Evaluation Metrics for Single Image Dehazing: A Domain-Specific Composite Evaluation Framework and Comprehensive Review
Ghazala Rafiq, Muhammad Rafiq, Ho-Youl Jung, Gyu Sang Choi · IEEE Access · 2026
Single image dehazing is a fundamental computer vision problem aimed at recovering haze-free images from atmospherically degraded observations. Despite steady algorithmic advances spanning physics-based priors, deep learning, and hybrid frameworks, evaluation in this field remains fragmented and inconsistent, with the community over-relying on pixel-fidelity metrics whose limitations are well documented but structurally unaddressed. This review places evaluation at its core, providing the first comprehensive and critical analysis of performance assessment methodologies for single image dehazing. We systematically examine full-reference metrics, no-reference metrics, task-driven evaluations, and subjective assessments across synthetic and real-world benchmarks, covering their theoretical foundations, computational properties, and domain-specific limitations.We identify and consolidate evaluation protocols for multiple application domains spanning safety-critical autonomous systems, specialized vision systems, remote sensing & environmental monitoring, and urban intelligence. The critical analysis reveals a persistent disconnect between metric convenience and real-world perceptual utility. To address this structurally, we propose a Domain-Specific Composite Evaluation Framework that integrates fidelity, perceptual, task-based, and subjective measures into a unified, application-aware scoring index (QSID), operationalized through a four-step metric selection decision flowchart and a domain-specific guideline specifying mandatory, and supplementary evaluation metrics for each application domain.We further identify key open challenges and outline a research agenda toward robust, holistic, and perceptually aligned bench marking for single image dehazing.