Frontal face reconstruction based on detail identification, variable scale self-attention and flexible skip connection

Reconstruction of the frontal face from the profile is of great significance for face recognition in complex scenes. The existing mainstream methods of face reconstruction, such as FF-GAN, CAPG-GAN, TP-GAN, etc., have made good progresses on improving the generator network, but fewer considerations on the identification of face details and the extraction of spatial context features. To address the problem, we propose the frontal face reconstruction based on the detail discrimination, variable scale self attention, and flexible skip connection (FR-DVF): designing a group of discriminators for multi-scale detail region identification, a novel encoder-decoder generator structure with a variable scale type of self-attention module, which inserts a max-pooling layer into the pathways of the traditional module to reduce its feature-dimension and computing-cost, and a flexible type of the skip-connections to alleviate the stiff property of the traditional connections between the encoder and decoder layers. After adding detail discrimination, variable scale self attention module, and flexible skip connection structure, the rank-1 recognition rate (\(\%\)) of DVF-FR in the database of M2FPA increased by 2.94, 1.93 and 1.67\(\%\), respectively, as well as that occurred in FERET.