【作者】 王蕊；

【导师】 王国宇；

【作者基本信息】 中国海洋大学 ， 计算机应用技术， 2014， 博士

【摘要】 经过散射介质（如雾、霾，水体等）成像的图像会严重退化，导致细节丢失、对比度下降以及颜色偏移失真。雾天和水下图像的退化主要是由于介质的衰减和散射所造成的，散射又分为前向散射和后向散射，前向散射导致图像细节的轻微模糊，而后向散射导致图像严重雾化。散射退化图像的复原在许多应用领域都有十分重要的作用，因此受到国内外众多研究人员的关注。不少学者采用提高设备性能的方法解决雾天及水下图像成像距离短的问题，但成本较高，难以广泛应用；采用图像处理方法来提高雾天及水下图像质量更加简单和适用。成像系统受散射介质的影响，图像退化原因复杂，对散射退化图像进行恢复时所需的先验知识必须完备。较早的复原方法通过多幅图像或人工交互获得更多的先验知识来进行去散射，也有不少基于测量的方法来获得散射介质的固有光学特性（IOP）参数并以此推导退化函数；这些方法不能有效解决现有退化图像实时性还原需求。本文针对以上问题深入研究了基于单幅图像且无需对实际成像环境中的介质进行实际环境测量的复原方法，主要工作和创新点包括以下几个方面：1）本文认为退化图像中的模糊程度和后向散射噪声本身包含了大量的与散射介质固有光学特性（IOP）相关的信息，结合微积分和物理极值条件的思想提出了散射介质分层分解模型来估计退化函数。假定每层厚度极小时后向散射噪声不相关，对分层模型进行无穷层数（即距离无穷远处）积分，然后利用纯后向散射背景（距离无穷远处如纯水体和天空区域）拟合得到以往需要实际测量的退化函数参数。2）散射退化与景深和散射介质浓度有关，本文开展了实际水池环境实验，分析和验证模型在这两方面的相关性：首先对不同浑浊程度的水体后向散射噪声进行详细分析，比较不同水质的模型参数；其次对实验中相同水环境一系列不同距离的图像进行测距分析，验证了光学距离与实际距离成正比例关系，因此在距离已知的情况下，分层模型还可以辅助求解和介质IOP相关的参数值。3）提出了一种迭代复原方法——“渐进反卷积”，完成图像去雾。通过模型推导出的退化函数和噪声信息可以结合多种反卷积方法对图像进行复原，本文提出了基于滤波器的线性渐进反卷积方法和基于全变分模型的非线性渐进反卷积方法;分别用这两种方法对实际雾天及水下图像进行复原，实验结果同目前广泛应用的两种单幅图像去雾方法进行了客观的比较和分析。4）全变分复原在迭代时使用图像全变分最小作为正则项约束条件，而基于传统滤波器的线性迭代过程也需要正则项约束，本文研究和比较了几种不同的清晰度评价函数作为正则项的有效性，提出最大熵正则化方法,结果表明这种方法适合没有退化前清晰目标图像作为参考的雾天及水下退化图像迭代复原。更多还原

【Abstract】 Images in a scattering medium (such as fog, water and smoke) are inevitably degraded,resulting in blurring details and color distortion as well as poor image contrast.Fog-degraded images and underwater images are both hazed due to the absorption andscattering of the scattering medium. The scattering can be divided into forwardscattering and backscattering, where the forward scattering leads to the light blur ofdetails and the backscattering causes the serious atomization. Image recovery in ascattering medium is highly desired for many applications and gains widely attentionsof the researchers: some researchers aim to solve the problem of increasing imagingrange by improving the equipment performance, which costs too much to be widelyused; some researches develop the image processing techniques to recover the hazeimage, which is considered to be more simple and powerful.Degradation of imaging system in scattering medium is complex; it needs full-fledged prior knowledge on the scattering medium to recover the haze image. Theearlier recovery methods were based on two or more images to obtain more priorknowledge of scattering, others were based on measuring to derive the degradationfunction; these methods cannot effectively solve the existing degraded image forreal-time requirement.In this paper, image recovery methods based on single image which are alsowithout measure of the scattering medium in the real environment are researched; themain work and innovations are listed as follows:1) The blur degradation and the backscattering noise themselves contain a largenumber of information related to the inherent optical properties (IOP) of the scatteringmedium. This paper proposes a multi-layered decomposition of the scattering volumeand models blurring function of a single-layer scattering volume based on manners of the calculus and the physical extremum problem. Supposing the backscattering noiseof each tiny singe-layer is uncorrelated, when the model is integrated to infinite, it canfit the degradation parameters just by the pure backscattering region.2) Since scattering degradation is associated with the depth of object andconcentration of the medium, this paper carried out an actual water environmentalexperiment to analyze the correctness of the multi-layered model on these two aspects.The experiment includes that: comparing the parameters of backscattering noise ofdifferent water; analyzing the depth correlation by images obtained at different rangesin the same water, and finding out the proportional relationship between the real depthand the parameter depth. What’s more, when the depth is known, this multi-layeredmodel can also help to solve the IOP related prior.3）This paper proposes an effective method called propagating deconvolution torecover single image degraded in a scattering medium. Parameters of thedeconvolution algorithm are estimated just from in-situ measurement of the purescattered background from one single image. Two kinds of propagating deconvolutionare used which are linear filter propagating deconvolution and nonlinear totalvariation propagating deconvolution. The results are analyzed and compared with theother two widely used single image dehaze methods.4）Total variation (TV) restoration uses the image TV minimization as theconvergence regularization in the iterations; While traditional linear filter also needs aconvergence criterion in the iterations of propagating deconvolution. This paperresearches several different criteria and come to a conclusion that image entropymaximization is suitable for the fog-degraded images and underwater images whosereference images are not available.更多还原