A general lightweight image super-resolution with sharpening enhancement and double attention network

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning-based single image super-resolution (SISR) achieves high performance but suffers from significant computational and storage costs.
• A need exists for lightweight SISR networks that maintain effectiveness while reducing resource demands for practical applications.

Purpose of the Study:

• To develop a general lightweight SISR network, named ESDAN (Enhancement and Sharpening with Dual Attention Network), that optimizes the trade-off between model complexity and performance.
• To introduce novel modules that enhance feature representation and reconstruction capabilities.

Main Methods:

• The proposed ESDAN network incorporates a Sharpening Enhancement Module (SEM) and a Dual Attention Upsampling module (DAU).
• SEM integrates Attention-Driven Feature Sharpening (ADFS) and Multi-Way Feature Enhancement (MWFE) to improve feature contrast and information reinforcement.
• DAU dynamically fuses shallow and deep features to boost reconstruction accuracy.

Main Results:

• Extensive experiments show ESDAN outperforms current state-of-the-art lightweight SISR methods.
• ESDAN demonstrates high versatility and effectiveness in various SISR-related tasks, including medical imaging (Alzheimer's disease brain MRI), stereo endoscopic images, and surveillance imagery.
• The source code is publicly available.

Conclusions:

• ESDAN offers an effective solution for lightweight single image super-resolution, balancing computational efficiency with high-quality image reconstruction.
• The network's design and demonstrated versatility suggest its broad applicability in diverse image super-resolution challenges.