Attention-Driven Graph Neural Network for Deep Face Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning, particularly convolutional neural networks (CNNs), has advanced face super-resolution (FSR).
• Existing FSR methods often overlook non-local correlations within face images, limiting performance.
• There is a need for improved feature extraction and relation modeling in FSR.

Purpose of the Study:

• To introduce a novel attention-driven graph neural network (AD-GNN) for enhanced face super-resolution.
• To improve discriminative feature extraction and model feature relationships by addressing limitations in current FSR approaches.
• To reconstruct facial details more effectively by leveraging non-local information.

Main Methods:

• Developed an end-to-end trainable attention-driven graph neural network (AD-GNN).
• Introduced a cross-scale dynamic graph (CDG) block to capture relationships between distant image patches across scales.
• Incorporated channel attention and spatial dynamic graph (CASDG) blocks, featuring a spatial-aware dynamic graph (SDG) unit for non-local self-similarity exploration.

Main Results:

• The AD-GNN effectively reconstructs facial details by utilizing information from similar, spatially remote patches.
• The method leverages structural information of faces for improved super-resolution.
• Extensive experiments on public benchmarks confirmed the superiority of AD-GNN over state-of-the-art FSR methods.

Conclusions:

• The proposed AD-GNN significantly advances face super-resolution capabilities.
• Modeling non-local correlations and spatial self-similarity is crucial for high-quality FSR.
• AD-GNN offers a promising direction for future research in image super-resolution and computer vision.