SIA OpenIR  > 光电信息技术研究室
面向探测识别的彩色图像去雾与评价方法研究
Alternative TitleColor Image Defogging and Defogged Image Quality Assessment for Detection and Recognition
宋颖超
Department光电信息技术研究室
Thesis Advisor罗海波
Keyword图像去雾 图像增强 天空识别 去雾图像评价 颜色对比度描述子
Pages124页
Degree Discipline模式识别与智能系统
Degree Name博士
2018-11-27
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract

本文主要针对面向探测识别的图像去雾与增强任务需要,在分析大气散射理论和雾天图像退化机理的基础上,归纳并总结了图像去雾和去雾效果评价研究领域现有的典型方法,对该领域尚未有效解决的多个关键问题展开了深入的研究。本文重点研究了四个方面的内容:天空光的准确估计、场景自适应去雾、去雾图像增强以及去雾效果评价,主要研究工作总结如下:(1)图像去雾效果定量评价方法。在面向探测识别的彩色图像去雾效果评价任务中,本文将去雾后图像对比度的提升效果作为主要关注的指标,提出了一种基于雾线理论的去雾图像颜色对比度描述子。针对图像去雾效果评价任务中无法获取同一场景的清晰图像作为评价参考的问题,本文从雾天图像退化的机理出发,以雾线理论为基础,在雾图中对像素进行雾线聚类,并用复原图像局部区域内不同颜色类的类间标准差来衡量复原图像的局部颜色对比度。设计了主观评价实验和合成图像评价实验,实验结果验证了所提出的评价指标的科学性。(2)场景自适应图像去雾方法。为了提高图像去雾算法对场景的自适应能力,实现不同场景下较优的自动去雾效果,本文在暗通道先验去雾方法的基础上加以改进,提出了一种场景自适应暗通道去雾方法。该方法根据图像的颜色和边缘特征自适应地调节暗通道的求解尺度,兼顾不同尺度复原图像的优点,使去雾后的图像色彩自然,对比度提升显著,且有效抑制了“白边”效应。此外,针对现有去雾方法在天空光估计过程中易出现估计点落入前景区域而造成复原图像偏暗和色彩失真的问题,本文从天空光的物理意义出发,提出了天空光估计的约束条件,为得到更为准确的天空光估计值提供依据。(3)雾天图像天空识别方法。为了得到更为准确的、与天空光物理意义相符的天空光估计值,提出了一种雾天场景天空识别方法。该方法以雾感知浓度和场景深度等与雾相关的特征来表征天空,再通过非平衡支持向量机(Imbalance-SVM)分类和相似性度量对图像块进行分类,最终识别出任意形状的天空。此外,为了训练模型和验证算法,建立了一个雾霾天气天空场景的数据集HazySky。通过对本文提出的天空识别方法在HazySky数据集以及公开数据集SkyFinder上进行的大量实验,结果表明,本文提出的方法对天气和光照条件的适应能力较强,不仅对雾天图像有效,对其它天气条件下的图像也能取得较高的天空识别精度。(4)基于有效天空识别的图像去雾与增强方法。为了实现雾天降质图像清晰度和对比度的全面提升,提出了一种基于有效天空识别的图像去雾与增强方法。将天空识别方法应用于图像去雾过程中的天空光估计,可以提高天空光估计的精度;将尺度自适应暗通道方法应用于图像去雾过程中的透射率估计,可以减小复原图像的“白边”效应;将自适应伽马校正方法应用于图像去雾后的增强,可以进一步提升复原图像对比度、改善视觉效果。实验结果表明,经该方法处理后的图像较去雾后不做增强处理的图像在信息熵和对比度等指标上均有大幅提升,且未引起明显的结构失真。

Other Abstract

Aiming at the demand of image defogging and enhancement tasks for detection and recognition, after analyzing atmospheric scattering theory and foggy image degradation mechanism, this paper summarizes the typical methods in the field of image defogging and defogged image quality assessment, and carries out deep research on several key issues that were not effectively solved in this field. This paper mainly focuses on four research contents: accurate estimation of sky light, scene adaptive defogging, defogged image enhancement and defogging effect assessment. The main research works are summarized as follows: (1) Method of quantitative assessment on image defogging effect. During the assessment task of color image defogging effect for detection and recognition, the improvement effect of contrast is taken as the main focus target, and a color contrast descriptor of defogged image based on the haze-line theory is proposed. Aiming at the problem that the foggy-free image of the same scene cannot be obtained as the evaluation reference during the process of image defogging effect assessment, based on the degradation mechanism of foggy image and the haze-line theory, this paper clusters the pixels in foggy image, and measures the local color contrast of restored image though computing the inter-cluster standard deviation of different color clusters within local image patch. Both subjective evaluation experiments and synthetic image evaluation experiments are designed and carried out, the experimental results verify the scientificity of the proposed evaluation indicator. (2) Method of scene adaptive image defogging. In order to improve the adaptive ability of image defogging algorithm to scene and achieve better automatic defogging effect in different scenes, this paper proposes an improved image defogging method based on the dark channel prior, namely scene adaptive dark channel defogging method. The method adjusts the scale of dark channel according to the color and edge features of image, and takes into account the advantages of defogging results achieved by different scales. The method achieves significant enhancement in contrast, natural restoration in color and great suppression in “halo” effect. Furthermore, in order to solve the problem of low brightness and color distortion caused by unreasonable sky light estimation, the constraints of sky light estimation are proposed from the physical meaning of sky light, which provide a basis for more accurate sky light estimation. (3) Method of sky detection in foggy image. A sky detection method for foggy scene is proposed in order to obtain a more accurate estimation of sky light according with its physical meaning. The method introduces several fog-relevant features that reflect the foggy perceptual density and the scene depth to characterize sky. Based on these features, the sky with arbitrary shape is detected by Imbalanced-SVM classifying and similarity measuring. Moreover, a sky dataset of foggy scene HazySky is built for model training and performance evaluation. To evaluate the performance of the proposed sky detection method, extensive experiments are conducted both on the HazySky dataset and the SkyFinder dataset. The experimental results demonstrate that the proposed method has strong adaptability to weather and lighting conditions, and achieves higher detection accuracy not only in foggy images but also in images of other weather conditions. (4) Method of image defogging and enhancement by effective sky detection. In order to improve the definition and contrast of foggy-degraded images, a defogging and enhancement method based on effective sky detection is proposed. This method, firstly, can improve the accuracy of sky light estimation by applying sky detection method to the sky light estimation step of image defogging process, secondly, can reduce "halo" effect of restored image by applying scale adaptive dark channel method to the transmission estimation step of image defogging process, and finally, can further improve the contrast and visual effect of restored images by applying adaptive gamma correction method to the image enhancement step after image defogging process. The experimental results demonstrate that the entropy and color contrast of the images processed by the proposed method are greatly improved than ones without enhancing after defogging, and the defogging process does not cause obvious structural distortion.

Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/23645
Collection光电信息技术研究室
Recommended Citation
GB/T 7714
宋颖超. 面向探测识别的彩色图像去雾与评价方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2018.
Files in This Item:
File Name/Size DocType Version Access License
面向探测识别的彩色图像去雾与评价方法研究(21405KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[宋颖超]'s Articles
Baidu academic
Similar articles in Baidu academic
[宋颖超]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[宋颖超]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.