Structure-Preserving Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Generative adversarial networks (GANs) have advanced single image super-resolution (SISR) by producing realistic details.
• Existing GAN-based SISR methods often struggle with structural distortions in recovered images.

Purpose of the Study:

• To develop a structure-preserving super-resolution (SPSR) method that mitigates structural distortions while retaining GAN-based realism.
• To enhance the ability of SISR models to accurately recover geometric structures.

Main Methods:

• Proposed SPSR with gradient guidance (SPSR-G), restoring high-resolution gradient maps and using a gradient loss for structural guidance.
• Introduced a learnable neural structure extractor (NSE) trained with self-supervised methods (contrastive prediction, jigsaw puzzles) for richer structure extraction.
• Developed model-agnostic methods applicable to existing super-resolution networks.

Main Results:

• SPSR-G and the NSE-enhanced SPSR demonstrated superior performance over state-of-the-art methods on five benchmark datasets.
• Achieved improved metrics including LPIPS, PSNR, and SSIM.
• Visual results confirmed enhanced structural integrity and natural detail generation in super-resolved images.

Conclusions:

• The proposed SPSR methods effectively preserve structures while generating perceptually pleasing details in single image super-resolution.
• Gradient guidance and learnable structure extraction offer significant improvements for GAN-based SISR.
• The model-agnostic nature allows broad applicability to existing SISR frameworks.