Scholarlink

<p>In large-scale visual systems, particularly in automated image triage and preprocessing pipelines, the reliability of downstream computer vision tasks depends on the quality of perceptual inputs and the interpretability of assessment outcomes. Existing blind image quality assessment methods are typically formulated as offline black-box predictors that produce a single scalar score. While indicating degradation, it provides limited actionable guidance, semantic awareness, and insufficient diagnostic information for system-level decision-making. To address these limitations, this work proposes an interpretable blind image quality assessment framework designed to serve as a perceptual monitoring component that provides diagnostic support for automated visual systems. The approach incorporates high-level semantic priors from a frozen vision–language model into a Vision Transformer-based assessment stream via feature-wise linear modulation, enabling content-aware quality evaluation. In parallel, a distortion diagnosis branch is jointly optimized to identify dominant degradation types and generate structured diagnostic cues that support adaptive restoration. Experiments on standard benchmarks demonstrate strong consistency with human subjective judgments, achieving Spearman’s rank correlation coefficients of 0.9509 on TID2013 and 0.9408 on KADID-10k. The model also operates at 65 frames per second, indicating a favorable balance between accuracy, interpretability, and efficiency.</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

Explainable blind image quality assessment with closed-loop semantic guidance and distortion diagnosis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月9日

查看全文
  • Chenye Song,
  • Fujiang Yuan,
  • Zhiwang Zhang

摘要

In large-scale visual systems, particularly in automated image triage and preprocessing pipelines, the reliability of downstream computer vision tasks depends on the quality of perceptual inputs and the interpretability of assessment outcomes. Existing blind image quality assessment methods are typically formulated as offline black-box predictors that produce a single scalar score. While indicating degradation, it provides limited actionable guidance, semantic awareness, and insufficient diagnostic information for system-level decision-making. To address these limitations, this work proposes an interpretable blind image quality assessment framework designed to serve as a perceptual monitoring component that provides diagnostic support for automated visual systems. The approach incorporates high-level semantic priors from a frozen vision–language model into a Vision Transformer-based assessment stream via feature-wise linear modulation, enabling content-aware quality evaluation. In parallel, a distortion diagnosis branch is jointly optimized to identify dominant degradation types and generate structured diagnostic cues that support adaptive restoration. Experiments on standard benchmarks demonstrate strong consistency with human subjective judgments, achieving Spearman’s rank correlation coefficients of 0.9509 on TID2013 and 0.9408 on KADID-10k. The model also operates at 65 frames per second, indicating a favorable balance between accuracy, interpretability, and efficiency.