Blind image quality assessment via deep learning

Area of Science:

• Computer Vision
• Artificial Intelligence
• Psychology

Background:

• Human image quality evaluation is often qualitative, not numerical.
• Existing learning-based image quality assessment (IQA) models convert qualitative descriptions to numerical scores, losing information.
• This irreversible conversion limits the accuracy of current IQA models.

Purpose of the Study:

• To develop a blind IQA model that learns qualitative evaluations directly.
• To output numerical scores for fair benchmarking of IQA metrics.
• To improve the accuracy and robustness of image quality assessment.

Main Methods:

• Images are represented using natural scene statistics features.
• A discriminative deep model classifies features into five qualitative grades (excellent, good, fair, poor, bad).
• A novel quality pooling method converts qualitative labels into numerical scores.

Main Results:

• The proposed classification framework is more natural than regression-based models.
• The model demonstrates robustness, particularly with small sample sizes.
• Experiments on popular databases verify the model's effectiveness and efficiency.

Conclusions:

• The developed blind IQA model provides a more accurate and robust method for evaluating visual quality.
• This approach aligns better with human qualitative judgment in image quality assessment.
• The model facilitates fair comparison and utilization of IQA metrics.