Unsupervised domain adaptation for disguised face recognition

Disguised face recognition (DFR) is an extremely challenging task due to the numerous variations that can be introduced with different disguises. Most existing disguised face recognition approaches follow a supervised learning framework. However, due to the domain shift problem, the Convolutional Neural Networks (CNN) model trained on one dataset often fail to generalize well to another dataset. In our attempt, we formulate the DFR as an unsupervised learning problem and propose a unified deep learning architecture Unsupervised Domain Adaptation Model (UDAM) with three merits. Firstly, UDAM is a unified deep architecture, containing a Domain Style Adaptation subNet (DSN) and an Attention Learning subNet (ALN), which jointly learn from end-to-end. Secondly, DSN is a well-design generative adversarial network which simultaneously translate the labeled image from the source to the target domain in an unsupervised manner and maintain the ID label after translation. Thirdly, ALN is a Convolutional Neural Network (CNN) for disguised face recognition with our proposed attention transfer strategy. Extensive experiments using Simple and Complex Face Disguise Dataset and the IIIT-Delhi Disguise Version 1 Face Database have demonstrated that the proposed method yields a consistent and competitive performance for disguised face recognition.