Scholarlink

<p>High Dynamic Range (HDR) images pose significant challenges to image quality assessment (IQA) due to their wide luminance range and complex perceptual characteristics. To address these challenges, this paper proposes a novel HDR-IQA model, termed Multi-Scale Wavelet Vision Transformer (MS-WaveViT), which integrates multi-scale wavelet analysis with a Vision Transformer backbone to effectively model the multi-level perceptual features of HDR content. First, we introduce a wavelet-domain attention mechanism that leverages Daubechies-4 (db4) wavelet decomposition to extract multi-scale and multi-band frequency components, yielding fine-grained and frequency-aware representations. Second, an HDR-aware enhancement module is designed to modulate feature responses according to the image luminance distribution, thereby strengthening the discrimination between bright and dark regions. Third, a HyperNet-TargetNet prediction framework is adopted to enable content-adaptive quality regression and improve the flexibility and generalization ability of the model. Extensive experiments conducted on two widely used HDR-IQA datasets demonstrate the superiority of the proposed method, achieving a PLCC of 0.9649 and an SRCC of 0.9693 on the Narwaria dataset, and further improving to a PLCC of 0.9807 and an SRCC of 0.9703 on the Korshunov dataset. Additional ablation and robustness analyses further verify the effectiveness of the proposed modules and the suitability of the db4 wavelet basis for HDR-IQA. In addition, the proposed framework is designed with practical computational deployment in mind: patch-based inference is naturally parallelizable on GPUs, and the reported complexity, throughput, and memory analysis further support its applicability to high-throughput HDR quality evaluation scenarios requiring efficient processing. These results highlight the effectiveness of MS-WaveViT in capturing critical quality cues and maintaining strong perceptual consistency with human subjective judgments in HDR images. The source code and trained models are publicly available at <a href="https://github.com/a947683707/MS-WaveViT.git">https://github.com/a947683707/MS-WaveViT.git</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-scale wavelet vision transformer with HDR-aware attention for high dynamic range image quality assessment

  • 期刊论文

  • 发表日期：

    2026年6月18日

查看全文
  • Rui Zhang,
  • Wu Dong,
  • Likun Lu,
  • Ziyi Zhou,
  • Tianqi Zhang,
  • Weipeng Niu

摘要

High Dynamic Range (HDR) images pose significant challenges to image quality assessment (IQA) due to their wide luminance range and complex perceptual characteristics. To address these challenges, this paper proposes a novel HDR-IQA model, termed Multi-Scale Wavelet Vision Transformer (MS-WaveViT), which integrates multi-scale wavelet analysis with a Vision Transformer backbone to effectively model the multi-level perceptual features of HDR content. First, we introduce a wavelet-domain attention mechanism that leverages Daubechies-4 (db4) wavelet decomposition to extract multi-scale and multi-band frequency components, yielding fine-grained and frequency-aware representations. Second, an HDR-aware enhancement module is designed to modulate feature responses according to the image luminance distribution, thereby strengthening the discrimination between bright and dark regions. Third, a HyperNet-TargetNet prediction framework is adopted to enable content-adaptive quality regression and improve the flexibility and generalization ability of the model. Extensive experiments conducted on two widely used HDR-IQA datasets demonstrate the superiority of the proposed method, achieving a PLCC of 0.9649 and an SRCC of 0.9693 on the Narwaria dataset, and further improving to a PLCC of 0.9807 and an SRCC of 0.9703 on the Korshunov dataset. Additional ablation and robustness analyses further verify the effectiveness of the proposed modules and the suitability of the db4 wavelet basis for HDR-IQA. In addition, the proposed framework is designed with practical computational deployment in mind: patch-based inference is naturally parallelizable on GPUs, and the reported complexity, throughput, and memory analysis further support its applicability to high-throughput HDR quality evaluation scenarios requiring efficient processing. These results highlight the effectiveness of MS-WaveViT in capturing critical quality cues and maintaining strong perceptual consistency with human subjective judgments in HDR images. The source code and trained models are publicly available at https://github.com/a947683707/MS-WaveViT.git.