【作者】 唐斌兵；

【导师】 王正明；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2009， 博士

【摘要】 不改变现存低分辨率成像系统,通过多幅互有位移、信息互补的低分辨率图像获取高分辨率图像的超分辨重构技术,一直是图像处理领域中研究热点之一,在刑事侦察、目标识别、医学诊断等领域具有广泛的应用。本文从图像超分辨机理分析入手,围绕图像的去噪、配准、重构模型及算法的设计,做了以下几方面的工作:分析了成像系统的降质原因,讨论了成像系统的模糊辨识问题,针对实测数据给出了一种有效估计系统降质点扩展函数的方法;分析了椒盐噪声的稀疏特性,以解的稀疏性为先验信息,给出了图像去噪的扩展模型和相应的求解算法。对于含椒盐噪声的图像,该模型具有显著的去噪效果,且能有效地保持图像特征。分析了运用子图集代替原图做配准的优势,提出了一种选取子图的方法,并对子图集进行了有效筛选。简化了相似性测度计算公式,分析了小波系数的特点,依据小波变换的多分辨分析特征,提出了一种图像由粗到精、逐步迭代的快速配准算法。该方法在稳定性和计算时间上,明显优于传统方法。基于寄生波纹先验提出了一种能抑制寄生波纹的超分辨重构方法。分析了寄生波纹产生的原因,针对配准误差、系统降质点扩展函数的估计误差以及噪声等引起的寄生波纹,采用增加寄生波纹惩罚项的方法,建立了图像超分辨重构的正则化模型。根据寄生波纹对图像不同区域的影响,给出了一种能够抑制寄生波纹的参数选取方法。相对于传统方法该方法在不影响图像边缘重构效果的基础上,能有效地抑制寄生波纹。提出了利用先验图像的灰度分布信息来进一步改善图像质量的超分辨率重构方法。分析了现有超分辨重构方法存在的不足,介绍了两种构造先验图像的方法。利用最小鉴别信息原理,引入先验图像信息作为超分辨重构的信息补充和约束依据,由此构造新的超分辨重构模型。基于置信策略,提出了一种选取正则化参数的方法。当选取的先验图像灰度分布信息与真实图像的相似时,其重构效果能够得到较大程度的提高;同时,相对传统的先验图像约束方法,该方法提出的灰度分布特征约束更加简单有效。更多还原

【Abstract】 Super-resolution image reconstruction from multiple inter-displacement and information-complemented low-resolution images is one of the research focuses in the image processing area, since a better-resolution image can be obtained based on this technique without changing current low-resolution imaging systems. It has comprehensive applications in the domain of penal reconnaissance, target recognition, medical diagnosis and so on.Starting with analyzing the degradation mechanism of images, this thesis focuses on image denoise, image registration and super-resolution reconstruction models and algorithms with the constraint of prior information. The main work is as follows.Firstly, this paper analyzed the degradation mechanism of imaging systems. The fuzzy identification issue of the imaging system is discussed. An effective method of estimating the system degradation point spread function is introduced based on real measurement data.Secondly, The characteristics of the salt and pepper noise was analyzed. The solution sparseness acted as the prior information. A new extended model was proposed with corresponding algorithm. Experiments showed that the model has a significant denoising effect on images with salt and pepper noise while maintaining important features of the images.Thirdly, the paper proposed a fast image registration algorithm based on wavelet transform. A method was given to select the sub-map effectively by using the efficiency sub-maps in place of the whole map to carry out registration. The t-test method was employed to filter the sun-maps. The formula to measure the similarity was simplified, and the wavelet coefficients features was also analyzed. Based on the multi-resolution analysis features of the wavelet transform, a fast registration algorithm using layered and iterated strategy was proposed. This method is proved to be much better than traditional methods in the aspects of both the accuracy and the calculation capacity.Fourthly, this paper proposed a method of super-resolution reconstruction using parasitic ripple as a priori information. The mechanism of parasitic ripple generation was studied. Based on the parasitic ripple caused by registration error, the PSF estimation error and noise, a super-resolution model was adopted by adding a penalty item of parasitic ripple which can restrain the parasitic ripple. In accordance with the effect of the parasitic ripple on different regions, a method of calculating the adaptive regularization parameter with which to suppress parasitic ripple was given. Compared with traditional smoothing-bound method, the new method can suppress the parasitic ripple without affecting the reconstruction of the edge of the image. Finally, the paper proposed a method of super-resolution reconstruction to enhance the image quality with the grey level distribution information of the prior image. The defects of the current super-resolution reconstruction methods were analyzed. Two methods of constructing a priori image were introduced. A priori image was applied to restraint the results of the super-resolution images. A new super-resolution model was constructed using the principle of minimum identification information to obtain a reconstructed image. The strategy based on confidence level gave an adaptive scheme to select the regularization parameters. When the grey level distribution information of the prior image is more silimar to that of the real image, the reconstruction results can be improved to a greater extent; at the same time, compared with traditional methods , the grey level distribution characteristics restraint proposed by the method mentioned above is proved to be simpler and more effective.更多还原