Efficient Image Super-Resolution via Self-Calibrated Feature Fuse.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning-based super-resolution (SR) networks offer convincing results but require significant computational resources.
• High computational demands and large parameter counts limit the application of SR technology on devices with limited computing power.

Purpose of the Study:

• To propose an efficient image super-resolution network (SCFFN) that balances performance and parameter count.
• To develop a novel self-calibrated feature fuse block (SCFFB) for recovering high-frequency image details.
• To accelerate network training and reduce computational complexity using an attention mechanism (U-SCA).

Main Methods:

• The proposed SCFFN utilizes a self-calibrated feature fuse block (SCFFB) for feature self-transformation and self-fusion.
• An attention mechanism, termed U-SCA, is integrated into the reconstruction part to enhance efficiency.
• The SCFFB aims to recover high-frequency details effectively while reducing computational burden compared to transposed convolution.

Main Results:

• Experimental results on public datasets demonstrate that SCFFN achieves comparable performance to existing SR networks.
• SCFFN requires significantly fewer parameters and computational resources than conventional methods.
• The U-SCA attention mechanism effectively reduces computational complexity without compromising reconstruction quality.

Conclusions:

• The proposed SCFFN offers an efficient solution for image super-resolution, suitable for resource-constrained devices.
• SCFFB and U-SCA are effective components for improving SR network efficiency and detail recovery.
• This work contributes to making advanced SR technology more accessible and practical.