【作者】 王晓甜；

【导师】 石光明；

【作者基本信息】 西安电子科技大学 ， 电路与系统， 2011， 博士

【摘要】 卫星遥感成像是对地观测的重要手段之一,在国家安全、国民经济、科学研究和人民生活等方面有重要作用。卫星遥感利用地物目标的辐射特性以及对电磁波的反射特性,实现对地表信息的间接获取。遥感卫星图像需要真实、客观的反映地物特征,然而,由于星载成像系统处于复杂的电磁环境,卫星内部电路系统众多,CCD器件的工艺缺陷等,在获取和传输图像的过程中会引入大量噪声,严重降低对地观测信息获取的图像质量,给后期的图像处理和人眼视觉判读带来障碍。因此,对卫星成像系统获得的遥感图像进行数据预处理,降低成像系统中引入的各种噪声,提高遥感图像的质量和准确度,为后期压缩编码、数据传输、判读解译、目标识别等工作能够顺利进行并取得良好结果提供前提和保证,是获取遥感图像后需要完成的首要工作。在星载成像系统获取图像的过程中,实际系统受到的噪声干扰多种多样,不同的环境干扰和系统缺陷会对遥感图像引入分布和统计特性截然不同的噪声,例如成像过程中受电磁干扰产生的服从高斯分布或泊松分布的热噪声、由于线阵CCD像元响应不均引起的具有一定方向性的条带噪声和由于成像系统的故障和缺陷及电气开关、继电器改变状态引起的脉冲噪声等。这些噪声具有明显不同的特征,对数据产生的干扰在图像中的表现形式也不同,使用一种降噪算法难以对它们进行有效的抑制。例如能够有效抑制高斯噪声的小波软阈值降噪算法在条带噪声和脉冲噪声面前就显得束手无策。因此,处理不同种类的噪声,必须针对它们各自的机理和特性分别加以消除,才能真正达到抑制图像噪声,提高图像质量的效果。另一方面,遥感图像信号有其自身显著的细节丰富、边缘和纹理方向性显著等特点,图像降噪的本质就是寻找信号和噪声之间的不同特征并利用这些特征将其分开。国内外目前还没有分析遥感图像的信噪特征,并利用这些特征针对不同噪声采用不同方法进行降噪的专门文献。本文在分析总结星载成像系统工作机理的基础上,探讨卫星遥感图像信号和噪声的源特性,并利用各自不同的源特性,将信号和噪声分开。从实际系统角度出发,对不同类型噪声分别加以处理。主要工作内容如下:1.星载可见光成像系统中噪声的来源和特性分析:在分析噪声产生机理的基础上介绍了成像过程中光电散粒噪声、读出噪声、条带噪声和脉冲等的来源和特点。对存在于遥感图像中的噪声进行分类,总结和探讨了泊松噪声、高斯噪声、条带噪声和脉冲噪声等各类噪声的统计特性和数学模型。2.线阵扫描CCD条带噪声抑制方法研究:分析研究线阵沿轨扫描CCD传感器内部每一个像元在光谱响应区内的响应函数不一致造成的遥感图像条带噪声。在分析遥感CCD图像孤立条带噪声产生原因和噪声模型的基础上,提出基于二维方向滤波器抑制条带噪声的方法。该方法利用窄带方向滤波器良好的图像纹理方向的频率选择特性,将条带噪声与图像其它信息分离到不同的方向滤波器子带内,并在水平信息子带内采用均值补偿方法去除条带噪声。实验结果证明它不仅能有效去除CCD图像中的随机孤立条带噪声,还能保持原始图像的辐射入射幅度,保留细节信息。较传统的方法相比,性能有较大的提高。3.基于信噪特征的图像高斯白噪声抑制方法研究:分析研究了成像过程中以读出噪声为代表的高斯白噪声特征,并利用遥感图像富含边缘和纹理等具有方向信息的二维数据特征,提出基于自适应方向提升小波变换的遥感图像降噪方法。该方法构建于传统提升小波变换框架之上,分析图像的局部方向信息,并使每一个提升步骤都沿着图像局部像素相关性最强的方向进行变换,实现局部最大去相关性,将图像中边缘和纹理产生的高频信息尽可能的压缩到低频子带,从而克服了传统小波缺乏灵活的方向选择性的缺点。方向提升小波将噪声产生的无方向性的高频信息留在高频子带,从而实现将图像的纹理和噪声在小波域更好的分开。通过阈值策略对高频系数进行处理,就可以实现在降噪的过程中尽可能的保留图像边缘信息,这一特点在富含边缘、纹理和小目标的遥感图像降噪中非常重要。4.基于两级插值的图像脉冲噪声去除方法研究:分析研究星载成像系统脉冲噪声的特点,提出一种基于非均匀下采样和分段自回归插值的去除图像脉冲噪声的新方法。根据脉冲噪声的特点:由随机非连续,幅度大的不规则脉冲或噪声尖峰组成。通过检测算法可以检测出未受脉冲噪声影响的像素。将未受噪声污染的像素点从含噪图像中提取来构成一幅不含脉冲噪声,分辨率为原来1/2的低分辨率图像。再通过分段自回归插值将低分辨率图像插回原始图像大小,得到恢复信号。该方法首次将下采样和插值的概念引入脉冲噪声抑制方法,充分利用未受噪声影响的像素和二维图像结构信息,恢复清晰图像。更多还原

