Task-Specific Normalization for Continual Learning of Blind Image Quality Models.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Blind Image Quality Assessment (BIQA) is crucial for evaluating image fidelity without reference images.
• Continual learning methods aim to adapt models to new tasks without forgetting previous ones, a challenge in BIQA.
• Existing BIQA methods struggle with adapting to new datasets while maintaining performance on previously learned tasks.

Purpose of the Study:

• To develop a simple and effective continual learning method for BIQA.
• To improve quality prediction accuracy and the plasticity-stability trade-off in BIQA models.
• To enhance robustness against variations in task order and length.

Main Methods:

• A novel continual learning approach for BIQA is proposed.
• Key technique involves freezing pre-trained deep neural network (DNN) convolution filters for stability.
• Task-specific normalization parameters are learned for plasticity, with each new IQA dataset assigned a prediction head.
• A lightweight K-means gating mechanism is used for weighted summation of predictions from all heads.

Main Results:

• The proposed method demonstrates improved quality prediction accuracy.
• It achieves a better plasticity-stability trade-off compared to previous techniques.
• The approach shows robustness to different task orders and lengths.
• Experiments on six IQA datasets validate the method's advantages.

Conclusions:

• The presented method offers a simple yet effective solution for continual learning in BIQA.
• Freezing DNN filters and learning task-specific normalization parameters is a successful strategy.
• The technique provides enhanced accuracy, stability, and robustness for adaptive BIQA systems.