A lightweight large receptive field network LrfSR for image super-resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) excel at single-image super-resolution (SISR).
• Existing CNN-based SISR methods often have high computational costs, large parameter counts, and significant latency.
• These limitations hinder the deployment of SISR on devices with constrained computational resources.

Purpose of the Study:

• To propose a lightweight network for image super-resolution (SISR) that overcomes the limitations of existing methods.
• To enhance the capture of pixel-to-pixel relationships and multi-scale information through a large receptive field.
• To improve image super-resolution quality using efficient attention mechanisms while minimizing network parameters.

Main Methods:

• Designed a lightweight large receptive field network for image super-resolution (LrfSR).
• Introduced an information distillation module (LrfDM) utilizing dilated convolutions to achieve a large receptive field.
• Incorporated efficient attention mechanisms (ECCA and SESA) to improve super-resolution performance.

Main Results:

• The LrfSR model achieved competitive Peak Signal-to-Noise Ratio (PSNR) values across multiple benchmark datasets (Set5, Set14, B100, Urban100, Manga109).
• Demonstrated superior performance compared to existing models like LKDN.
• Showcased effective extraction of high-frequency features and successful fusion of multi-scale information.

Conclusions:

• The LrfSR model effectively balances high-quality image reconstruction with limited computational resources.
• Explored the potential of large receptive fields in lightweight SISR networks.
• Provides a viable solution for deploying advanced super-resolution on resource-constrained devices.