A Swin transformer-based framework for digital media image quality assessment

Abstract

No-reference image quality assessment (NR-IQA) is an important task in image processing and is essential for automatically monitoring image quality during content distribution. Images captured under uncontrolled conditions may contain multiple authentic distortions, and their perceived quality depends on multiple dimensions, including pixel-level distortion features, semantic content structure, and perceptual aesthetics. Existing NR-IQA methods exhibit notable limitations in flexible multi-scale feature extraction, joint modeling of technical quality and perceptual aesthetics, and robust representation learning when subjective annotations are scarce. To address these issues, we present a Swin Transformer-based NR-IQA method comprising three core modules: a Multi-scale Window-adaptive Feature Extraction module (MW-SFE), which dynamically adjusts the window size and aggregates multi-granularity features across scales; a Quality-Aware Dual-Branch Evaluation Network (QA-DBN), which learns complementary technical-quality-oriented and aesthetics-oriented representations and adaptively integrates them through gated fusion; and a Content-Quality Contrastive Learning enhancement module (CQ-CL), which constructs content-level and quality-level contrastive objectives to alleviate the scarcity of subjective annotations. Experiments on two public NR-IQA datasets, LIVE-itW and KonIQ-10k, demonstrate that the proposed method improves Spearman Rank-Order Correlation Coefficient (SRCC) by 3.9% and 3.3% and Pearson Linear Correlation Coefficient (PLCC) by 3.6% and 3.2%, respectively, over the best-performing comparison method, validating its effectiveness for no-reference image quality prediction.