No-reference image quality assessment based on global awareness.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Image Quality Assessment (IQA) is crucial but challenging for distorted images without reference.
• Existing methods struggle with accurately perceiving quality in authentic distorted images.

Purpose of the Study:

• To develop an advanced no-reference image quality assessment (NR-IQA) model for authentic distorted images.
• To improve the accuracy and generalization ability of IQA models.

Main Methods:

• Utilized the Swin Transformer (ST) for effective feature extraction.
• Designed a Global Self-Attention Block (GSAB) to integrate spatial and channel information.
• Incorporated a Transformer block to capture long-range dependencies.
• Employed a Dual-Branching structure for final quality score prediction.

Main Results:

• The proposed method outperformed existing approaches on four synthetic and two authentic datasets.
• Experimental results were weighted by dataset size, confirming the model's effectiveness.
• The method demonstrated strong generalization ability in tests.

Conclusions:

• The novel NR-IQA method shows significant improvements over current state-of-the-art techniques.
• The model's robust generalization ability makes it suitable for real-world applications.
• Future work will involve releasing the code for broader accessibility.