DFVNet: Real-Time Disguised Face Verification

A face is arguably the most conspicuous biometric for recognizing a person, yet its identification can be an abstruse task if the subject tries to impersonate someone else using different disguises. In this work, we present DFVNet, a deep learning framework to help identify disguised faces in real-time with greater accuracy. The proposed architecture employs three neural networks to achieve its goal: the first one does real-time image matting on the given image dataset to extract the image foreground without the use of a trimap at 63 FPS, the second one uses a spatial fusion convolutional network to extract 18 facial key-points whose heatmap prognosis are coincided using optical flow, which are then processed and their relative angles used as input in the third one i.e. the deep neural network for subject identification. Our method mitigates the anomalies induced by the disguises, outperforming the previously used methods for the same.