SIA OpenIR  > 机器人学研究室
可见光散射环境下的图像恢复算法研究
Alternative TitleResearch on Image Restoration Algorithms in Visible-Light Scattering Environment
崔童1,2
Department机器人学研究室
Thesis Advisor唐延东
Keyword大气光亮度权重先验 隐分割边界限定 吸收透射率 时空导向滤波 强散射水下图像恢复
Pages99页
Degree Discipline模式识别与智能系统
Degree Name博士
2019-05-19
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract本文从可见光散射环境成像的物理机制出发,以大气散射吸收理论和散射成像物理模型为基础,对雾图像和水下图像进行了物理建模,提出了有效的图像去雾、视频去雾和强散射水下图像清晰化算法。实验证明所提出算法能够获得高清晰度、高对比度且色彩自然的恢复结果,与现有先进算法相比具有一定的优越性。本文的主要工作包括以下几个方面:(1)基于大气光亮度权重先验的图像去雾算法。在雾环境成像模型中,雾介质透射率代表无雾图像辐射强度的权重,其精度直接决定去雾结果的优劣。为了得到准确的透射率,我们提出了一个大气光亮度权重先验LWP(Luminance Weight Prior)。不同于大多现有算法重点关注于无雾图像辐射强度的衰减程度,LWP先验从研究大气光强度的衰减程度入手解决透射率估计问题,所获得的大气光权重透射率能够取得较好的去雾效果。为了进一步提高去雾结果的质量,我们提出了一个基于物理先验的优化方程对LWP先验进行后处理。该后处理优化方程能够自适应地调节大气光亮度权重的取值范围,保证大气光亮度权重矩阵连续性,在浓雾区域可有效提高无雾图像的对比度和纹理清晰度。(2)基于隐分割的边界限定透射率估计和权重L1-范数正则化的图像去雾算法。为了解决暗通道先验在高亮区域低估透射率而在非高亮区域高估透射率的缺陷,利用雾介质透射率的本质特征,我们提出了基于隐区域分割的透射率估计方法。该透射率估计方法通过统一的限定条件自然地将图像分割为高亮区域和非高亮区域,并分别对这两类区域的透射率进行修正,得到准确的初始透射率。为了使透射率满足局部一致性,我们提出了权重的L1-范数正则化优化方程对初始透射率进行优化。所提出的图像去雾算法能够有效提升图像清晰度和对比度,获得高颜色保真度的无雾图像,相比于目前主流的算法取得了较好的效果。(3)基于融合透射率和时空导向滤波的视频去雾算法。根据雾的光学成因,为了弥补经典去雾模型忽略大气吸收衰减的缺陷,我们提出了一个包含吸收透射率的去雾模型。我们基于比尔-兰伯特法则和所提出的亮度饱和度比LSR(Luminance-Saturation Ratio)先验对吸收透射率进行建模,并将此吸收透射率与隐分割边界限定方法所求得的散射透射率相融合,对雾介质透射率进行全面的描述。所提出的融合透射率能够有效地抑制由于忽略吸收衰减而在近景处产生的过饱和噪声。针对视频去雾算法中常产生闪烁噪声的问题,我们利用前后帧间的相关性,将导向滤波扩展到时域空间,提出了时空导向滤波(ST-GIF)能量方程对透射率进行优化。所提出的视频去雾算法能够有效提升图像对比度,抑制近景处的过饱和噪声,满足实时处理需求。与几种目前主流的视频去雾算法相比,本文算法在近景区域的去雾结果色彩更加自然、真实。(4)基于光场数据的水下图像恢复算法。针对强散射水下环境,我们提出了一个包含后向散射的水下图像恢复模型。为了获得浑浊水介质透射率,我们利用光场相机能够提供不同视角图像及粗略深度信息的优势,提出了一个基于光场数据切分重聚焦的透射率估计方法。我们将所获得的初始透射率与散焦、匹配信息相融合获得优化的深度信息,能够有效提升初始透射率精度。在强散射水下环境中,本文算法可获得对比度更高、纹理更加丰富且色彩保真度较高的恢复结果。
Other AbstractStarting from the physical mechanism of visible-light scattering environment imaging, based on the theory of atmospheric scattering absorption and the physical model of scattering imaging, in this paper, we physically modeled the haze images and underwater images. Several effective image dehazing, video dehazing, and strong scattering underwater image restoration algorithms have been proposed. The experimental results show that the proposed algorithms can achieve high definition, high contrast, and natural color restoration results, and the proposed algorithms are superior to the state-of-the-art algorithms. The main work of this paper includes the following aspects: (1) Image dehazing algorithm based on the atmospheric light luminance weight prior (LWP). In the hazy environment imaging model, the transmission of the haze medium represents the weight of the radiance intensity of the haze-free image, and its accuracy directly determines the quality of the dehazing results. In order to obtain accurate transmission, we proposed an atmospheric light luminance weight prior (LWP). Different from most existing algorithms, which focus on the attenuation of the radiance intensity of haze-free images, to estimate the transmission, the LWP prior starts with studying the attenuation degree of atmospheric-light intensity, the obtained atmospheric-light weight transmission can achieve good dehazing effects. In order to further improve the quality of dehazing results, a post-processing optimizing function based on physical prior in charging of optimizing LWP is proposed. This post-processing optimization function can adjust the value range of atmospheric brightness weight adaptively, ensure the continuity of atmospheric brightness weight matrix, while effectively improving the contrast and texture clarity of haze-free images, especially in dense haze regions. (2) Image dehazing algorithm based on latent region-segmentation boundary-constraint transmission estimation and weight L1-norm regularization. To solve the defects of dark channel