Data-driven Face Hallucination by Inverse Degradation Neural Network

Face hallucination refers to the technology that inferring its potential corresponding high-resolution (HR) image from the input low-resolution (LR) facial image. At present, most face hallucination algorithms improve reconstruction performance by optimizing models. However, the common approach will out of operation when meeting more complex problem, etc, the input image contains degraded pixels (noise), their reconstruction performance will drop sharply. In order to solve the problem, we propose an inverse degradation neural network (IDNN), which can mine the essential features of the images under data-driven. In this network, we design different network structures in different task stages. Firstly, the more accurated face structure is generated by the denoising network in the LR space. But the details from the face image is lacked in this stage. In order to further enhance the face image details, we utilize the reconstruction network to restore the missing details. The experimental results on FEI face database show that IDNN outperforms some state-of-the-art approaches in subjective and objective measures.