s-LWSR: Super Lightweight Super-Resolution Network.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Deep learning models excel at single image super-resolution (SISR) but require substantial parameters and computational complexity.
• High computational demands limit the application of current SISR models on resource-constrained devices like mobile phones.

Purpose of the Study:

• To propose a flexibly adjustable, super lightweight super-resolution network (s-LWSR) for efficient image enhancement.
• To overcome the limitations of high computational complexity in existing deep learning-based SISR methods.

Main Methods:

• Designed a high-efficient U-shape based block with an information pool for effective multi-level feature abstraction.
• Employed a compression mechanism utilizing depth-wise separable convolution to reduce parameters with minimal performance loss.
• Optimized the network by removing specific activation layers to preserve more information and enhance final performance.

Main Results:

• The proposed s-LWSR network demonstrates efficient feature extraction and parameter reduction.
• Experiments confirm that s-LWSR achieves performance comparable to more complex deep learning super-resolution (DL-SR) methods.
• The network operates with limited parameters and computational operations, validating its lightweight design.

Conclusions:

• s-LWSR offers an effective solution for single image super-resolution on devices with limited computational and storage resources.
• The proposed architectural innovations enable significant reductions in model complexity without sacrificing performance.
• This work contributes a practical and efficient deep learning model for real-world image enhancement applications.