Aggregative Adversarial Network for Still-to-Video Face Recognition
Jin Wei, Chen Ying · 2020 5th International Conference on Computer and Communication Systems (ICCCS) · 2020
Aiming to tackle the problems of blur, noise, pose variations and occlusion in still-to-video face recognition, a novel aggregative adversarial network is proposed to aggregate each sequence of video frames into a still frontal image with high-quality and without occlusion. The network is composed of three modules. The aggregation module and the discriminator module form a competitive relationship which is capable of improving visual quality of synthesized images. The discriminator module distinguishes the synthesized images true or fake, and meanwhile predict its identity in adversarial ways. A perceptual loss combining with the recognition loss is introduced in the recognition module to guarantee the synthesized images produce discriminative features for recognition. Experimental results prove that the recognition accuracy and computational complexity of the proposed network is significantly superior to the state-of-the-arts.