【作者】 邱宇；

【导师】 田逢春；

【作者基本信息】 重庆大学 ， 电路与系统， 2011， 博士

【摘要】 随着数码设备的普及,数字图像已成为人们获取信息的主要手段。然而在图像获取、处理、压缩、传输、存储以及复制的过程中,不可避免地会引入噪声,从而降低图像质量。图像去噪的主要目标是滤除其中的随机噪声,同时尽可能地保留图像细节信息和避免添加滤波失真。双边滤波由于其算法结构简单,计算复杂度低且易于工程实现等特点,近年来得到了广泛的关注。本论文在分析双边滤波(BF)内在特性的基础上充分挖掘双边滤波的潜力,通过与小波分析、主成份分析(PCA)技术以及图论等结合,对算法进行改进,使之适合于不同类型图像、不同类型噪声去噪以及对多模图像的图像锐化。具体的研究工作可概括如下:针对灰度图像的高斯型随机噪声去除,提出了一种自交叉双边滤波算法。借鉴交叉双边滤波算法中灰度测度权重在参考图像中计算这一思想,将带噪图像首先通过预滤波器得到预降噪图像,并令其作为参考图像计算灰度测度权重,再在原始带噪图像上面运用交叉双边滤波去噪。理论分析和仿真实验结果表明,预滤波器采用转换域去噪算法时,能很好地克服双边滤波和转换域去噪算法在噪声去除和伪像抑制方面的内在缺陷,最终结果在客观评价指标方面和主观视觉质量方面不仅高于原始BF算法,同时也高于预滤波器的输出参考图像。在比较离散小波变换和非下采样小波变换后,本文采用非下采样小波阈值去噪作为预滤波器。在不考虑计算时间的场合下,还可采用曲线波阈值去噪作为预滤波器,并基于非邻域均值滤波(NL-means)算法中图像子块相似性的思想,以图像子块间相似性代替单点像素间相似性的自交叉双边滤波算法能更进一步提高图像峰值信噪比(PSNR)和消除曲线波阈值去噪所产生的划痕状伪像。针对彩色图像和多模图像的高斯型随机噪声去除,提出一种结合PCA的多模图像自交叉双边滤波去噪算法。利用PCA,首先提取出所有图像分量的主成份。由于主成份是各分量的一个线性组合,因此主成份分量所含噪声低于各原始分量。再对主成份分量施加预滤波,得到参考图像,继而对所有分量进行交叉双边滤波得到最终降噪输出。实验表明,各种预滤波器中非下采样小波(UWT)阈值去噪适用于大多数的多模图像,在PSNR和主观视觉方面均能获得满意的结果。常见的噪声类型除加性高斯白噪声以外,还有脉冲噪声。针对脉冲噪声中的一般形式——随机值脉冲噪声和加性高斯噪声的混合噪声模型,提出一种基于谱图理论的图像去噪滤波器。从数据聚类角度出发,利用图的邻接矩阵,基于双边滤波中几何测度权重的思想,对图的每个节点赋予权重,构造新的内聚度方程,计算各像素点对于主簇的隶属程度,最后输出邻域窗口像素主簇的质心得到去噪后的图像。算法性能还可通过滤波器叠加的方式得到进一步提升。实验表明,算法在PSNR值和主观视觉质量上优于现有的混合噪声去噪算法。除图像去噪以外,图像锐化也是提高图像质量的方法之一。通过在非锐化掩模(UM)中引入不同形式的灰度测度权重,对非锐化掩模实施非线性化处理,实现了对强弱边缘的不同滤波操作。针对多模图像的特点,指出了多模图像的锐化目标,提出了一种针对多模图像的边缘保存交叉锐化算法。实验结果表明,算法具有很好的主观视觉质量,而且在增强细节的同时不会增强噪声和产生光晕现象。针对带噪多模图像,采用先去噪、再锐化、再去噪的3级滤波器级联方式进行增强。去噪方面将自交叉双边滤波结合到Dual Bilateral Filter (Dual BF),提出一种应用于多模图像去噪的交叉Dual BF算法。仿真实验表明,该级联滤波器能够充分利用多模图像的特点,有效地滤除噪声和增强图像边缘,并消除围绕边缘的噪声扰动。在采用相干光源照明的常规4f光学系统中,输出图像极易受到镜头及CCD上的灰尘污点的影响从而造成图像降质。针对该问题,将图像降质原因划分为加性随机噪声、污点、光源不均匀性影响以及系统的低通特性,从而提出一种简化系统模型。基于该模型,利用系统输入全白图像时的输出结果作为先验信息,在假定一次实验中污点和光源保持不变的基础上,确定其分布。基于该先验信息,提出一种邻域区域自适应的双边滤波算法,可以实现同时去除噪声和污点的目的。基于光学实拍图像和人工合成图像的实验表明,算法能够在保持图像细节的同时较好地去除噪声,进而恢复图像,实现主观视觉质量和PSNR值的提高;同时在污点污染严重的情况下算法仍然具有较好的鲁棒性。更多还原

