Cascaded Hallucination-Classification Deep Network for Low-Resolution Face Recognition in the Wild

Abstract

Low-resolution face recognition (LR FR) has become an active research subarea due to its significances for real applications. Conventional low-resolution face recognition approaches meet challenges like noise affection and lack of effective features with LR faces. In this paper, we propose a deep learning method for LR FR. Our convolutional neural network (CNN) model directly learns an end-to-end classification on LR faces. Different from normal CNN for high-resolution (HR) face recognition, ours integrates a lightweight hallucination network mapping LR images into HR ones. Furthermore, we concatenate the hallucination and classification networks so that the training propagation is operated in one model, which largely boosts the performance over basic CNN and separate two-step models. Besides, our model is robust to varying poses and illuminations in the wild, and also portable to embedded system for its memoryand energy-saving features.