Application of human visual perception paradigm for automated inspection of monochromatic security documents

A digital inspection machine has been developed to detect and evaluate printing flaws in U.S. currency using a methodology extensible to general graphical images. A flaw detection model was compiled from previous research by others, derived from a panel of experienced judges who rated a set of flawed exemplars from 1 to 10 according to the perceived noticeability of the flaws found on each note. The flaw model relates a numerical grade to the location, area and contrast of printing defects. The objective grading scale is a significant addition to previous currency inspection devices. The inspection machine acquires a high-resolution image of a test note and subtracts a reference image from it to isolate the print flaws. There are two causes of large amplitude errors in the resulting difference image: (1) photometric errors caused by acceptable fluctuations in illumination and scene reflectance, and (2) spatial errors arising from normal variations in the printed pattern of the reference and test samples. Photometric errors magnify the flaw amplitude in the difference image, while spatial errors create topological artifacts related to surface structure. Photometric errors are mitigated by intensity normalization, which redistributes the grey levels in the test image to match a defined standard reflectance. (This technique does not change the feature distribution in the test image; the flaws are unchanged by normalization.) Spatial errors are caused by positional misalignment between the test and reference notes. A subpixel spatial normalization algorithm was devised to align the two images. The misalignment is complex because the samples may have subtle distortions, imperceptible to the eye, due to mechanical stresses from the printing process. The algorithm fits an elliptic paraboloid surface to the cross-correlation coefficients about a prominent topological feature common to both images. The fractional displacement of a feature in the test image relative to the reference is determined from the location of the surface peak. After the test image is normalized photometrically and spatially with the reference image, the absolute difference of the two images is taken. The difference image contains only those features of the test note not found on the reference. To further reduce spatial errors, a structural filter is developed to control detection sensitivity in the image on a point-by-point basis. The sensitivity coefficients are weighting factors at each point and range from 0 to 1.0. The coefficients are determined during calibration from the measured variations in a set of known acceptable notes. In order to detect low-contrast flaws, the noise floor in the difference in\age must be less than the flaw magnitude. The registration and filter combination reduces the noise floor of the difference image to an acceptable level. Consequently, flaws isolated in the difference image are unaffected by normal variations in illumination, position, orientation, and printing conditions. Finally, the difference image is screened by size to separate actual printing flaws from any remaining artifacts using connectivity analysis. The size, contrast, and location of the flaws is measured, and the flaw severity grade representing the test note quality is calculated from the flaw parameters.