The Application of Visual Saliency Models in Objective Image Quality Assessment: A Statistical Evaluation

Area of Science:

• Computer Vision
• Image Processing
• Human-Computer Interaction

Background:

• Objective image quality assessment (IQA) models aim to predict human perception of image quality.
• Integrating visual attention, specifically saliency models, into IQA shows potential for improved reliability.
• Current understanding of how saliency model performance impacts IQA metric reliability is limited.

Purpose of the Study:

• To statistically evaluate the added value of computational saliency in objective IQA.
• To investigate the relationship between saliency model performance and their contribution to IQA metrics.
• To provide guidance on selecting and applying saliency models for IQA.

Main Methods:

• Utilized 20 state-of-the-art computational saliency models.
• Employed 12 well-established image quality metrics (IQMs).
• Conducted an exhaustive statistical evaluation of saliency model integration into IQMs.

Main Results:

• Significant performance gains in IQMs were observed when incorporating saliency models.
• The predictive accuracy of saliency models for human fixations did not directly correlate with their benefit to IQM performance.
• Saliency model dependence, IQM dependence, and image distortion type significantly influence performance gains.

Conclusions:

• Computational saliency offers demonstrable benefits for objective IQA.
• The effectiveness of saliency models in IQA is complex and not solely dependent on their fixation prediction accuracy.
• The study provides crucial insights and a public testbed for advancing saliency-based IQA.