Lightweight Single Image Super-Resolution with Selective Channel Processing Network.

Area of Science:

• Computer Vision
• Deep Learning
• Image Restoration

Background:

• State-of-the-art single image super-resolution (SR) methods often have high computational costs due to numerous parameters.
• Existing SR networks require significant computational resources during inference, limiting their practical application.

Purpose of the Study:

• To develop a more lightweight and resource-friendly SR network.
• To reduce parameter count and computational consumption in deep learning-based SR.

Main Methods:

• Introduced a selective channel processing network (SCPN) utilizing a selective channel processing module (SCPM).
• SCPM dynamically learns channel significance during training for efficient inference.
• Proposed a differential channel attention (DCA) block to enhance high-frequency information restoration.

Main Results:

• The SCPN model achieved superior results on natural and remote-sensing image super-resolution benchmarks.
• The method demonstrated a slim model size with fewer than 1 million parameters.
• Achieved a favorable trade-off between computational cost and performance.

Conclusions:

• The SCPN model offers an efficient solution for image super-resolution with reduced computational requirements.
• The proposed SCPM and DCA modules contribute to improved performance and efficiency.
• The methodology shows potential for extension to other computer vision tasks.