【Abstract】 With the popularization of digital equipment, digital images have become the main means of access to information. However, digital images are often corrupted by noise during acquisition, processing, compression, transmission, storage, and reproduction, any of which may result into a degraded image quality. The goal of denoising is to remove the noise while preserving, as much as possible, the signal features and make-up, if necessary, the filtering distortion, also known as artifact.Bilateral filter has received extensive concern in recent years due to its simplicity of algorithm structure, economical computational complexity and easiness of implementation. After theoretical analysis of intrinsic properties of bilateral filter, the potential of bilateral filter is fully exploited. By introducing the wavelet, principal component analysis (PCA) technology and graph theory into bilateral filter, the bilateral filtering based denoising algorithm is then improved in order to be fit for different types of images and remove different types of noise. In addition, the sharpening algorithm for multi-modal images can also be improved by employing the concept of bilateral filtering. The main results obtained in this dissertation can be summarized as follows:For Gaussian random noise removal in gray-scale image, inspired by the principle of cross bilateral filter, in which the radiometric similarity is calculated in the reference image instead of in the noisy one, a joint non-linear filter which can be called as self-cross bilateral filter (SCBF) is proposed. The noisy image is first inputted into a preliminary filter to obtain a pre-denoised image. Then the original noisy image is denoised by the cross bilateral filter using the pre-denoised output as the reference image. Theoretical analysis and simulation experiment results show that if the transform domain denoising algorithm is adopted in the preliminary filter, the inherent defects of bilateral filter and transform domain denoising algorithm are suppressed in the aspects of noisy removal and artifact restraint, and their advantages are enhanced. The final output has higher quality than the outputs of both the original bilateral filter and the preliminary filter both in objective and subjective criteria. After comparing the discrete wavelet transform (DWT) and the undecimated wavelet transform (UWT), we choose the UWT thresholding to be the preliminary filter in this dissertation. In the case of not considering calculating time consumption, the curvelet thresholding can also be used as the preliminary filter and the idea of patch similarity in NL-means algorithm can be introduced into the calculation of radiometric weight to replace the single pixel similarity in bilateral filter for suppressing the scratch shaped artifacts produced by the curvelet thresholding and improving the PSNR value further.For Gaussian random noise removal in color image and multi-modal images, a PCA based SCBF is proposed. First, the PCA is applied to all the components of input image for extracting the principal component (monochromatic image). Since the principal component is the linear combination of all the components, noise is expected to be relatively reduced in this monochromatic image. Next, the principal component is properly smoothed by a preliminary filter and the output is regarded as a reference image. Finally, the SCBF with the smoothed principal component is applied to all the image components to get the final outputs. The experiment results show that among some preliminary filters, the UWT thresholding is useful for effective denoising of various multi-modal images. The results are shown to be satisfactory both from the aspects of PSNR value and visual quality.In addition to additive white Gaussian noise (AWGN), the impulse noise removal is also a common task in denoising. For the removing of mixture of Gaussian and random impulse noise– a general type of impulse noise other than the salt and pepper noise, a graph-spectral method based denoising filter is proposed from the viewpoint of data clustering. By using the adjacency matrix in graph, motivated by the idea of geometric distance in bilateral filter, each node is given a geometric weight, and a new cohesiveness equation is then constructed. By calculating the membership of the cohesiveness equation, the most dominant cluster is extracted from the set of pixels in each window and the centroid of those pixels values is outputted as the denoised image. Denoising ability can be further improved by cascading some filters. Experiment results show that the proposed method outperforms the existing mixture noise removal algorithms in term of PSNR value and visual quality.Except for image denoising, image sharpening is also a commonly used way to improve image quality. By introducing different form of radiometric weight into the unsharp masking (UM), the UM is nonlinearized to achieve different filtering purposes for edges with different extent of sharpness. Based on the characteristics of multi-modal images, the goal of multi-modal images sharpening is discussed, an edge-preserving cross-sharpening algorithm for multi-modal images is then proposed. It is shown by the experiment results that the proposed method can make satisfactory visual quality and enhance the details while not boost up noise and produce halos around large edges. For the noisy multi-modal images sharpening, a cascade filtering framework with three steps of denoising-sharpening-denoising is proposed. In the step of denoising, by combining the SCBF with dual BF, a cross dual BF for multi-modal images is proposed. Experiment results show that the cascade filtering framework can take full advantage of the characteristics of multi-modal images to effectively remove noise, enhance the edges and eliminate the noise disturbance around the edges.In an optical 4f system with coherent illumination source, output images are easily contaminated by dusts and spots in the surface of lens and CCD. With the image degradation factors being classified by additive stochastic noise, spots, illumination non-uniformity and system low-pass characterization, a simplified model is then proposed. Based on this model, the output result is used as a priori information when the input is a full white image to calculate the distribution of spots and illumination non-uniformity of which stabilization is assumed in an experiment. An improved neighborhood shape adapted bilateral filter is therefore introduced to denoise and remove spots. It can be shown by the optical and synthetic image experiments that, the proposed method can effectively reduce noise and restore image with edge sharpness being preserved. The robust characterization of the proposed method is embodied by the simulation results even under serious spot-contamination circumstances.更多还原