A flexible high-accuracy calibrating method is proposed for projector-camera systems. To reduce the interferences between the projected pattern and printed pattern, a pure white pattern is projected on the printed pattern board as the background image, and then a background subtraction process is applied. Through analysing the remaining noises, a Bayesian classifier-based adaptive filter is proposed to eliminate them and maximally reserve the information around corners with a minimum error probability to obtain accurate corner correspondences. Without the need for any additional high-precision devices, this method only uses a checkerboard calibration board. The experiment results show that the maximum reprojection error is <;0.4561 pixels and the root mean square of the reprojection error is 0.1422 pixels.