Benchmarking single-image dehazing methods: a systematic evaluation across diverse scenarios and datasets

Abstract

Single-image dehazing has witnessed remarkable progress in recent years, with numerous algorithms proposed to enhance visibility under adverse weather conditions. However, existing review articles often lack systematic performance comparisons under standardized experimental settings, making it challenging to objectively assess the true capabilities and limitations of different approaches. This paper presents a comprehensive review and benchmark for single-image dehazing across diverse scenarios. We establish a multi-dimensional benchmarking framework encompassing five widely used datasets–Foggy Cityscapes, ITS, O-Haze, RTTS, and ACDC–which span a broad spectrum of synthetic and real-world haze conditions. We systematically evaluate 17 representative dehazing algorithms from 2009 to 2025 by strictly adhering to consistent training configurations and parameter settings on each dataset, ensuring a fair and objective comparison across all evaluated methods. We assess performance across multiple dimensions: reference-based metrics (Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM)), no-reference metrics (Natural Image Quality Evaluator (NIQE), Fog Aware Density Evaluator (FADE)), computational efficiency, and practical impact on downstream tasks (e.g., object detection). Our results reveal that high-performing models on synthetic benchmarks often encounter generalization bottlenecks in real-world scenarios. Specifically, the evaluation identifies a divergence between reconstruction accuracy and functional utility, suggesting that excessive optimization for pixel-wise metrics may not always benefit downstream perception. This benchmark offers a comprehensive decision matrix, providing valuable insights for selecting and developing dehazing methods for robust, real-world computer vision systems.

Subjects