Abstract
Face image synthesis has advanced rapidly in recent years. However, similar success has not been witnessed in related areas such as face single image super-resolution (SISR). The performance of SISR on real-world low-quality face images remains unsatisfactory. In this paper, we demonstrate how to advance the state-of-the-art in face SISR by leveraging style-based generator in unsupervised settings. For real-world low-resolution (LR) face images, we propose a novel unsupervised learning approach by combining style-based generator with relativistic discriminator. With a carefully designed training strategy, we demonstrate our converges faster and better suppresses artifacts than Bulat’s approach. When trained on an ensemble of high-quality datasets (CelebA, AFLW, LS3D-W, and VGGFace2), we report significant visual quality improvements over other competing methods especially for real-world low-quality face images such as those in Widerface. Additionally, we have verified that both our unsupervised approaches are capable of improving the matching performance of widely used face recognition systems such as OpenFace.
Highlights
With recent advancements in deep learning algorithms [1], Single Image Superresolution (SISR) has seen a significant advance in performance in terms of objective metrics like peak signal-to-noise-ratio (PSNR)
We will talk about generative adversarial networks (GAN) and show that adding a discriminator can significantly improve the visual quality of the superresolved image [2]
The invention of GAN opens the door to construct a whole new class of powerful generative models which have found numerous applications in low-level vision including SISR, face image synthesis, and style transfer. 2.2.2 Single image super-resolution using generative adversarial networks In SRGAN [2], the authors showed that using a GAN-based architecture for the task of single image super-resolution leads to noticeable improvements in terms of subjective visual quality despite the sacrifice on traditional objective quality metric such as PSNR
Summary
With recent advancements in deep learning algorithms [1], Single Image Superresolution (SISR) has seen a significant advance in performance in terms of objective metrics like peak signal-to-noise-ratio (PSNR). Most of the existing deep learning algorithms for solving SISR problems are categorized as supervised; in that they rely on paired high-resolution (HR) and low-resolution (LR) images to optimize the neural network weights. The HR images are downsampled using algorithms (like bicubic downsampling) to create the corresponding LR ones These artificially created LR data deviate significantly from the complex real word degradation model and with that a rapid decrease in performance is observed when neural networks trained on artificial LR that are tested on real-world LR images [3]. To make our solution work for real-world LR face images, we borrow ideas from recent advances in style transfer [4] and image synthesis [5]. We will first review some related works including convolutional neural network (CNN), generative adversarial networks (GAN), image synthesis, and style transfer. We will show that both our approaches outperform previous state-of-the-art ones
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
