VCRNet: Visual Compensation Restoration Network for No-Reference Image Quality Assessment.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Generative Adversarial Networks (GANs) guided by the free-energy principle have advanced no-reference image quality assessment (NR-IQA).
• However, GANs struggle with image restoration tasks, particularly for severely degraded images, hindering accurate quality reconstruction modeling.
• Existing methods fail to establish a robust relationship between distorted and restored image quality.

Purpose of the Study:

• To propose a novel Visual Compensation Restoration Network (VCRNet) for NR-IQA.
• To address the limitations of GAN-based methods in handling severely destroyed images.
• To accurately build the quality reconstruction relationship between distorted and restored images.

Main Methods:

• VCRNet employs a non-adversarial model for efficient distorted image restoration.
• It integrates a visual restoration network and a quality estimation network.
• Key components include a visual compensation module, optimized asymmetric residual blocks, and an error map-based mixed loss function to enhance restoration capabilities and utilize multi-level restoration features for quality estimation.

Main Results:

• The proposed VCRNet demonstrates superior performance in image quality prediction.
• Experimental results on seven benchmark IQA databases confirm state-of-the-art accuracy.
• The method effectively handles the restoration task for severely degraded images.

Conclusions:

• VCRNet offers an effective alternative to GAN-based NR-IQA methods, particularly for challenging restoration tasks.
• The network accurately models the relationship between image distortion and perceived quality.
• The developed method achieves leading performance in NR-IQA.