Dual-Path Deep Fusion Network for Face Image Hallucination.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep learning methods for face hallucination often rely on complex face priors, increasing computational cost.
• Existing methods for super-resolution (SR) can be expensive and laborious due to the need for detailed facial feature representations.

Purpose of the Study:

• To develop a simple yet effective face image super-resolution (SR) method that eliminates the need for additional face priors.
• To improve the efficiency and quality of super-resolved face images.

Main Methods:

• A dual-path deep fusion network (DPDFN) was proposed, comprising a global memory subnetwork (GMN) and a local reinforcement subnetwork (LRN).
• GMN uses recurrent dense residual learning for wide-range context excavation, capturing holistic facial shape.
• LRN focuses on patch-wise mapping relations for local facial components, learning from low-resolution (LR) to high-resolution (HR) space.

Main Results:

• The DPDFN successfully generates high-quality face images by fusing global and local facial information.
• Experimental results on face hallucination demonstrated superior visual effects compared to state-of-the-art methods.
• Performance on face recognition tasks using VGGface and SCFace datasets also showed significant improvements.

Conclusions:

• The proposed DPDFN offers a computationally efficient and effective solution for face image super-resolution.
• The network's ability to learn global and local facial features without explicit priors advances the field.
• DPDFN achieves superior performance in both visual quality and objective metrics for face hallucination and recognition tasks.