Answers are grounded in this paper and cite their sources.
No-reference image quality assessment (NR-IQA) is crucial for evaluating visual quality in real-world scenarios where pristine reference images are unavailable. Most existing deep learning-based methods adopt convolutional neural networks (CNNs) with local receptive fields to extract structural features, yet they suffer from limitations in modeling global statistical distortions. In contrast, pure mathematically modeled features, such as SVD (singular value decomposition), exhibit promising capability for characterizing global statistical degradations. Nevertheless, these mathematical strategies cannot capture fine-grained semantic information. Inspired by the distortion-dependent bimodal attention mechanism of the human visual system, we propose a novel dual-branch network named CNN-SVD Adaptive Gating Network (CSAG-Net). The framework integrates a lightweight CNN stream for local texture and structural feature extraction, and an SVD stream that operates on the perceptually uniform CIELAB color space to encode global degradation patterns via SVD computing. A learnable adaptive gating mechanism dynamically fuses the two heterogeneous feature streams in a dimension-wise manner, mimicking the HVS’s content-adaptive attention allocation. Extensive experiments on five benchmark datasets—including LIVE, CSIQ, TID2013, LIVEC, and KonIQ-10k—demonstrate that CSAG-Net achieves superior or highly competitive performance against state-of-the-art NR-IQA methods. Ablation studies confirm the complementary benefits of the SVD branch and the superiority of adaptive gating over fixed fusion strategies. With a compact parameter count and low computational cost, the proposed method offers an effective trade-off between prediction accuracy and efficiency.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: