Continual Learning for Blind Image Quality Assessment.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• The rapid increase in image data and novel distortions challenges existing blind image quality assessment (BIQA) models, particularly their adaptability to subpopulation shifts.
• Current methods of training on combined datasets are not scalable and difficult to update with new data.

Purpose of the Study:

• To formulate continual learning for BIQA, enabling models to learn sequentially from diverse image quality assessment (IQA) datasets.
• To address the limitations of scalability and adaptability in current BIQA training paradigms.

Main Methods:

• Proposed a continual learning framework for BIQA, defining five desiderata and three criteria for prediction accuracy, plasticity, and stability.
• Introduced a method using a shared backbone with a new prediction head per dataset, employing a regularizer to prevent catastrophic forgetting.
• Calculated the overall quality score via a weighted summation of predictions from all heads.

Main Results:

• Demonstrated the effectiveness of the proposed continual learning method for BIQA.
• Showcased significant improvements compared to standard training techniques, both with and without experience replay.
• Validated the model's ability to adapt to new data while retaining knowledge from previous datasets.

Conclusions:

• The proposed continual learning approach offers a scalable and adaptable solution for blind image quality assessment.
• This method effectively mitigates catastrophic forgetting, allowing BIQA models to continuously improve with new data.
• The publicly available code facilitates further research and application of continual learning in image quality assessment.