Reparameterizable Multibranch Bottleneck Network for Lightweight Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Deep convolutional neural networks (CNNs) for single image super-resolution (SISR) face computational challenges on edge devices.
• Existing lightweight SR networks often struggle with oversmoothed results and high parameter counts.

Purpose of the Study:

• To propose a lightweight SR network for efficient deployment on edge computing devices.
• To improve the quality of reconstructed images by addressing oversmoothing and enhancing structural details.

Main Methods:

• Introduced a reparameterizable multibranch bottleneck module (RMBM) with bottleneck residual block (BRB), inverted bottleneck residual block (IBRB), and expand-squeeze convolution block (ESB).
• Developed a novel peak-structure-edge (PSE) loss function to improve image structure similarity and reduce oversmoothing.
• Optimized and deployed the network on edge devices with Rockchip Neural Processor Unit (RKNPU) for real-time performance.

Main Results:

• The proposed network achieved a model size of 98.1K, significantly outperforming advanced lightweight SR networks.
• Demonstrated superior objective evaluation metrics and subjective visual quality on natural and remote sensing image datasets.
• Enabled real-time SR reconstruction on edge devices without additional computational cost during inference.

Conclusions:

• The proposed lightweight SR network effectively balances performance and computational cost for edge deployment.
• The RMBM and PSE loss are key innovations for achieving high-quality, real-time image super-resolution.
• The network shows great potential for applications requiring efficient image enhancement on resource-constrained devices.