Expanding Training Data for Facial Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• High-quality training data is crucial for learning-based facial image super-resolution (SR).
• The performance of SR models heavily relies on the similarity between training and testing data.
• Existing methods lack strategies for optimizing training datasets for specific facial SR tasks.

Purpose of the Study:

• To propose a novel approach for expanding training datasets to improve facial image super-resolution.
• To enhance the similarity between training data and specific testing inputs.
• To develop a method applicable to both patch-based and global facial SR techniques.

Main Methods:

• Proposed three constraints: local structure, correspondence, and similarity constraints.
• Generated new training data by expanding local patches with varying parameters.
• Applied these constraints to create a more relevant training set for facial SR.

Main Results:

• Demonstrated significant improvements in facial image super-resolution quality.
• Validated the effectiveness of the proposed training data expansion method on benchmark and real-world images.
• Showcased the applicability of the expanded data to diverse facial SR approaches.

Conclusions:

• Training data expansion is an effective strategy for enhancing facial image SR.
• The proposed constrained data generation method yields superior SR results.
• This approach offers a versatile solution for improving facial SR across different methodologies.