Perceptual no-reference image quality assessment with meta-learning by graph representation learning and multi-scale...

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• No-reference image quality assessment (NR-IQA) aims to predict image perceptual quality without a reference image.
• Existing NR-IQA methods struggle with long-range dependencies and content fidelity.
• Challenges include diverse distortion types, varying distortion levels, and preserving image content.

Purpose of the Study:

• To develop a perceptually-driven NR-IQA framework addressing limitations of existing methods.
• To improve the accurate prediction of perceptual quality aligned with the human visual system (HVS).
• To enhance the extraction of distortion-aware features and model hierarchical relationships in image distortions.

Main Methods:

• Utilized a meta-learning paradigm to pre-train a self-calibrated convolutional backbone for adaptive feature extraction.
• Introduced a graph representation learning module with a graph convolutional network to encode relationships between distortions and content.
• Employed dual supervision from a distortion-type discriminator and a distortion-level regressor for joint optimization.

Main Results:

• Achieved significant performance improvements on four benchmark datasets, with average SROCC and PLCC gains of 3.6-36.6% over traditional methods.
• Demonstrated consistent performance gains over existing deep learning-based NR-IQA approaches.
• Ablation studies confirmed the effectiveness of proposed components in creating discriminative and generalizable distortion representations.

Conclusions:

• The proposed NR-IQA framework effectively captures complex image distortions and their hierarchical relationships.
• The method closely mirrors human perceptual judgments, offering superior image quality assessment.
• The integration of meta-learning and graph representation learning provides a robust and generalizable solution for NR-IQA.