A Noise Robust Face Hallucination Framework Via Cascaded Model of Deep Convolutional Networks and Manifold Learning

Face hallucination technique generates high-resolution clean faces from low-resolution ones. Traditional technique generates facial features by incorporating manifold structure into patch representation. In recent years, deep learning techniques have achieved great success on the topic. These deep learning based methods can well maintain the middle and low frequency information. However, they still cannot well recover the high-frequency facial features, especially when the input is contaminated by noise. To address this problem, we propose a novel noise robust face hallucination framework via cascaded model of deep convolutional networks and manifold learning. In general, we utilize convolutional network to remove the noise and generate medium and low frequency facial information; then, we further utilize another convolutional network to compensate the lost high frequency with the help of personalized manifold learning method. Experimental results on public dataset show the superiority of our method compared with state-of-the-art methods.

1. Introduction

Facial details of human face are required in many computer vision applications, e.g., person identification in video surveillance systems. However, face images captured by surveillance camera are often noisy and small, leading to the limited information available for reference. To tackle this challenge, face hallucination technique, a domain-specific super-resolution problem aimed at generating high-resolution (HR) face images from low-resolution (LR) inputs has been widely studied recently.