A Deep Reconstruction CNN for Illumination-robust Face Image Recovery and Recognition
Liping Yang, Bin Feng Yang, Xiaohua Gu · 2018
We propose a deep reconstruction convolution neural network approach for illumination-robust face image recovery and recognition. Our reconstruction approach learns an end-to-end mapping between non-uniform illumination and uniform illumination face images. The architecture of the network is functional divided into four stages, i.e., feature decomposition, multi-scale feature mapping and fusion, nonlinear mapping and reconstruction. We illustrate that the incorporating of multiscale feature mapping and fusion stage in the network significantly remove the unwanted illumination information of face images. Besides, a gradient constraint is also added to the overall loss function of the proposed approach to preserve detailed texture information of face images. Face reconstruction and face recognition experiments are conducted in two typical illumination-sensitive data sets, extended YaieB and eMU-PIE. The reconstructed results demonstrate that the proposed approach can efficiently separate out the unwanted illumination information while keeping the discriminant details of face images. Meanwhile, the proposed approach can significantly enhance the face recognition performances of traditional feature descriptors.