Active Fine-Tuning From gMAD Examples Improves Blind Image Quality Assessment.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep neural networks (DNNs) have advanced image quality assessment (IQA).
• Existing DNN-based IQA models show high correlation but can be vulnerable to adversarial examples in group maximum differentiation (gMAD) competitions.
• Blind image quality assessment (BIQA) methods specifically assess image quality without a reference image.

Purpose of the Study:

• To improve the robustness and generalizability of DNN-based BIQA models.
• To leverage gMAD examples to identify and rectify weaknesses in BIQA models.
• To develop an active learning strategy for refining BIQA models using targeted image examples.

Main Methods:

• Pre-trained a DNN-based BIQA model using noisy annotators and synthetic distortions.
• Identified gMAD examples by comparing the baseline BIQA model with full-reference IQA methods.
• Collected human quality annotations for selected gMAD images in a controlled lab setting.
• Iteratively fine-tuned the BIQA model using gMAD-derived human ratings and existing databases.

Main Results:

• Demonstrated the feasibility of the active learning scheme on a large-scale unlabeled image set.
• The fine-tuned BIQA model achieved improved generalizability in gMAD.
• Performance on previously seen databases was maintained without degradation.
• gMAD examples effectively revealed model weaknesses and guided refinement.

Conclusions:

• gMAD examples are valuable for improving BIQA methods by highlighting vulnerabilities.
• The proposed active learning approach enhances BIQA model robustness and generalizability.
• Iterative refinement using gMAD and human feedback offers a promising direction for future BIQA research.