Single image super-resolution via Image Quality Assessment-Guided Deep Learning Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep learning (DL) networks have significantly advanced super-resolution (SR) performance.
• Existing DL-based SR methods often struggle to balance perceptual quality and distortion metrics.
• A need exists for SR methods that optimize both visual appeal and fidelity.

Purpose of the Study:

• To propose a novel image quality assessment (IQA)-guided single image super-resolution (SISR) method using DL.
• To achieve an optimal tradeoff between perceptual quality and distortion in SR results.
• To introduce a flexible and robust architecture for SISR.

Main Methods:

• An IQA network is integrated into the DL architecture to extract perception features from SR outputs.
• A fused loss function combines IQA-derived perception loss with traditional pixel-wise loss.
• An interactive training model using a cascaded network addresses heterogeneous datasets, and a pairwise ranking hinge loss handles limited training samples.

Main Results:

• The proposed IQA-guided SISR method demonstrates a superior tradeoff between perceptual quality and distortion compared to existing SISR techniques.
• Extensive experiments validate the effectiveness of the approach.
• The architecture proves adaptable and not limited to specific network models.

Conclusions:

• The developed IQA-guided DL architecture offers a promising approach for SISR.
• The method effectively balances perceptual quality and distortion, enhancing SR image results.
• The proposed architecture provides a flexible foundation for future SISR research.