Blind light field image quality assessment based on deep meta-learning.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Deep convolutional neural networks (DCNNs) are popular for light field image quality assessment (LFIQA) but require extensive data.
• Existing DCNN methods struggle with long-range dependencies and small-sample datasets common in LFIQA.
• Data augmentation in DCNNs for LFIQA yields unsatisfactory performance.

Purpose of the Study:

• To develop an efficient LFIQA metric that overcomes data limitations and captures spatial-angular information.
• To leverage meta-learning for effective small-sample learning in LFIQA tasks.
• To enhance the performance of LFIQA metrics by utilizing self-attention mechanisms.

Main Methods:

• Utilized the Swin Transformer's self-attention for capturing spatial-angular information in light field images.
• Employed meta-learning to acquire shared prior knowledge across diverse distortion types in LFIQA.
• Fine-tuned a quality prior model on specific LFIQA tasks for rapid model development.

Main Results:

• The proposed LFIQA metric demonstrated high consistency with subjective quality scores.
• The method significantly outperformed several existing state-of-the-art LFIQA approaches.
• Achieved robust performance in small-sample LFIQA scenarios.

Conclusions:

• The combination of Swin Transformers and meta-learning offers a powerful solution for LFIQA, especially with limited data.
• This approach effectively addresses the limitations of traditional DCNN-based methods in LFIQA.
• The proposed metric provides a reliable and efficient tool for assessing light field image quality.