CTCNet: A CNN-Transformer Cooperation Network for Face Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep Convolutional Neural Networks (CNNs) have advanced face super-resolution but face limitations.
• Existing methods struggle with computational cost and fidelity due to limited receptive fields.
• Multi-task learning requires extensive dataset annotation and prior networks increase complexity.

Purpose of the Study:

• To propose an efficient CNN-Transformer Cooperation Network (CTCNet) for superior face super-resolution.
• To address limitations of existing CNN-based methods in fidelity and computational efficiency.
• To improve the restoration of both local facial details and global facial structure.

Main Methods:

• Developed a CNN-Transformer Cooperation Network (CTCNet) with a multi-scale encoder-decoder backbone.
• Introduced a Local-Global Feature Cooperation Module (LGCM) integrating a Facial Structure Attention Unit (FSAU) and Transformer block.
• Designed a Feature Refinement Module (FRM) and a Multi-scale Feature Fusion Unit (MFFU) for enhanced feature processing and fusion.

Main Results:

• CTCNet demonstrated significant improvements in face super-resolution tasks.
• The proposed modules effectively promoted consistency in local detail and global structure restoration.
• Evaluations showed superior performance compared to state-of-the-art methods across various datasets.

Conclusions:

• CTCNet offers an efficient and effective solution for face super-resolution.
• The integration of CNN and Transformer architectures enhances facial image reconstruction quality.
• The novel modules contribute to improved fidelity and naturalness in super-resolved faces.