No-Reference Image Quality Assessment: An Attention Driven Approach

Area of Science:

• Computer Vision
• Image Processing
• Human-Computer Interaction

Background:

• No-reference image quality assessment (NR-IQA) predicts image quality without a pristine reference.
• Human visual system (HVS) perception involves predicting pristine images and attending to distorted ones.
• Attention mechanisms, including foveal vision and eye movement, are crucial for perceptual quality estimation.

Purpose of the Study:

• To develop an attention-driven NR-IQA method inspired by the free-energy principle and HVS properties.
• To leverage reinforcement learning (RL) for training the attention policy.
• To improve the accuracy of perceptual image quality prediction.

Main Methods:

• An attention-driven NR-IQA model using reinforcement learning (RL) was implemented.
• The model learns a policy to attend to multiple image regions concurrently.
• Region observations are aggregated using a weighted average inspired by robust averaging strategies.
• Rewards for policy learning are derived from distortion type classification and perceptual score estimation.

Main Results:

• The proposed method demonstrates superiority across multiple benchmark datasets (LIVE, TID2008, TID2013, CSIQ).
• The attention mechanism effectively guides the model to relevant image regions for quality assessment.
• Reinforcement learning successfully optimizes the policy for accurate quality prediction.

Conclusions:

• The attention-driven RL approach offers a promising direction for NR-IQA.
• Mimicking HVS attention significantly enhances image quality assessment performance.
• The method provides accurate perceptual quality predictions without requiring original reference images.