【作者】 张祥光；

【导师】 刘智深；

【作者基本信息】 中国海洋大学 ， 地图学与地理信息系统， 2009， 博士

【摘要】 水下目标探测是发展海洋工程，进行海洋研究与开发的重要手段，一直是海洋探测的重要课题。它不但对国民经济具有很大的推动作用，而且在国家安全上也具有重要的意义。如何解决在水下环境中对目标进行高分辨率成像，一直是海洋领域重要的研究内容。针对这些问题，本文将图像超分辨率重构技术应用于水下图像处理中，以期解决水下图像分辨率较低的问题。图像及视频序列的超分辨率重建，是近年来图像处理领域的一个研究热点，不仅在理论上具有重要意义，在实用中也有迫切需求。本论文的主要工作围绕基于水下降质模型的图像及视频序列的超分辨率重建展开，具体内容包括：(1)水下降质模型的建立。主要讨论了两种水体PSF的建立方法，第一种是理论推导的方法，重点针对非相干光照明方式，通过测得的水体abc参数，建立水体解析式的PSF，其中非相干光照明可建立解析式PSF表达式，而相干光照明目前仍停留在原理分析阶段，无法得到具体的解析式PSF表达式，可作为下一步工作的重点；第二种是实验的方法，由于最终得到的水下数据是PSF和噪声共同作用的结果，所以在构造PSF矩阵前应先进行滤除噪声，然后再利用实验数据获取PSF矩阵，这里针对相干光和非相干光两种不同的照明方式可以获得特定水体的PSF矩阵，用以后续的数据恢复。(2)图像数据预处理。针对水体吸收和前向散射导致图像对比度下降、后向散射导致图像模糊以及光电设备引入的噪声等问题，提出了：(a)在水体PSF指导下的逆滤波消模糊算法；(b)保细节的图像亮度调整算法；(c)基于噪声特征的非线性除噪算法。该预处理部分主要为后续超分辨率重构提供尽可能好的低分辨率图像源；(3)单帧图像的超分辨率重构。主要讨论了三个方面的内容：(a)针对图像中奇异点对重构效果造成巨大影响这一问题，通过引入PCNN的简化模型ICM奇异点快速检测机制对图像中的非高斯奇异点进行检测；(b)针对ICM检测器检测出奇异点的特征，构造了改进的极值中值非线性滤波器，对非高斯奇异点进行处理；(c)在插值重构方面，针对图像中的高频细节和低频平坦区域分别采用不同的插值算法，对传统插值算法进行了改进；(4)多帧图像的超分辨率重构。主要分为两部分：第一部分讨论了频域中的图像配准算法，主要针对频谱部分混叠现象，借助于非混叠低频信息中含高频分量的先验信息进行频域配准；第二部分着重研究了基于改进Keren配准算法的空域重构方法和极大似然估计重构中的改进最速下降法，前者针对Keren算法基于小角度的泰勒级数展开所带来的配准误差，提出了一种基于六参数仿射变换的改进Keren算法，该算法相比原始算法的刚体模型能够在大角度偏移情况下获得更精准的配准效果；后者为超分辨率重构算法实时性的实现提供了一种思路。更多还原

【Abstract】 Underwater target detecting is an important technique for the development of theocean engineering and exploration,also a significant task of the ocean detecting.Itplays a substantial role not only for the civil economy but also for the nationalsecurity.The formation of the super-resolution underwater image is a significant topicin ocean detecting field.In order to solve the lower-resolution underwater image,theimage super-resolution reconstruction technique is applied to the underwater imageprocessing.For decades,super-resolution reconstruction is a hot topic in image andvideo processing which is theoretically important as well as practically urgent inmany fields.This dissertation focuses on the super-resolution of image and videosequence based on the degraded model of underwater image.And the study is carriedout in the following aspects:(1) The establishment of underwater degraded model.There are two methods arediscussed in this part about establishment of the water PSF.The first method is basedon the theoretic organon.For incoherent illumination situation,abc parameters ofwater transmission system are measured and its PSF can be modeled in an analyticalformula,while the PSF can not be achieved when in coherent illumination conditions,which is still in theory research stage.This will be the further research work in thisdissertation.The second means for modeling is experimental method.On theconsideration that the acquired data underwater is determined by water PSF and noise,so the first step is data filtering to compress noise and then model the PSF matrix withthe filtered data.For both incoherent and coherent illumination environment,theirunderwater PSFs are achieved,which can be used for future image restoration.(2) Image preprocessing:in the view of illumination absorption by water andforward dispersion which cause image contrast degradation,and backward scatteringwhich blurs image,also noise caused by photoelectricity,three preprocessing methodsare proposed herein,(a) inverse filtering method to overcome image blurring on thebasis of water PSF;(b) image intensity adjustment with the precondition of imagetexture protection.(c) Nonlinear filter methods based on noise attribute.Thepreprocessing is to offer good SNR image for later super-resolution reconstruction. (3) Super-resolution reconstruction for single frame.Three aspect contents aredebated herein,(a) for the sake that outlier image pixels can cause great aftermath forimage reconstruction,the PCNN simplified model ICM is adopted to detect noneGaussian outlier image pixels;(b)According to the characteristics of these outlierpixels,an improved extremurn-median nonlinear filter is proposed to remove thesenone Gaussian outlier image pixels;(c) traditional interpolation methods are improved,in which high and low frequency image areas are interpolated with differentalgorithm.(4) Super-resolution reconstruction for multi-frames.There are two parts for thistopic:the first part is about image registration in frequency domain,which is on thebasis that none overlapped low frequency parts contain high frequency informationand the registration is carried out in frequency domain.The second part introduces theimproved Keren registration algorithm in spatial domain and improved steepestdescent method for MLE reconstruction.The former improved method is to overcomethe error caused by Taylor series extended in small angle in Keren method,and theimprovement is based on affine transformation with 6 parameters,which can achievebetter registration results than old method when larger angled shift occurred forrigid-body;the later improved method offers a new clue of realization in real time forsuper-resolution reconstruction algorithm.更多还原