Learning multiple linear mappings for efficient single image super-resolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Learning-based super-resolution (SR) methods can restore high-resolution (HR) images from low-resolution (LR) inputs.
• Existing popular SR algorithms are often computationally intensive, limiting their real-world applicability.

Purpose of the Study:

• To propose a novel, computationally efficient single image SR method.
• To address the limitations of time- and space-intensive SR algorithms.

Main Methods:

• A multiple linear mappings (MLM) approach is proposed to transform LR feature subspaces into HR subspaces.
• LR images are partitioned into linear subspaces, with LR and HR subdictionaries learned based on shared representation coefficients.
• A fast nonlocal means algorithm is employed for similarity-based regularization to reduce artifacts.

Main Results:

• The proposed MLM-based SR method demonstrates superior quantitative and qualitative performance compared to existing application-oriented SR methods.
• The approach achieves fast and stable SR recovery.
• The method maintains relatively low time and space complexity.

Conclusions:

• The novel MLM-based SR method offers an efficient and effective solution for image super-resolution.
• The technique provides a practical alternative for resource-limited settings requiring high-resolution image restoration.
• Combining MLM with nonlocal means regularization enhances SR quality while preserving computational efficiency.