No-Reference Image Quality Assessment Combining Swin-Transformer and Natural Scene Statistics

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• No-reference image quality assessment (NR-IQA) models predict image quality without a reference image.
• Existing NR-IQA methods struggle to simultaneously capture global and local image information and mitigate information loss from resizing.

Purpose of the Study:

• To develop an advanced NR-IQA model that overcomes limitations of current methods.
• To improve the accuracy and robustness of image quality assessment by integrating multi-scale feature extraction and natural scene statistics.

Main Methods:

• A novel model combining Swin-Transformer for multi-scale feature extraction and Natural Scene Statistics (NSS) to address resizing-induced information loss.
• Incorporation of a feature enhancement module, deformable convolution, and dual-branch attention to refine feature representation and focus on critical image areas.
• Utilized a normalized loss function to enhance model convergence and stability.

Main Results:

• The proposed model achieved state-of-the-art performance on six diverse image quality assessment datasets (synthetic and authentic).
• Demonstrated superior performance compared to the DACNN method, achieving Spearman rank correlation coefficients of 0.922 on KADID and 0.923 on KonIQ.
• Showcased excellent capability in handling both synthetic and authentic image scenes.

Conclusions:

• The integrated Swin-Transformer and NSS approach offers a robust and accurate solution for NR-IQA.
• The model's ability to simultaneously process global and local information and compensate for resizing artifacts leads to improved alignment with human perception.