prior that underestimates the transmission in the highlighted regions and overestimates the transmission in the non-highlighted regions, we proposed a transmission estimation method based on latent region-segmentation boundary-constraint. Using the essential characteristics of the haze medium transmission, this transmission estimation algorithm naturally divides a hazy image into highlighted regions and non-highlighted regions through an integrated constraint condition, corrects the transmission in these two kinds of regions, and obtains high accuracy scattering transmission. In order to make the transmission satisfy the local consistency, we proposed a weighted L1-norm regularization energy function to optimize the initial transmission. The proposed image dehazing algorithm can effectively improve the clarity and contrast of images, obtain the haze-free images with high color fidelity. Compared with the current mainstream algorithms, the proposed algorithm can achieve better results. (3) The video dehazing algorithm based on fusion transmission and spatio-temporal guided image filter (ST-GIF). According to the optical cause of haze, to solve the defect of the atmospheric scattering dehazing model ignoring atmospheric light absorption attenuation, we proposed a novel dehazing model that includes absorption transmission. We modeled the absorption transmission based on the Beer-Lambert rule and the proposed luminance-saturation ratio (LSR) prior and fused this absorption transmission with the scattering transmission obtained by latent region-segmentation boundary-constraint method to comprehensively describe the transmission of haze medium. In the close shot, the fusion transmission can effectively suppress the over-saturation noises which are led by ignoring the absorption attenuation. In order to solve the flicker noises which often generate in video dehazing algorithms, we extended the guided image filter to the time domain space and proposed the spatio-temporal guided image filter (ST-GIF) by using the correlative information between the frames in the video. The proposed video dehazing algorithm can effectively improve the image contrast, suppress over-saturation noise in the close shot, and meet the requirements of real-time processing. Compared with several current mainstream video dehazing algorithms, the colors of our dehazing results are more realistic and natural. (4) Underwater image restoration algorithm based on light field data. A novel underwater image restoration model with backscatter is proposed for strong scattering underwater environment imaging. In order to obtain the transmission of the turbid water medium, taking advantage of the optical field camera's ability to provide images from different perspectives and rough depth information, we proposed a transmission estimation method based on the shearing and refocusing light field data. We combined the obtained original transmission with the defocus and correspondence cues to get the refined depth. The fusion depth can effectively improve the accuracy of initial transmission. The proposed algorithm can obtain the recovery results with higher contrast, richer texture and higher color fidelity in the strong scattering underwater environment.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/25157
Collection机器人学研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院大学
Recommended Citation
GB/T 7714
崔童. 可见光散射环境下的图像恢复算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
可见光散射环境下的图像恢复算法研究.pd(24867KB)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
[崔童]'s Articles
Baidu academic
Similar articles in Baidu academic
[崔童]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[崔童]'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.