【作者】 刘海霞；

【导师】 方欣欣；

【作者基本信息】 内蒙古工业大学 ， 控制理论与控制工程， 2006， 硕士

【摘要】 本文对汽车牌照识别前的汽车图像进行小波去噪研究。首先,建立了汽车图像受到雨点、喷溅泥污染后的数学模型,并对汽车图像加噪。第二,论述了去噪图像质量的评价标准,特别是论述了图像残差及其灰度直方图评价图像质量的方法。第三,选择最优小波基。运用去噪后图像与含噪图像峰值信噪比的增值和去噪后图像保留能量的计算结果,得出结论:(1)对于加入噪声水平为25的零均值归一化高斯噪声的汽车图像,sym4～sym8小波、db8～db15小波、coif2～coif5小波、bior2.8、bior4.4、bior6.8小波去噪效果较好;(2)对于加入噪声强度为0.2的椒盐噪声的汽车图像,sym8、db15、coif5、bior6.8小波较好。第四,选择阈值函数。对加入噪声水平1～60的零均值归一化高斯白噪声汽车图像分别用软硬阈值函数小波去噪,得出了若干种典型小波去噪后PSNR随噪声水平变化的软、硬阈值曲线对比图,图中表明:两曲线存在交叉点(临界值),当噪声水平小于“临界值”时,硬阈值去噪效果好,反之软阈值去噪效果好;对加入噪声强度为0.01～0.6的椒盐噪声汽车图像,分别采用软、硬阈值函数小波去噪,得出了若干种典型小波基去噪后PSNR随噪声强度变化的软、硬阈值曲线对比图,图中表明:两曲线也存在“临界值”,但此“临界值”较小。当噪声强度小于“临界值”时,硬阈值去噪效果好,反之软阈值去噪效果好。第五,研究阈值设置策略,运用bior6.8小波,对含噪汽车图像进行三层小波分解,运用图像残差及其灰度直方图的图像去噪质量分析方法,研究层内各方向细节与各层阈值的设置策略。研究结果表明:第一层小波系数中所含的有用信息少;各层对角细节小波系数所含的有用信息少。提出了分层局部阈值的设定策略,对两种噪声模型加噪的汽车图像去噪结果表明:在有效去除噪声的同时,增强了汽车图像必须保留的特征—水平与垂直的边缘和纹理。更多还原

【Abstract】 The paper study on wavelet denoising for motorcar images before motor’s registrations are recognized.First, I have set up mathematical models of motorcar images blured by raindrop and spurted slurry and appended two kinds of noises to motorcar images.Secondly, I have studied how to evaluate images’quality, especially, how to evaluate a method of images’quality with image residuals and histograms.Thirdly, I have selected of best wavelet functions. Added values of PSNR between noising image and denoising image , as well as saved energy after denoising is applied ,their computing results indicate: (1) Effect that a motorcar image appended normal zero mean gauss noise with level 25 is better by following wavelet functions: sym4～sym、db8～db15、coif2～coif5、bior2.8、bior4.4、bior6.8 ;(2) Effect that a motorcar image appended peper & salt noises with density 0.2 is better by following wavelet functions: sym8、db15、coif5、bior6.8.Fourthly, I have applied a suitable threshold function. After a image appended normal zero mean gauss noises from level 1 to level 60 is denoised by some typical wavelet functions with soft and hard threshold function, two soft and hard threshold curves are drawn ,whose PSNR is changed with noise level. The charts of curves indicate that two curves have one crossing (critical value), when noise levels are smaller than“critical value”; hard threshold function is better, or, soft threshold function is better. after a motorcar image appended peper & salt noises from density 0.01 to 0.6 is denoised by some typical wavelet functions with soft and hard threshold function, two soft and hard threshold curves are made ,whose PSNR change with noise density. The charts of curves indicate that two curves have one crossing (critical value) too, when noise density are smaller than“critical value”, results are as same as the image appended gauss noises.Fifthly, How to set up threshold. by bior6.8 wavelet ,a motorcar image include of noise is decomposited three levels. According to image residuals and histograms, threshold selecting on every details of one level and on every levels is researched. results indicate: there are poor useful informations in every details of first level and in diagonal更多还原