【作者】 丁海勇；

【导师】 卞正富；

【作者基本信息】 中国矿业大学 ， 地图制图学与地理信息工程， 2008， 博士

【摘要】 图像超分辨率重建是指利用同一场景的不同观测角度、不同观测时间或不同传感器获取的低分辨率图像序列重建出一幅或多幅高空间分辨率图像的技术。由于在图像获取过程中会因为大气扰动、目标景物和传感器之间的相对运动以及传感器光学系统自身因素的影响,图像中存在模糊、附加噪声等质量退化现象,使得图像的空间分辨率较低。超分辨率技术能够融合多幅图像中存在的信息,重建出质量较高的高分辨率图像。重建过程分为图像配准、插值、模糊函数估计、重建计算等步骤,本论文对图像序列配准、图像复原、图像中的附加噪声、重建算法等问题进行了研究。图像序列的配准方法进行了讨论,基于配准算法的基本理论提出了利用Powell优化算法和图像的多分辨率分解相结合的配准参数估计方法,首先对待配准图像和参考图像进行多分辨率分解,得到图像的金字塔表示,在最粗层上给出配准参数的初始估计值,然后利用Powell优化算法估算出该层上的配准参数值作为上一层的初始值,这样一直进行下去,在原始图像中估计出配准参数值。数值实验表明该算法能够准确估计预先设计的配准参数,对噪声和模糊退化干扰具有一定的稳健性。在对图像模糊函数分析后论文提出了基于神经网络的图像自适应复原算法,当模糊退化过程已知时,该算法能够估计出较好的图像复原效果。论文还讨论了图像中的噪声为泊松噪声和脉冲噪声时的图像超分辨率重建问题,分别建立了重建模型,利用数值实验说明了算法的有效性。对小波分析理论及其在超分辨率重建中的算法进行了研究,提出了小域多通道超分辨率重建算法,利用数值实验对算法进行了验证。基于小波系数的统计模型改进了已有文献中的小波域超分辨率重建算法。对遥感图像超分辨率重建问题进行了讨论,利用ETM+图像中的全色图像作为辅助图像,对多光谱图像进行了重建。首先建立重建模型,讨论了模型中参数的估计方法,提出了利用已知图像的局部方差对正则项系数进行估计的算法。将观测到的多光谱图像作为低分辨率图像序列,认为是由对应的高分辨率多光谱图像序列经过模糊变换、下采样和附加噪声后得到的,高分辨率图像为观测到的全色图像,可以看作是待估计多光谱图像序列的线性组合后得到的。将该算法进行了数值实验,并将实验结果与主成份图像融合及小波方法图像融合结果进行了比较,说明该算法是切实可行的。最后通过实例对超分辨率重建的不确定性问题进行了讨论。更多还原

【Abstract】 Image super-resolution (SR) reconstruction refers to the techniques to reconstruct one or more images with high resolution (HR) from the low resolution image sequence which was taken about the same scene from different viewpoint, different time or from different sensors. Due to the influence of air turbulence, the relative motion between the object and the sensors and the influence of optical systems, the images obtained are unavoidable degraded by blurring, additive noises which decrease the spatial resolution. SR technique can be used to fuse the information contained in the image sequence to reconstruct a HR image with higher quality. The process of SR reconstruction can be seen as the following steps: image sequence registration, interpolation, blur function identification and reconstruction algorithm. In this thesis, the author did some work on the following problems: image registration, image restoration, image reconstruction under different noise model, remote sensing image SR reconstruction algorithm.Firstly, the image registration problem was discussed and a new registration algorithm based on the Powell optimal algorithm and the image’s pyramid decomposition was proposed to estimate the registration parameters. The images were decomposed using Gaussian pyramid method and then the initial registration parameters were initialized at the bottom level, then the Powell algorithm was used on the object function to estimate the parameters in this level. The estimated values were then used at another level as the initialized values to estimate the new parameters. This process was carried through at the finest level and the parameters were estimated finally. Numerical experiment shows that this algorithm can estimate the predetermined parameters with higher accuracy.Then, the blur function was discussed and after that an image restoration algorithm based on adaptive neural network was given. When the blur function was known, this algorithm could give better restoration results. This thesis also discussed the SR reconstruction algorithm when the LR images were degraded by different noise models such as Poisson noise and impulse noise. For different noise model, the reconstruction formulas were given and numerical experiments were made to show the performance of these algorithms.In this thesis the wavelet transform method was discussed and the SR algorithm based on the wavelet theory was given. Multi-channel SR algorithm was presented and the numerical experiments were given. Based on the statistics model of wavelet coefficients, the modified SR algorithm was proposed.Finally, this thesis discussed the problem of remote sensing image SR reconstruction. The panchromatic image in ETM+ was used to reconstruct the multi-spectral images. The reconstruction model was established and the parameter estimation algorithms were discussed. A new adaptive regularization parameters estimation algorithm based on the local variance of the images was proposed. The observed multi-spectral images were seen as the low resolution image sequences which were the degraded versions of the corresponding multi-spectral images with higher spatial resolution. The degradation process was blurring transform, under sample and additive noise. The panchromatic image was seen as the high resolution observed image which can be seen as the combination of unknown multi-spectral high resolution images. The task of SR reconstruction for remote sensing images was to merge the spatial information of the panchromatic image into the multi-spectral images. Numerical experiments were made to verify the performance of this algorithm and the reconstructed results were compared with that of principal component analysis fusion algorithm and wavelet transform fusion algorithm which demonstrate validity of the proposed algorithm. Finally, the uncertainty problem of SR reconstruction was discussed using some examples.更多还原