Learning Decoupled Features With Perceptual Distillation for Blind Image Quality Assessment

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Current Blind Image Quality Assessment (BIQA) models struggle with weak supervision due to the complexity of image distortions and semantics.
• Subjective scores, used as optimization targets, represent overall quality but fail to capture diverse perceptual cues effectively.

Purpose of the Study:

• To develop a novel framework for BIQA that disentangles content-aware and distortion-aware features.
• To improve the accuracy and robustness of image quality assessment models.

Main Methods:

• Proposed a Decoupled Feature Learning (DFL) framework utilizing global-local input pairs to decompose entangled features.
• Implemented a perceptual knowledge distillation strategy with a Just-Noticeable-Difference (JND) model for feature transfer.
• Introduced a local distortion-guided attention module to integrate decoupled perceptual features.

Main Results:

• The DFL framework achieved superior performance over state-of-the-art methods on eight benchmark datasets.
• Demonstrated the framework's flexibility in enhancing the perceptual capabilities of other Transformer variants.
• The proposed approach effectively learns compact global content-aware and local distortion-aware features.

Conclusions:

• The DFL framework offers a robust solution for BIQA by effectively learning disentangled perceptual features.
• The proposed methods significantly advance the field of image quality assessment.
• The framework's adaptability suggests broad applicability in related computer vision tasks.