Purpose: Visual backward masking is a method employed in many psychophysical experiments. We examined the physiological basis of backward masking by measuring magnetic fields at the scalp of subjects while they were peforming a scene recognition task. Methods: A match-to-sample task was used to evaluate recognition performance for briefly presented and masked natural images. On each trial, a digitized image was displayed on a projection screen for 37 ms or 92 ms. The target was immediately followed by a random pattern mask which was displayed for 500 ms. In two other condtions either only the mask or only the target (unmasked) were displayed. In the subsequent query phase the target was shown together with a distractor image. The subjects had to indicate the target image (2AFC). During presentation, evoked magnetic fields were recorded with a CTF 151 channel whole-cortex MEG-system. Results: Reducing presentation duration from 92 ms to 37 ms degraded recognition performance from 97% to 67% correct. The MEG data revealed that for the unmasked and mask-only conditions the activity in the first 80-120 ms was concentrated in the channels over occipital cortex. For the long target duration (92 ms), the mask had no effect on the initial occipital activity caused by the target. However, for the short target duration, occipital processing of the mask briefly interfered with processing of the target. Conclusions: The results indicate that during the first 40 ms of visual processing of a natural scene new information arriving in the early visual areas can lead to a profound degradation of recognition performance, correlating with the temporal overlap of target and mask signals in occipital cortex. Later processing stages, beyond 180 ms, seem to be unaffected by the mask.