Multi-scale feature selection network for lightweight image super-resolution.

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Super-resolution (SR) methods using deep convolutional neural networks (CNNs) offer high performance but require substantial computational resources.
• This limits their application on resource-constrained, low-budget devices.

Purpose of the Study:

• To develop a novel, lightweight super-resolution network (MFSN) suitable for real-world low-budget devices.
• To improve feature extraction and fusion for enhanced SR performance in a computationally efficient manner.

Main Methods:

• Proposed a Multi-Scale Feature Selection Network (MFSN) with a core Multi-Scale Feature Selection Block (MFSB).
• MFSB utilizes a coarse-to-fine receptive field strategy and wide-activated residual units.
• Integrated a Scale Selection Module (SSM) with adaptive receptive field adjustment and a Comprehensive Channel Attention Mechanism (CCAM) for dynamic feature fusion.

Main Results:

• The proposed MFSN demonstrated superior performance compared to existing lightweight SR methods.
• The MFSB effectively extracts multi-scale features, and the CCAM enhances feature representation through dynamic channel weighting.

Conclusions:

• The MFSN offers an effective and efficient solution for super-resolution tasks on devices with limited resources.
• The proposed MFSB and CCAM contribute to improved performance in lightweight deep learning models for computer vision.