基于雾密度感知增强的去雾图像质量评价方法

Dehazed Image Quality Assessment Method Based on Fog Density Perception Enhancement

去雾图像质量评价方法在图像去雾算法研究中发挥着重要作用,目前去雾质量评价方法存在雾密度感知能力不足的问题。提出一种基于雾密度感知增强的去雾图像质量评价(FDP-DIQA)方法,针对自然雾图的成像特点,从对比度信息及色彩信息2个方面提取6个特征,并结合以往的图像去雾研究结果提取3个特征,构成完整的去雾感知特征,增强对雾密度的感知能力。随后,结合雾图与去雾图像之间全局和低对比度区域的结构特征及块效应特征,对经平均池化形成的22维特征采用LightGBM进行模型构建,实现去雾图像质量评价。实验结果表明,FDP-DIQA方法在合成雾图数据集和自然雾图数据集上的表现优异,其加权平均后的斯皮尔曼等级相关系数、皮尔森线性相关系数、均方根误差分别为0.9623、0.9583、2.0981,大幅领先于同类方法,与人类主观评价有较高的一致性。

The Dehazed Image Quality Assessment(DIQA)method plays an important role in research on the image dehazing algorithm.At present,the dehazing quality assessment method has the problem of insufficient fog density perceived ability.In this regard,a new DIQA method called FDP-DIQA is proposed.According to the imagery characteristics of natural fog images,six features are extracted from the two aspects of contrast information and color information,and three features are extracted from the previous image dehazing research results to form a complete dehazing perception feature and enhance the perceived ability of fog density.Then,combined with the structural features and block effect features of the global and low-contrast regions between the fog map and defogged image,LightGBM is used for model construction of the 22-dimensional features formed by average pooling to realize quality evaluation of the defogged image.Experimental results show that the FDP-DIQA method performs well on both synthetic and natural fog map datasets,and its weighted average Spearman's Rank Correlation Coefficient(SRCC),Pearson Linear Correlation Coefficient(PLCC),and Root Mean Square Error(RMSE)are 0.9623,0.9583,and 2.0981,respectively,which is a significant improvement over existing similar methods and has a high consistency with human subjective evaluation.