Perception-Distortion Balanced Super-Resolution: A Multi-Objective Optimization Perspective

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Achieving high perceptual quality and low distortion in super-resolution (SR) is challenging.
• Existing SR methods struggle to balance conflicting objectives like perceptual and reconstruction losses.
• Gradient-based optimizers like Adam face difficulties with contradictory loss functions.

Purpose of the Study:

• To address the perception-distortion trade-off in image super-resolution.
• To develop a novel optimization method that effectively balances competing objectives.
• To improve both perceptual quality and reconstruction fidelity in SR models.

Main Methods:

• Formulated the perception-distortion trade-off as a multi-objective optimization problem.
• Developed a hybrid optimizer integrating a gradient-free evolutionary algorithm (EA) with gradient-based Adam.
• Designed a fusion network to merge an EA-Adam-generated population of models.

Main Results:

• The proposed EA-Adam optimizer effectively balances perception and distortion in SR.
• A population of optimal models with diverse perception-distortion preferences was obtained.
• The fusion network successfully merged models, enhancing the perception-distortion trade-off.
• Experimental results show improved perceptual quality and reconstruction fidelity compared to competitors.

Conclusions:

• The novel EA-Adam optimizer offers a superior approach to balancing perception and distortion in image super-resolution.
• The developed fusion network effectively consolidates diverse model strengths for enhanced SR performance.
• This method provides a promising direction for future research in image restoration tasks.