These analyses test the hypothesis that contour integration is distinct from simple, non-random pattern recognition, and that it can be studied fruitfully using a novel combination of methodologies. Real, human observers and ideal observers perform a classification task in which the two stimulus patterns are systematically corrupted by noise to determine the effect of that noise on observers' responses. This technique, response classification, is used to determine how observers use the available information in an array of line elements to discriminate between an aligned contour and a non-random (orthogonal) pattern. The comparison between performance of several ideal observers and the human observers reveals that template-matching models might not account entirely for the human observers' responses. Beyond showing that human observers perform the task using the local relations between neighboring elements, we also determine the circumstances under which grouped pairs of elements are integrated perceptually into more extended contours. These circumstances are related theoretically to co-occurrence statistics derived from natural images between pairs of small line elements using the parameters of distance, orientation difference, and direction. The complete set of analyses further our understanding of how humans perceive visual contour.