Adversarial Reconstruction CNN for Illumination-Robust Frontal Face Image Recovery and Recognition

Liping Yang, Bin Yang, Xiaohua Gu

OAI: oai:igi-global.com:268847 • DOI: 10.4018/IJCINI.20210401.oa2

This article proposes an adversarial reconstruction convolution neural network (ARCNN) for non-uniform illumination frontal face image recovery and recognition. The proposed ARCNN includes a reconstruction network and a discriminative network. The authors employ GAN framework to learn the reconstruction network in an adversarial manner. This article integrates gradient loss and perceptual loss terms, which are able to preserve the detailed and spatial structure image information, into the overall reconstruction loss function to constraint the reconstruction procedure. Experiments are conducted on the typical illumination-sensitive dataset, extended YaleB dataset. The reconstructed results demonstrate that the proposed ARCNN approach can remove the illumination and shadow information and recover natural uniform illuminated face image from non-uniform illuminated ones. Face recognition results on the extended YaleB dataset demonstrate that the proposed ARCNN reconstruction procedure can also preserve the discriminative information of face image for classification task.

Cognitive Informatics Computer Science and Information Technology Cognitive Informatics