【作者】 王伟；

【导师】 陆佩忠；

【作者基本信息】 复旦大学 ， 计算机应用技术， 2012， 博士

【摘要】 图像复原是图像处理领域的重要问题,一直是研究的热点和难点。由于环境条件限制及成像设备的物理局限性,图像在摄取、传输和存储的过程中,不可避免地会受到噪声污染和模糊化,使获得的图像质量下降。为得到清晰图像,基于滤波、正则化和微分方程的图像复原方法目前发展迅速,研究成果突出,但针对高密度冲击噪声和混合噪声的滤波算法效果仍然不够理想,现有的图像去模糊和图像增强算法对边缘和细节保持能力还难达到完全满意的效果。本文主要采用基于微分的方法,首先概述了图像复原的数学模型和经典算法,然后针对高密度冲击噪声去除问题、混合噪声滤波与通用噪声滤波问题、图像去模糊过程中的细节保持问题和图像增强问题分别展开研究,具体取得了以下结果：1)研究了高密度冲击噪声污染图像的复原问题,提出了基于局部离群因子的滤波算法。该算法首次将数据挖掘领域的局部离群因子概念用于图像冲击噪声的识别和检测,结合边缘区分噪声检测算法BDND和有向加权中值滤波,提出了LOFBDND算法。该算法对于高密度冲击噪声检测准确率很高,滤波后复原的图像在客观指标评价和主观视觉质量上都优于现有经典算法。2)研究了混合噪声污染图像的复原问题,提出了基于偏微分的自适应通用滤波算法。该算法采用加权图像局部统计信息计算局部差分因子LDF,并利用LDF进行图像噪声的识别和检测,然后将LDF加入P-M各项异性扩散模型中,结合图像局部梯度信息进行扩散函数的自适应控制。该方法对椒盐噪声、冲击噪声、高斯噪声和高斯-冲击混合噪声都有很好的滤波效果。3)研究了模糊图像的复原问题,提出了基于分数阶微分的图像去模糊算法。该算法根据分数阶微分掩模算子对模糊核的不敏感特性,将分数阶微分作为图像复原处理中的一个正则化项,实现对图像复原病态问题求解过程的正则化限制,并结合全变分方法,建立基于分数阶微分和全变分的去模糊模型FDTV。FDTV结合了分数阶微分和全变分在图像处理领域的优点,复原的图像不仅有更清晰的轮廓、边缘和细节,还对图像有一定的锐化增强。4)研究了降质图像的对比度增强问题,提出了基于分数阶微分的多尺度域图像增强算法。该算法先将图像在多尺度域中进行分解,构建图像金字塔,然后采用分数阶微分对图像金字塔层进行自适应增强,最后利用增强后的金字塔层重建图像以实现对比度增强。该方法增强后的图像在客观指标评价和主观视觉特性上都优于直方图均衡、频域图像增强和非线性滤波等传统算法。更多还原

【Abstract】 Image restoration is a key problem in image processing, and it has been the most difficult and focus question. Because of the environmental restraint and equipment physical limitations, image will suffer noise pollution and blur during image acquisition, transmission and storage. Noise and blurring will lead to image degradation. In order to get clear recovery image, lots of methods have been proposed. Recently, filtering method, regularization methods and partial differential equation (PDE) method have been quickly developed with good results. Otherwise, the result of high-density impulse noise removal or universal noise removal is not good enough. Despite high effectiveness in image deblurring and image enhancement, most of the existing algorithms tend to over smooth the image details.In this paper, the proposed methods are mainly based on differential. Firstly, mathematical model and typical methods of image restoration are overviewed. Subsequently, some new algorithms aimed to high-density impulse noise removal, universal noise removal, image deblurring and image enhancement with good edges and image details preservation are discussed respectively. The main results are given as follows:1) An efficient switching median filter aimed on high-density impulse noise removal is proposed based on Local Outlier Factor (LOF). LOF, which is popularly used in data mining, is firstly used in impulse noise detection and discrimination. LOFBDBD algorithm is designed by using boundary discriminative noise detection (BDND) combining with directional weighted median filter. LOFBDBD has high noise detection accuracy even if the noise level is above50%and provides better performance than many other median filters for noise image restoration.2) A novel method is presented for universal noise removal based on PDE. Local Difference Factor (LDF), which is computed locally from intensity values of image pixels in a neighborhood using weighted statistics, is used in noise identification. Furthermore, LDF is added in classic P-M model to control the diffusion process adaptively incorporating with local gradient. This method has great image quality by efficiently removing salt-and-pepper noise, impulse noise, Gaussian noise and mixed noise.3) A new Fractional Differential (FD) based image deblurring approach is presented. At first, the application of FD in image edge detection and image enhancement is discussed. Secondly, based on FD mask’s insensitivity to various blurring kernel, FD is adopt as an additional regularization term in image deblurring cost function to restrict the solution of the ill-posed image restoration problem. FDTV model is built by combining FD and total variation (TV) constraint. Because of comprehensive the advantages of both FD and TV in image processing, the recovery images by FDTV have good image quality, such as clear contour and edge, rich details and a certain degree of sharpness enhancement.4) A Fractional Differential (FD) based image enhancement method in multi-scale domain is presented. Firstly, the image is decomposed according to the Laplacian pyramid transform. Then, FD is adopted to enhance image pyramid levels. The enhanced image is got by image reconstruction using enhanced pyramid finally. Experimental results show that proposed image enhancement method is markedly superior to many other popular image enhancement algorithms in subjective visual effects and some objective quality assessments.更多还原