminate visual textures created by randomly mixing small squares of different brightnesses. Such textures are called grayscale scrambles. Some examples are shown in Fig. 1. This research addressed the following questions: How many field-capture channels in human vision are differentially sensitive to grayscale scrambles? How does each of these field-capture channels respond to different grayscales? In these experiments, participants strove to discriminate different types of grayscale scrambles. New inferential power was achieved by (1) using a task in which participants had to detect the location of a small patch of scramble in a large background of different scramble, and (2) varying the predominant quality defining the target scramble across different conditions to enable participants to use top-down attention to optimize the grayscale filters used in different conditions. The results were well described by a model proposing that human vision has four field-capture channels differentially sensitive to grayscale scrambles: the (previously characterized) blackshot channel, sharply tuned to the blackest grayscales; a (previously unknown) gray-tuned channel whose sensitivity is zero for black rises sharply to maximum sensitivity for grayscales slightly darker than mid-gray then decreases to half-height for brighter grayscales; an up-ramped channel whose sensitivity is zero for black, increases linearly with increasing grayscale reaching a maximum near white; a (complementary) down-ramped channel whose sensitivity is maximal for black, decreases linearly reaching a minimum near white. Estimates of the sensitivity functions of these four field-capture channels for the three participants are plotted in Fig. 2. The discovery of these four field-capture channels is likely to have far-reaching impact in fields ranging from interface design to visual neurophysiology to ophthalmology. The centroid paradigm for analyzing feature-based attention. When a tour guide suggests, "look at the reds in this canvas," he is prompting his listeners to select information from the painting based on a visual feature distributed broadly across space. This is an example of "feature-based attention." One of the main goals of the project was to develop the centroid paradigm, a powerful new experimental method for analyzing feature-based attention. In any centroid experiment, the stimuli are random clouds of some sort of items. Fig. 3 illustrates a trial from an experiment using dots of different grayscales as items. On this trial, the participant is trying to ignore the bright dots and click on the centroid of all the dark dots. The participant faces a blank screen (Fig. 3a) and presses a button to start the trial; then (Fig. 3b) the stimulus cloud is flashed (e.g., for 0.1 sec.) and is followed by a mask (Fig 3c) to obliterate the stimulus from sensory memory; then a blank screen with a cursor appears (Fig. 3d), and the participant moves the cursor (Fig. 3e) to click on the location he/she judges to be the centroid of the target dots. After entering his/her response, the participant receives feedback (Fig. 3f). Fig. 4 shows attention filters achieved by two participants. The lefthand (righthand) plots show the weights exerted by different grayscales on centroid judgments when the task was to click on the dark (bright) dots and ignore the bright (dark) dots. Open circles show filters achieved on trials without distractors. Note that both participants achieve attention filters highly selective for the target dots in each condtion. However, (1) they cannot completely ignore distractor dots, and (2) in striving to ignore distractors, they underweight low contrast target dots. Only 200 trials (around 10 minutes of work) were required to produce the plot in each of these four panels. The centroid paradigm thus provides a very powerful new tool for probing the processes of feature-based attention.
