Face Hallucination by Attentive Sequence Optimization with Reinforcement Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Face hallucination is a specialized super-resolution task aiming to create high-resolution (HR) face images from low-resolution (LR) inputs.
• Traditional patch-wise methods process facial regions independently, neglecting global image interdependencies.

Purpose of the Study:

• To develop a novel attention-aware face hallucination (Attention-FH) framework using deep reinforcement learning.
• To improve face super-resolution by exploiting global facial interdependencies and recurrently attending to image patches.

Main Methods:

• Implemented a deep reinforcement learning approach with an attention-aware framework (Attention-FH).
• Developed a recurrent policy network for dynamic region selection and a local enhancement network for patch hallucination and state updating.
• Jointly trained both networks by maximizing a long-term reward reflecting the overall HR image quality.

Main Results:

• The Attention-FH framework demonstrated superior performance compared to state-of-the-art methods.
• Achieved significant improvements on in-the-wild face images, even with substantial pose and illumination variations.

Conclusions:

• The proposed Attention-FH framework effectively enhances face hallucination by leveraging attention mechanisms and global context.
• This approach offers a significant advancement in generating high-quality HR faces from LR inputs, particularly in challenging real-world scenarios.