Automatic Detection of Demosaicing Image Artifacts and Its Use in Tampering Detection

摘要

Even novice people can nowadays do convincing forged images. However, most forgeries alter the underlying statistics of the image, and in particular the slight artifacts caused by the demosaicing method. Demosaicing transforms the undersampled image acquired by the CFA into a three channel color image. Two problems arise: The first one is to identify the underlying CFA configuration, which classifies pixels according to whether they acquired a red, green or blue value. The second one is to detect anomalies in the regularity of the “demosaicing artifacts” that may reveal tampered image regions. We review the state of the art of detection methods, but point out that none of the proposed methods yields guaranteed detections for the CFA configuration or for the tampered areas. The methods generally yield an output that must still be evaluated visually. We therefore introduce an a contrario method that yields guaranteed detections, namely detections with a very low number of false alarms (NFA). Obtaining such an NFA is a useful complement to existing detection methods and should enable these methods to be included into automatic image evaluation processes.