Objective Integration of local image components into global shapes is crucial for perceiving coherent objects in visual scenes.Previous studies indicate that V1 is involved in mediating contour integration—a delayed component of neuronal responses in V1 is predictive of perceptual saliency of contours.The contour-related responses are subject to the influences of attentional control and perceptual learning, which suggests task-dependent gating of contextual interactions in V 1 via feedback connections from higher cortical areas.Methods To test this speculation directly, we chronically implanted Utah arrays into retinotopically matched regions in macaque V1 and V4, and trained the animal to detect visual contours formed by collinear bars embedded in a background of randomly oriented bars.Results By comparing neuronal activity recorded simultaneously in these two areas in response to the same contour stimuli in the detection task, we could replicate earlier findings in V1 (Li et al., Neuron 2006, 2008)—neuronal responses that were strongly correlated with the length of the global contours did not appear until 100 ms after stimulus onset, about 60 ms longer than the response latencies of V1 neurons to the stimuli.The contour activated neuronal responses in V4, however, showed a different time pattern.At the response outset, which was on average about 15 ms later than that in V1, neurons in V4 already carried information about contours of different lengths, although the contour-related responses were weaker than those seen in the late components of both V1 and V4 responses.Moreover, as the contour-related responses evolved with time in these two cortical areas, contour activated responses reached a maximum strength first in V4, about 15 ms earlier than V1.Conclusion Our observations suggest that V4 detects the presence of a global configuration of contours earlier than V1, and that the countercurrent processing via feedforward and feedback connections between lower and higher cortical areas greatly amplifies the contour information.This processing scheme is in agreement with a computational model showing simultaneous encoding of objects and their parts at different levels of detail through bottom-up and top-down interactions (Epshtein et al., PNAS 2008).