What is it about?
This work studies how to automatically evaluate image quality when no original reference image is available. This is important for many real-world applications, such as image enhancement, compression, and machine vision systems, where reference images are often inaccessible. We propose a new image quality assessment model that combines both global and local visual information. The global branch captures the overall image structure and long-range relationships, while the local branch focuses on fine-grained distortion details. These complementary features are progressively integrated to provide more accurate quality predictions. We also observe that images with similar semantic content often exhibit highly consistent perceptual quality under the same distortion conditions. Based on this finding, we introduce a semantic-aligned quality transfer method, which transfers quality labels across semantically similar images to construct an augmented training dataset. This enables the model to learn richer distortion-quality relationships and improves its generalization ability.
Featured Image
Photo by Wesley Tingey on Unsplash
Why is it important?
Accurate no-reference image quality assessment remains challenging because subjective quality depends on both local distortion patterns and global semantic context. Existing methods often emphasize only one aspect, limiting their robustness across diverse distortion types. Our work is important because it addresses both model design and data scarcity simultaneously. The proposed global-local framework captures distortions at multiple perceptual scales, while the semantic-aligned quality transfer strategy expands training supervision without requiring costly human annotation. This provides a more scalable and practical solution for developing reliable image quality assessment systems, with potential applications in image restoration, compression optimization, and quality-aware visual processing.
Perspectives
A key insight of this work is that perceptual quality is not entirely image-specific; it can be meaningfully transferred across semantically similar content under identical distortions. This observation motivated us to rethink how subjective quality annotations can be reused more efficiently. From a broader perspective, we believe future image quality assessment research should move beyond purely architectural improvements and pay greater attention to data construction strategies. Combining effective feature modeling with principled label transfer offers a promising direction for building more generalizable and data-efficient quality assessment models.
Xiaoqi Wang
Sun Yat-Sen University
Read the Original
This page is a summary of: Global-Local Progressive Integration and Semantic-Aligned Quality Transfer for No-Reference Image Quality Assessment, ACM Transactions on Multimedia Computing Communications and Applications, May 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3815779.
You can read the full text:
Resources
Contributors
The following have contributed to this page
