A Multi-Fidelity Data Fusion Approach Based on Semi-Supervised Learning for Image Super-Resolution in Data-Scarce...

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Traditional image super-resolution (SR) methods require extensive paired low- and high-resolution (LR-HR) data.
• Data scarcity, distribution inconsistencies, and missing high-frequency details challenge the generalization and robustness of existing SR techniques.
• Developing SR methods that perform well with limited high-fidelity data is crucial for practical applications.

Purpose of the Study:

• To propose a novel semi-supervised learning-driven multi-fidelity fusion (SSLMF) method for image reconstruction.
• To reduce the reliance on high-fidelity data in image super-resolution tasks.
• To enhance data efficiency and reconstruction quality in data-scarce scenarios.

Main Methods:

• Integration of multi-fidelity data fusion (MFDF) for global structural constraints and information compensation using low-fidelity data.
• Application of semi-supervised learning (SSL) to minimize dependence on labeled high-resolution (HR) samples by utilizing abundant unlabeled multi-fidelity data.
• Validation on benchmark functions and image reconstruction tasks to assess performance with limited high-fidelity samples.

Main Results:

• SSLMF effectively models linear and nonlinear relationships across multi-fidelity data.
• The proposed framework demonstrates high performance and improved data efficiency even with a limited number of high-fidelity samples.
• Successful application in image reconstruction and a case study in audio restoration highlights cross-disciplinary potential.

Conclusions:

• SSLMF offers a robust solution for image reconstruction, particularly in data-scarce environments.
• The method significantly enhances reconstruction quality and data efficiency compared to traditional approaches.
• SSLMF presents a novel and efficient approach for image super-resolution with limited high-fidelity data.