Face Hallucination With Finishing Touches

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Generating high-quality frontal face images from low-resolution, non-frontal images is crucial for facial analysis.
• Existing methods often focus on either super-resolution or frontalization, not both simultaneously for unconstrained images.

Purpose of the Study:

• To present VividGAN, a novel generative adversarial network (GAN) designed for simultaneous super-resolution and frontalization of tiny, non-frontal face images.
• To enable seamless enhancement of daily-life, unconstrained face images for improved facial analysis.

Main Methods:

• VividGAN employs a two-level approach with coarse-level and fine-level Face Hallucination Networks (FHnets) and two discriminators (Coarse-D and Fine-D).
• The coarse-level FHnet generates a basic frontal high-resolution (HR) face, while the fine-level FHnet refines details using facial component priors and a facial component-aware module.
• Two discriminators ensure both global facial structure and fine-grained details are accurately reconstructed.

Main Results:

• VividGAN successfully generates photo-realistic frontal high-resolution (HR) face images from low-resolution (LR) non-frontal inputs.
• The method demonstrates superior performance in downstream tasks, including face recognition and expression classification, compared to state-of-the-art approaches.

Conclusions:

• VividGAN offers a powerful solution for simultaneously super-resolving and frontalizing unconstrained, low-quality face images.
• The proposed approach significantly enhances the quality of facial images, leading to improved performance in critical facial analysis applications.