【作者】 吴静；

【导师】 王珂；

【作者基本信息】 吉林大学 ， 信号与信息处理， 2005， 硕士

【摘要】 论文以去除噪声、保留图像的边缘特征点为出发点,以提高PSNR、重构图像清晰度为目的进行研究。由于小波变换后的图像能量主要分布在低频部分,噪声基本上分布在高频部分,而图像的边缘信息是图像最为有用的高频信息。因此,利用传统的方法去噪时,虽然能较好的去除图像中的噪声,但不能较好的保留图像的边缘信息。另外,全局阈值法有“过扼杀”小波系数的倾向,使图像的信息被去除。针对传统去噪方法不能保留边缘特征和全局阈值法“过扼杀”小波系数的缺点,论文首先对图像进行边缘检测。由于在去噪方面采用了小波阈值去噪方法,为了更好的利用小波变换的特点,把小波边缘检测应用于图像去噪,通过小波边缘检测方法确定边缘特征点的位置。在对小波变换后的各高频子带进行阈值处理时,非边缘特征点进行阈值处理,而边缘特征点所在位置的小波系数不受阈值去噪的影响,在阈值去噪时保持不变。其次在阈值选择上,论文采用了基于Bayesian风险估计推出的阈值公式,计算得到的阈值进行去噪的。理论分析和实验结果表明,与传统的去噪方法相比,本文方法能较好的保留图像的边缘信息,且提高了图像的峰值信噪比。更多还原

【Abstract】 An image is often corrupted by noise (refers to additive white Gaussian noise in this paper) in its acquisition. The goal of denoising is to remove the noise while retaining as much as possible the important signal features. So, in order to increase PSNR and definition of the constructed images, image denoising and edge information maintained are first considered. According to this, idea of image denoising of this paper is presented: image thresholding denoising based on edge detection and Bayesian estimation. The method put forward is depended on these: (1)Traditionally, the approaches of denoising are achieved by linear processing such as Wiener filtering, which can restrain the high-frequency parts and decrease noise but may destroy image details. They have many defects to unstatic signals. (2)Nonlinear techniques are used in image denoising. After wavelet decomposing, most of the energy of the original image is concentrated in the low-frequency subband, while noise is distributed in high-frequency components. Edge information is high-frequency information, so traditional thresholding shrinkage will throw off the useful edge information as cutting off those high-frequency noises, which blur sharp edges. Those wavelet coefficients of an image that are corresponding to image’s edges are first detected by the method of wavelet edge detection. The detected wavelet coefficients will be protected from denoising. (3)It is important to choose thresholding. Global thresholding can throw off the important information of image, while local thresholding with locality is calculated by local components, which provided better performance on denoising. Ideal thresholding from thresholding Eq. is achieved just depended on Bayesian risk estimation. There are two stages of the procedure which operates as follows: Wavelet edge detection is in first stage. (1) Smooth A part of the noise is cut off through smoothing the noisy image, which may reduce the calculation as edge detection. The details of the image are misty ,which was created by linear filter , while it can be overcame by median filter. At the same time, median filter is adapt to cut off the isolated noise. Median filter which size of the window is 3×3 is adopted in this paper. (2)proceeding the image by wavelet edge detection According to the method of the wavelet edge detection expressed in chapter 4, image by pre-proceeded is detected through multi-scale wavelet edge detection. The locations of edge are confirmed, and the coefficients of the edge are maintained, while others are proceeded through the wavelet thresholding. The thresholding procedure is in second stage. (1)denoising image of the wavelet transform Denoising image is decomposed through two-dimensional WT to obtain low-frequency subband and detail subbands. The energy of the noise reduces 90% when the scale is 3. (2)Every threshold of denoising of detail subbands is computed, according to T? = σ? 2σ?X based on Bayesian estimation. According to Eq. (5.18-5.22), every parameter is calculated. (3) wavelet-thresholding denosing processing To get rid of the noise in the detail subbands, wavelet coefficients of non-edge are processed by Eq. obtained from (2). While wavelet coefficients of edge are not processed by thresholding. According to this, wavelet coefficients are obtained. According to the comparison between hard-thresholding and soft-thresholding, soft thresholding is applied in this paper. ??ω? j , k= ??? 0si gn (ωj,k)(ωj,k?T?)ωωjj ,,kk≥<TT?? (4)Through inverse wavelet transform, denoised image can be reconstructed.ABSTRACT Sum up, the main research works of the paper are as follows: (1)It is introduced that the methods of the image edge detection. In view of the multi-scale features of wavelet transform, wavelet edge detection is prior to the traditional methods, according to the comparison of simulation between traditional and wavelet edge detection. (2)It is analyzed that traditional wavelet denoising algorithms (thresholding shrinkage and several thresholding functions). Global thresholding can throw off the important information of image, while local thresholding with locality is calculated by local components, which provided better performance on denoising. Threshold Eq. is cited depends on Bayesian risk estimation. (3)According to existing problems, the method of this paper is presented. It is proved that the method is valid by the simulation. Several conclusions are drawn as follows: (1)In view of the multi-scale features of wavelet transform, wavelet edge detection can maintain image edge features over the traditional methods. The latter can detect the edge which is discontinuous and cannot handle the details. The former can delineate the better details (such as human’s eye pupils, hat edge, and pillar etc.). The method can reserve the image edge when cutting off the noise. It is value to important edge information. (2)Local thresholding procedure removes noise by thresholding only wavelet coefficients located non-edge, while keeping the coefficients of the edge. (3)The denoising algorithm of this paper is valid through simulation experiments. After “Woman”and “Lena”image processed, the results are compared to other denoising algorithms. It is found that image contour is clearer. After denoising, the method can increase PSNR up to 1~ 2dB. It is proved that the method of the paper is valid.更多还原