Disk features of positive or negative polarity were superimposed on images of stationary, Gaussian, uncorrelated noise. In one detection task, the three observers rated the likelihood that specified image locations contained a feature of known polarity rather than the uniform noise background. In a second detection task, the same observers rated the likelihood that each specified location contained a feature of unknown polarity (dark or bright), and then indicated which was the more likely polarity, assuming that a feature was present. In the third task (polarity discrimination) observers rated the likelihood that the feature, known to be present in each location, was positive rather than negative in polarity. Independent samples of images varied the feature's contrast to manipulate the observers performance within each task. An index of the observer's detection or discrimination accuracy, obtained from the measured ROC curve for each observer, task, and condition, was compared to the calculated value of the same index for the realized cross-correlator. With noisy images of this type, the cross-correlator is an "ideal observer" that yields a physically optimal decision variable. In all three tasks, the observers' performance indices were closely proportional to the cross-correlator's (slope about 0.57). This indicates an' "observer efficiency" similar to the levels we measured in tasks using CT images, for which the cross correlator is suboptimal. Observers' detection of the positive-contrast and negative-contrast features did not differ, in either the known-- or unknown--polarity tasks.