We investigated human sensitivity to vertical mirror symmetry in noise patterns filtered for narrow bands of variable orientations. Sensitivity is defined here as the amount of spatial phase randomization corresponding to 75% correct performance in a 2AFC detection task. In Experiment 1, sensitivity was found to be high for tests patterns of all orientations except those parallel to the axis of symmetry. This implies that corresponding mirror-orientations (e.g. -45 and +45 degrees ) are combined prior to symmetry detection. In Experiment 2, observers detected symmetry in tests of variable orientation in the presence of either non-symmetric or symmetric masks filtered for orientations either parallel or perpendicular to the axis. Observers were found to be primarily affected by masks of the same orientation as the test, thus suggesting that symmetry is computed separately in distinct mirror-orientation channels. In Experiment 3, observers detected a symmetric test of variable height and width embedded in random noise. Data revealed that mirror symmetry is computed over a spatial integration region (IR) that remains approximately constant in area but whose height-to-width aspect ratio changes from 20:1 to 2:1 as orientation is varied from parallel to perpendicular to the axis. We compare human data against that of an ideal observer to identify key factors that limit visual performance and discuss the implications for the functional architecture of symmetry perception. We also propose a multi-channel model of symmetry detection that combines the output of oriented spatial filters in a simple and physiologically plausible manner. Particular emphasis is placed on the notion that changes in the shape of the IR with orientation compensate for changes in information density and partially equate performance across orientations.