We present an approach for the detection of crease in banknotes using 3D image processing. A line-scan camera equipped with two lenses implements a stereo vision system acquiring images of banknotes transported perpendicular to the stereo baseline. Depth data is obtained from stereo matching and local curvature features are calculated from the estimated 3D banknote surface model. Mean and Gaussian curvature measures are calculated for each banknote pixel and median values are used to characterize whole banknotes with respect to their fitness for circulation. The scale at which the curvature description provides most meaningful information with respect to the banknote fitness grade is identified by pyramidal decomposition of the depth map. The approach is validated based on banknotes with varying crease which were labeled by experts into several fitness grades.