【Abstract】 Remote sensing technology is a key to conduct the observation towards ground objects, which proclaim its importance in many areas, such as national security, national economy, scientific research, life of individuals, and so forth. Satellite remote sensing systems receive electromagnetic wave radiation to reflect the real feature of ground objects through indirect observation accurately. However, enormous noise exists during image acquisition and delivery due to the complicated electromagnetic environment out side space, along with the numerous circuit systems inside the satellite and the defects of CCD. It is well known that the existence of noise strongly influences the quality of observation results which surely lead great difficulties to the incoming image processing or manual interpretation. Therefore, effective remote sensing image denoising is a very important preprocessing schedule before anyother image processing can be carried out efficiently.During the acquisition of remote sensing images, there are many reasons which result in different kinds of noise. The characteristics of noise caused by different reasons can be quite distinct from each other. The thermal noise caused by electromagnetic interference or temperature changes obeys Gaussian or Poisson distribution. The striping noise caused by line array CCD image input system response inequality lies in certain direction in the image. The impulsive noise introduced by defection and malfunction of imaging system or state alternation of electric switches and relays is a kind of extreme disturbance, which can totally erase part of the image pixels. Different kinds of noise have distinct features and they behave differently in noised images. It is impossible to handle different kinds of noise by any individual algorithm. For example, the wavelet based shrinkage algorithms, which have been proved to be quite efficient in suppressing Gaussian noise, become invalid in front of striping noise or impulsive noise. In this case, we must research each noise characters and mechanisms individually to pursue good denoising results. Moreover, the key factor lies in any successful denoising method is to find the different characteristics between signal and noise. As for remote sensing image, the obvious characteristic is the abundant detail information and 2-D directional information. So far, there were no such specified documents to research the noise suppression method based on synthetic analysis towards the characteristics of different kinds of noise.In this paper, based on the exploration of working mechanism in satellite remote sensing imaging system, the distinct source characteristics between remote sensing image and noise are discussed. Then, according to diverse characteristics of different noise, respective denoising methods are proposed for each one of them. The main work of this dissertation can be summarized as follows:1. Satellite imaging system noise source and character analyzing. This paper introduces, summarizes and probes into the distinct source and characteristics of different kinds of noise imported in imaging acquisition and delivery, including Gaussian noise, striping noise, and impulsive noise.2. Destriping method researching in linear array CCD system. This chapter analyzes striping noise caused by unequal response function of each pixel in linear array CCD sensors. After discussing the characteristic of random striping noise, a 2-D directional filter based de-striping method is proposed. By utilizing the outstanding frequency selecting feature of the narrow channel directional filter, this method divides striping noise into another sub-band aside from image signal and point-wise noise, then use average compensation algorithm in horizontal sub-band to remove striping noise. Experiment results prove that not only can it remove isolated random striping noise, but also can it retain detailed information in CCD image.3. Zero-mean Gaussian noise suppression research. Based on Signal & Noise characteristic analysis in remote satellite images, a Gaussian noise suppression method based on adaptive directional lifting (ADL) is proposed by means of utilizing the rich directional information induced by textures and edges. The proposed method is constructed based on the scheme of traditional lifting scheme. By integrating local direction information in each lifting steps, ADL compresses most image high frequency information induced by edges and textures into low frequency sub-band. This procedure can overcome the drawbacks of traditional wavelet on lacking of direction selecting elasticity. Since ADL can compress most edge and texture information into low frequency sub-band, leaving only point-wise noise in high frequency sub-band, the thresholding processing can effectively remove noise while retain image information.4. Impulsive noise removal method research. Analyze impulsive noise in satellite imaging system, propose a nonlinear down sampling and subsection auto-regressive interpolation based impulsive noise removal algorithm. At the first step, this algorithm constructs a half resolution image without impulsive noise from the noised image by means of non-uniform down-sampling. After that, the piecewise auto-regressive interpolation method is used to restore low resolution image into the original size. The proposed algorithm innovates in introducing down sampling and interpolation concept into impulsive noise removal, which fully utilizes every uncontaminated pixels and two-dimensional structural information wihtin to achieve better denoising results.更多还原