雾霾图像清晰化算法综述

PDF下载HTML

Title:

Review of hazy image sharpening methods

作者:

王科平¹, 杨艺¹, 费树岷²

1. 河南理工大学 电气工程与自动化学院，河南 焦作 454003;
2. 东南大学 自动化学院，江苏 南京 210096

Author(s):

WANG Keping¹, YANG Yi¹, FEI Shumin²

1. School of Electrical Engineering and Automation, He’nan Polytechnic University, Jiaozuo 454003;
2. School of Automation, Southeast University, Nanjing 210096, China

关键词:

图像清晰化; 图像去雾; 不适定问题; 图像降质; 大气散射模型; 深度学习; 无模型; 非均匀雾图

Keywords:

image sharpening; image dehazing; ill-posed problem; image degradation; atmospheric scattering model; deep learning; model-free; uneven hazy image

分类号:

TP391.41

DOI:

10.11992/tis.202110029

摘要:

雾霾图像不仅影响视觉效果，而且模糊不清晰的图像容易为后续识别、理解等高层次任务带来困难。雾霾图像清晰化问题是一个典型的不适定问题，其成像过程难以精确建模，消除图像中的雾霾面临巨大的挑战。近年来，研究者提出大量的图像去雾算法克服雾霾引起的图像降质退化，为全面认识和理解图像清晰化算法，论文对其进行梳理和综述。首先，对雾霾图像清晰化算法进行整理，根据雾霾退化过程是否有模型支持，将清晰化算法分为基于Retinex模型、大气散射模型去雾算法和无模型图像去雾算法。大气散射模型是有模型算法中主流模型，本文详细剖析了模型成像机理，并根据其成像机制揭示大气散射模型容易受大气浓度均匀分布假设的限制，较难处理非均匀雾霾图像问题。基于深度学习的无模型图像去雾算法则不仅可以应对非均匀雾霾图像，而且去雾性能获得了极大地提升。其次，本文汇总了近年来常用去雾数据集，从数据集适应范围、规模、可扩展性等多个维度进行总结。并根据雾霾图像形成方式，对人工合成雾霾数据集和真实拍摄数据集分别从定性和定量的角度探讨了数据集对图像去雾算法的影响。

Abstract:

Hazy images not only induce visual effects but also easily introduce difficulties to subsequent high-level tasks, such as image recognition and understanding. Image dehazing is a typical ill-posed problem, and accurately modeling the imaging process is difficult. Therefore, eliminating the haze in the image faces enormous challenges. Researchers have proposed numerous methods to overcome the hazy image degradation caused by haze. First, this paper summarizes the image dehazing methods to understand and organize them. Whether the haze degradation process is supported by the model, the clarity algorithm is generally divided into a Retinex-based model, atmospheric scattering model defogging algorithm, and model-free image dehazing algorithm. The atmospheric scattering model is the typical model-based dehazing. The imaging mechanism of the model is comprehensively analyzed. However, addressing the non-uniform hazy image problem is difficult because the atmospheric scattering model is easily restricted to an assumption that the atmospheric concentration is distributed uniformly. The deep learning-based model-free dehazing algorithm not only deals with the non-uniform hazy image but also gains a considerable improvement in dehazing performance. Second, this paper summarizes the commonly used image dehazing data sets in recent years and compares the data sets from multiple dimensions, such as the scope of application, scale, and expandability. Moreover, the influence of a synthetic hazy data set and the data set of images shot in reality on the image dehazing algorithm is qualitatively and quantitatively discussed in accordance with the formation mode of hazy images.