R2GDN: RepGhost based residual dense network for image super-resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Existing super-resolution networks face challenges with computational complexity and memory usage.
• Deployment on edge computing devices is hindered by resource limitations.

Purpose of the Study:

• Introduce a novel lightweight image super-resolution reconstruction network.
• Mitigate computational complexity and memory consumption issues.
• Optimize network architecture for edge computing environments.

Main Methods:

• Developed a lightweight reparameterization layer for efficient feature utilization.
• Designed the RGAB module for deep feature extraction, retaining dense connections and local residual learning.
• Implemented feature reuse and structural reparameterization techniques.

Main Results:

• The R2GDN network shows significant reductions in model parameters (approx. 95%) and improved inference speed (86.8%) on edge devices compared to performance-oriented methods.
• Outperforms lightweight super-resolution algorithms with a lower parameter count and a 0.74% SSIM improvement on the BSD100 dataset for 4x reconstruction.
• Demonstrates an effective balance between network performance and complexity.

Conclusions:

• R2GDN offers a computationally efficient and effective solution for image super-resolution on edge devices.
• The proposed architecture successfully addresses the trade-off between performance and resource constraints.
• R2GDN represents a significant advancement in lightweight deep learning models for image reconstruction.