VCRNet

VCRNet performs no-reference image quality assessment by restoring severely distorted images with a non-adversarial visual restoration network and estimating image quality from multi-level restoration features.

Key Features:

• Non-Adversarial Model: Utilizes a non-adversarial approach to manage restoration of severely distorted images, overcoming challenges faced by GAN-based methods.
• Visual Restoration Network: Restores distorted images and incorporates a visual compensation module and optimized asymmetric residual block to enhance restoration capability.
• Visual Compensation Module: Improves the relationship between distorted and restored images to support accurate quality reconstruction.
• Optimized Asymmetric Residual Block: Enhances the restoration network's ability to handle complex distortions.
• Error Map-Based Mixed Loss Function: Refines restoration by focusing optimization on error minimization via an error map-based mixed loss.
• Quality Estimation Network: Estimates image quality by leveraging multi-level restoration features derived from the visual restoration network.
• Multi-Level Restoration Features: Extracts features from multiple restoration levels to enable precise no-reference quality estimation, particularly for severely degraded images.
• Validation on IQA Databases: Demonstrated image quality prediction performance across seven representative IQA databases.

Scientific Applications:

• No-Reference Image Quality Assessment (NR-IQA): Predicts perceived image quality without requiring an undistorted reference image.
• Restoration for Severely Degraded Images: Restores and assesses images with severe distortions where GAN-based methods struggle.
• Benchmarking in IQA Studies: Provides a framework for evaluating NR-IQA performance across multiple IQA databases.

Methodology:

Employs a non-adversarial architecture composed of two components—a Visual Restoration Network (with a visual compensation module and optimized asymmetric residual blocks) trained with an error map-based mixed loss—and a Quality Estimation Network that predicts image quality from multi-level restoration features.