【作者】 陈健；

【导师】 吴昊天；

【作者基本信息】 华南理工大学 ， 计算机技术（专业学位）， 2024， 硕士

【摘要】 在水下环境所拍摄的图像有时会呈现出模糊与颜色偏差现象。这是因为水中悬浮微粒的影响,使得从空气照射到水下的光在传播到相机的过程中发生了散射与吸收作用。因此,光的散射与吸收作用会导致水下拍摄到的图像质量发生退化,进而造成了图像在使用上的困难。针对上述问题,国内外学者提出了许多水下图像增强方法。目前,基于图像形成模型的方法对退化水下图像的质量恢复,起到一定程度的积极作用。然而,现存方法在图像形成模型的参数估计中存在偏差,这些偏差的参数影响到水下图像的质量恢复,恢复图像存在模糊与颜色偏差现象。此外,尽管基于指标的分类数据集在水下图像增强中得到广泛使用,然而,现有的分类数据集在构建过程中使用到存在偏差的质量指标,这些偏差的质量指标影响到图像数据的分类。针对上述问题,本文基于一种已有的图像形成模型和一种综合色度(Colorfulness)、对比度(Contrast)和雾浓度(Fog density)的图像质量度量(简称CCF)开展水下图像增强研究,主要的研究内容如下:(1)针对传统的基于图像形成模型的水下图像增强方法中,模糊度图的不准确估计问题,本文提出了一个全局性的模糊度图估计模型。由于该模型采用了数据驱动的策略,并且考虑到图像的退化是一个全局性的过程,所提出的模型能有效的估计图像的模糊度图,并能用于图像形成模型中背景光参数和透射率参数的估计。此外,为了解决背景光的估计偏差所导致图像质量不高的问题,在背景光参数的估计中,本文还提出了一项综合性的估计策略,综合使用三个候选的背景光中间值来确定最终所估计的背景光参数。根据所估计的背景光参数和模糊度图,在使用图像形成模型恢复图像时,能有效的解决图像模糊与颜色偏差问题。实验结果表明,所提出的基于图像形成模型的水下图像增强方法对于不同退化水平的图像都有比较好的效果。(2)为了进一步探究现存的水下图像数据集,本文开展了一项水下图像数据集的对比研究。考虑到现存数据集在数据的收集方式上,或是图像收集的数量与新鲜度上,仍然存在着提升的空间。作为现有数据集的补充,本文提出了一个基于CCF度量的水下图像数据集。由于CCF度量综合考虑了水下图像的色度、对比度和雾浓度,在使用CCF度量将所收集到的水下图像进行划分时,能有效的划分成具有不同退化水平的图像子集。当这些具有不同退化水平的图像子集用于评估水下图像增强方法时,能更好的分析水下图像增强方法实际的表现情况。通过十一种水下图像增强方法在所提出数据集上的运用,以及四个数据集关于数据多样性的评估实验结果,表明所提出的数据集能有效的用于评估水下图像增强方法,并且提出的数据集还能用于水下图像的视觉任务中。此外,实验所使用的十一种方法所生成的部分水下图像,根据实验的结果补充到所提出的数据集中,以此提高数据集的新鲜度。综上所述,在使用本文提出的基于图像形成模型的水下图像增强方法时,退化图像的清晰程度不仅得到了改善,而且图像的整体质量也得到了提升。除此之外,本文提出的基于CCF度量的水下图像数据集,不仅提高了现有数据集的新鲜度,在空间信息与颜色多样性上更具优势,而且CCF度量的使用让数据集内的图像划分更加准确。更多还原

【Abstract】 Images taken in underwater environment sometimes present low visibility and color deviation.This phenomenon is caused by the movement of suspended particles in water,which leads to the light scattering and absorbing from air to underwater.Therefore,the quality of underwater images is degraded due to the scattering and absorption of light,leading to the difficulties of using those images.In response to the above problems,domestic and foreign scholars have developed many underwater image enhancement approaches.At present,methods based on the image formation model have made a positive effect in restoring degraded underwater images.However,the parameters required by these methods were inaccurately estimated,resulting in undesirable image restoration.In addition,although the metric-based image sets have been widely used in underwater image enhancement,some inaccurate metrics were still utilized in the construction of those metric-based image sets.However,After the construction of metric-based image sets,the evaluation of image enhancement methods is affected by those inaccurate metrics.Consequently,research on underwater image enhancement based on an image quality metric that combines colorfulness,contrast,and fog density(referred to as CCF)and an image formation model have been conducted in thesis.The main work of this thesis is as follows:(1)Due to the inaccurate estimation of blurriness maps in traditional underwater image enhancement methods,a global blurriness map estimation model have been developed.Since a data-driven strategy is used during the model training and the global image degradation have been taken into account in the model,the blurriness map can be effectively estimated and used for determining background light and transmission map parameters in image formation model.In addition,in order to solve the low quality of images caused by inaccurate background light estimation,a multi-scale estimation strategy have been developed to determine the final background light parameter.According to the estimated background light and blurriness map,the problem of low image visibility and color deviation is effectively solved when using the image formation model to restore the image.Experiment results show that the images with different degradation levels can be well restored by the proposed method.(2)In order to further explore existing underwater image datasets,a comparative study of underwater image datasets have been conducted in this thesis.Considering existing datasets deficient in data acquisition strategy,or lack of quantity/freshness to some extent,we propose an underwater image set based on CCF metric as a supplement to existing datasets.Since the chroma,contrast and fog density of underwater images are comprehensively combined in the CCF metric,the proposed image set can be effectively divided into three image subsets by the CCF metric.When these image subsets with different degradation levels are used to evaluate underwater image enhancement methods,the actual performance of underwater image enhancement methods can be better analyzed and realized.Through the performance of eleven underwater image enhancement methods on the proposed image set,as well as the evaluation results of four datasets and the proposed image set,it is shown that the proposed image set can be effectively used to measure underwater image enhancement methods and can be used in some underwater image vision tasks,such as underwater image object detection.In addition,some underwater images generated by eleven methods are included in the proposed image set so as to improve the freshness of the proposed image set.In summary,the clarity of degraded images is not only improved and the overall quality of those images is enhanced by using the proposed underwater image enhancement approach based on image formation model.In addition,the freshness,spatial information and color diversity of existing datasets can be increased in the proposed CCF-based image set,while the classification of subsets in proposed image set is more accurate by using the CCF metric.更多还原