A Document Presentation Attack Detection Scheme with Optical Flow under a Flashlight

The presentation attack of a document image can be launched by simple printing/displaying and capturing operations. It has posed a serious threat to the existing document authentication system. However, the generalization performance of existing document presentation attack detection (DPAD) methods is not satisfactory over document images with diverse contents and qualities. In this work, we propose a novel DPAD method that exploits a new data modality observable in the video recorded under a flashlight, named the optical Flow-based Authentication Network (F-AutheNet). Our F-AutheNet extracts both the spatial and optical flow information from video frames. The spatial data carries information on low-level image features, while the optical flow data shows the variations of reflective light under different materials. The features extracted from both data modalities are fused through a cross-modal focal loss. In the experiment, we collect a comprehensive student card video database (SCVD) under a flashlight, comprising more than 18K genuine, printer-captured, and screen-recaptured video frames. Experimental results show that, by incorporating features extracted from the optical flow data, the generalization performance of generic DPAD backbones has improved by a large margin under some challenging experimental protocols. Specifically, the average EER metrics have been reduced by 2.73 and 2.84 percentage points under cross-domain experiments covering samples of different capturing distances (i.e., qualities) and contents, respectively.