Contrast detection thresholds are known to increase with background co
ntrast, a phenomenon called contrast masking. We found that, under som
e conditions, observers improved their masked detection performance by
repetitive practice of a masking experiment. This learning effect res
ulted in a cancellation of suprathreshold contrast masking within the
contrast range measured. A two-alternative forced-choice discriminatio
n paradigm was used, with stimuli consisting of Gabor signals as maske
rs and target, presented at the same location and time. Untrained obse
rvers showed increased detection thresholds with increasing mask contr
ast for suprathreshold mask contrasts, but perceptual learning caused
an elimination of this classical effect, with masked thresholds reachi
ng the no-mask level and below. Learning did not decrease, but rather
somewhat increased, discrimination thresholds when target and mask sha
red the same Gabor signal parameters. Performance improvement was foun
d to be specific for orientation and mask configurations, though it di
d transfer between mirror symmetric mask configurations and between ey
es. These results argue against a static transducer function-based acc
ount for contrast masking and are consistent with a theory assuming mu
ltiple feature-based interactive network capable of long-term gain mod
ifications. (C) 1997 Elsevier Science Ltd.