DETECTABILITY OF DEGRADED VISUAL SIGNALS: A BASIS FOR EVALUATING IMAGE-RETRIEVAL PROGRAMS

Abstract

We report here the first experiments in a long-range program for investigating the effectiveness of image-retrieval or image-enhancement procedures. We employed a signal-detection mode of observer response, and our stimuli were computer-generated, pointed, abstract forms that we call quadrigons. Four values of signal-to-noise ratio were provided by varying the amount of roundedness of the interior and exterior angles of these forms. Linear blur, grain magnification, and figure-surround contrast ratio were the other independent variables in our factorial design. For each quadrigon, observers gave scaled expressions of confidence that the photographed object was originally pointed (the "signal present" condition). Scores from 12 observers, who were each exposed twice to a set of 500 quadrigons, yielded receiver operating characteristics (ROC functions) that are sensitive, quantitative indicators of the discriminabilities of the stimuli. By this method, also, an observer's criterion state (his degree of willingness to guess that a signal was present) was evaluated and removed as a contaminating factor. Signal-to-noise ratio had a strong and systematic effect upon signal detection accuracy when the effects of all other variables were combined. Linear blur, grain size, and contrast ratio each affected observer performance greatly. All two-way, three-way, and four-way interactions among the independent variables were highly significant; each source of image degradation had differential effects upon every other variable, and combinations of the variables had additional effects. The results are complex, but they provide useful implications for image processing that is designed to enhance information retrieval. We can, for instance, define the effects upon signal detectability when linear blur is reduced by specified amounts, or grain size is reduced, or contrast conditions are improved. We can, moreover, specify certain optimum combinations of values for these variables.