s-LWSR

s-LWSR implements a lightweight deep-learning single image super-resolution (SISR) network that reduces computational complexity and parameter count to enable high-resolution image reconstruction on resource-constrained devices such as mobile and portable medical imaging systems.

Key Features:

• High-Efficiency U-Shape Based Block: Employs a high-efficient U-shape based block that constructs an information pool and integrates multi-level data from early processing stages for comprehensive feature abstraction.
• Compression Mechanism (depth-wise separable convolution): Uses depth-wise separable convolution to minimize parameter count while retaining model capacity.
• Optimized Activation Layers: Removes selected activation layers to preserve information within the model and reduce information loss during processing.
• Computational Efficiency for Resource-Constrained Devices: Reduces computational complexity and resource demands to enable deployment on devices with limited computing power.
• Adjustable Model Complexity: Provides flexible adjustment of network complexity to balance computational cost and super-resolution output quality.
• Comparable Performance to DL-SR: Achieves performance levels comparable to more resource-intensive deep learning super-resolution (DL-SR) methods while using fewer resources.

Scientific Applications:

• Single Image Super-Resolution (SISR): Enhances spatial resolution of low-resolution images for downstream analysis in imaging studies.
• Mobile Bioinformatics Applications: Enables high-resolution image reconstruction within mobile bioinformatics workflows under device constraints.
• Remote Sensing: Improves resolution of remotely acquired imagery where onboard computational resources are limited.
• Medical Imaging on Portable Devices: Facilitates super-resolution reconstruction for medical imaging modalities deployed on portable or bedside devices.

Methodology:

Constructs an information pool via a high-efficient U-shape based block to integrate multi-level features, applies depth-wise separable convolution for compression, removes select activation layers to retain information, and allows flexible adjustment of network complexity to balance efficiency and output quality.