面向真实场景的无参考图像质量评价研究综述

No-Reference Image Quality Assessment for Real-World Scenarios: A Review

摘要: 移动摄影、社交媒体与图像处理技术的发展，使图像质量评价由受控合成失真研究逐步转向真实场景无参考质量评价。本文聚焦真实场景下的无参考图像质量评价(NR-IQA)，界定任务边界，归纳复合失真、空间非均匀、质量感知对图像内容与语义的依赖、跨数据集域偏移及主观评分不确定性等挑战，并比较代表性数据集与评测协议。按照技术路线演进，系统梳理自然场景统计与浅层学习、卷积神经网络(CNN)与内容自适应建模、排序与自监督学习、跨域学习、Transformer及视觉语言模型等方法，并从数据集内性能、跨数据集泛化、分数校准与不确定性估计、局部质量证据和语言解释等方面分析其优势、局限与适用条件。本文认为，真实场景NR-IQA应由追求单一数据集内精度，转向兼顾泛化性、校准性、可解释性与实际应用可行性。

Abstract: The widespread adoption of mobile photography, the growth of social media, and advances in image processing technologies have shifted image quality assessment (IQA) research from the study of synthetic distortions under controlled conditions toward no-reference quality assessment in real-world scenarios. This review focuses on real-world no-reference image quality assessment (NR-IQA), delineates its task boundaries, summarizes key challenges, including composite distortions, spatial non-uniformity, the dependence of quality perception on image content and semantics, cross-dataset domain shifts, and uncertainty in subjective scores, and compares representative datasets and evaluation protocols. Following the evolution of technical approaches, we systematically review natural scene statistics and shallow learning, convolutional neural networks and content-adaptive modeling, ranking and self-supervised learning, cross-domain learning, Transformers, and vision-language models. We analyze their strengths, limitations, and applicability in terms of within-dataset performance and cross-dataset generalization, score calibration and uncertainty estimation, local quality evidence, and language-based explanations. We argue that NR-IQA in the real-world should move beyond the pursuit of accuracy on a single dataset and instead consider cross-dataset generalization, reliable score calibration, interpretability, and practical applicability.