PROJECT TITLE :

Face Frontalization Using an Appearance-Flow-Based Convolutional Neural Network

ABSTRACT:

Face Recognition (FR) is a difficult problem because of the wide range of facial expressions. Non-frontal faces can be transformed into frontal ones so that FR is possible. Facial pixel values alter when faces are rotated. As a result, current CNN-based algorithms are able to synthesise frontal faces in colour space. As a result, the synthetic frontal faces lose subtle facial textures because of this non-linear learning difficulty in a colour space. Nonfrontal-frontal pixel changes are mostly induced by geometric modifications (rotation, translation, etc.) in space, according to our findings in this work. To ease the learning process, we intend to learn the nonfrontal-frontal facial conversion in the spatial domain rather than the colour domain. As a result, we present a convolutional neural network for face frontalization based on appearance-flow (A3F-CNN). Non-frontal and frontal faces have a deep relationship that A3F-CNN learns to establish. Frontal faces are created by intentionally transferring pixels from the non-frontal one once the correlation is established. In this approach, the synthetic frontal faces can retain the delicate textures of the face. An appearance-flow-guided learning technique is presented to promote training convergence. Face mirroring is used to deal with the self-occlusion problem, as is generative adversarial network loss to produce a more lifelike face. Face synthesis and pose-invariant FR are subjected to extensive testing. When it comes to creating photo-realistic faces in both controlled and uncontrolled lighting settings, our method outperforms all other methods. Furthermore, our FR performance on the Multi-PIE, LFW, and IJB-A databases is quite competitive.