No-Reference Image Quality Assessment by Hallucinating Pristine Features.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• No-reference image quality assessment (NR-IQA) is crucial for evaluating visual experiences without pristine originals.
• Existing NR-IQA methods often struggle with generalization across diverse image distortions and datasets.

Purpose of the Study:

• To propose a novel NR-IQA method leveraging feature-level pseudo-reference (PR) hallucination.
• To enhance the characterization of visual quality by exploiting perceptually meaningful features and natural image statistics.
• To achieve accurate and generalizable image quality predictions.

Main Methods:

• A feature-level PR hallucination framework is developed.
• Mutual learning with pristine references is used to train PR features from distorted images.
• Triplet constraints are employed to ensure discriminative PR features.
• An invertible neural layer facilitates feature disentanglement for quality prediction.

Main Results:

• The proposed method demonstrates superior performance on four popular IQA databases.
• Cross-database evaluation confirms the high generalization capability of the method.
• The approach effectively predicts image quality using hallucinated pseudo-reference features.

Conclusions:

• The feature-level PR hallucination approach offers an effective solution for NR-IQA.
• The method exhibits robust performance and excellent generalization across different image datasets.
• This work advances the field of objective image quality assessment.