Comparison of Full-Reference Image Quality Models for Optimization of Image Processing Systems.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Objective image quality assessment (IQA) models are typically validated against human judgments.
• Existing perceptual datasets are heavily used, risking overfitting and limiting benchmark effectiveness.

Purpose of the Study:

• To evaluate IQA models as optimization objectives for image processing algorithms.
• To compare the performance of eleven full-reference IQA models across various low-level vision tasks.

Main Methods:

• Training deep neural networks using eleven full-reference IQA models as objectives.
• Applying these networks to four image processing tasks: denoising, deblurring, super-resolution, and compression.
• Conducting subjective testing on the processed images to assess perceptual quality.

Main Results:

• Ranking of IQA models based on their effectiveness in optimizing image processing tasks.
• Identification of relative advantages and disadvantages of different IQA models.
• Empirical data on the perceptual performance of IQA models in practical applications.

Conclusions:

• IQA models can be effectively used as objectives for optimizing image processing.
• Subjective testing provides a robust method for ranking IQA model performance.
• Recommendations for desirable properties for future IQA model development are proposed.