Scale Invariant Domain Generalization Image Recapture Detection

Recapturing and rebroadcasting of images are common attack methods in insurance frauds face identification spoofing, an increasing number detection techniques were introduced to handle this problem. However, most them ignored the domain generalization scenario scale variances, with inferior performance on shift situations, normally exacerbated by intra-domain inter-domain variances. In paper, we propose a alignment framework (SADG) address these challenges. First, adversarial discriminator is exploited minimize discrepancies image representation distributions among different domains. Meanwhile, exploit triplet loss as local constraint achieve clearer decision boundary. Moreover, global relationship regularization force representations same class across scales be undistinguishable. Experimental results four databases comparison state-of-the-art approaches show that better can achieved using our framework.