Learning Dual Memory Dictionaries for Blind Face Restoration

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Blind face restoration is challenging due to complex, unknown degradations.
• Existing methods often separate generic and specific restoration, limiting performance and applicability.
• A unified model for both generic and specific restoration is highly desirable.

Purpose of the Study:

• To develop a unified model for blind face restoration that adaptively handles generic and specific restoration scenarios.
• To improve the photo-realistic performance of blind face restoration using explicit feature memorization.
• To introduce a new high-quality dataset for high-resolution specific face restoration.

Main Methods:

• Proposed DMDNet (Dual Dictionary Network) that explicitly memorizes generic and specific facial features.
• Utilized a generic dictionary for general facial priors and a specific dictionary for identity-specific features.
• Implemented a dictionary transform module for feature fusion and leveraged multi-scale dictionaries for coarse-to-fine restoration.

Main Results:

• DMDNet achieved state-of-the-art performance in both quantitative and qualitative evaluations.
• The model generated more photo-realistic results on real-world low-quality images.
• Introduced CelebRef-HQ, a new high-quality dataset for high-resolution face restoration research.

Conclusions:

• DMDNet offers an effective and unified approach to blind face restoration.
• The dual dictionary mechanism successfully integrates generic and identity-specific facial details.
• The proposed method demonstrates significant improvements in restoring degraded faces, with potential for broad applications.