【作者】 谢鹏；

【导师】 姜黎；

【作者基本信息】 湘潭大学 ， 集成电路工程， 2019， 硕士

【摘要】 随着各种数字仪器以及数码产品的普及,图像和视频已经成为人类生活中最常用的信息载体。生活中一幅简单的图像可以包含着大量的信息,这也使得图像成为了人们获取外界原始信息的主要途径。然而图像在获取、传递和保存的过程中会不可避免地带入各种噪声,从而使得图像质量降低。被噪声污染的图像不仅仅会影响人们的视觉体验,还会对后续的图像处理造成极大的影响。图像去噪在图像处理中仍然具有很大的挑战性,很多研究者都在思考如何才能在更彻底地去除噪声,并且最大程度地保留原始的图像信息。本论文主要研究了卷积神经网络在图像去噪中的应用,主要工作内容包括:（1）设计了一个基于深度卷积神经网络的图像去噪模型。为了更深层次地提取图像的特征,同时解决随着网络深度增加导致的学习效率低下的问题,本文引入了残差网络。目前大多去噪方法是通过训练不同的模型来处理不同强度的噪声,这样就需要先预估噪声强度,然后再选用与噪声强度相对应的模型来进行去噪处理。本文使用一个训练好的MLP^[1]模型来对“Lena”进行测试,在标准差σ分别为10,25,50和70的高斯白噪声下,取得的峰值信噪比（Peak signal to noise ratio,PSNR）分别为28.87dB,31.28dB,17.66dB和13.82dB。从这样的一个结果中可以看出,只在一种噪声强度下训练出来的模型不能很好处理其他强度的噪声。本文设计的卷积神经网络模型使用不同噪声强度的数据训练,能同时兼顾处理不同强度的噪声。（2）本文去噪方法将卷积神经网络与平稳小波变换结合,在去噪同时能更好地保留原始的图像信息。目前许多深度学习去噪算法都利用卷积神经网络强大的非线性拟合能力,去学习含噪图像到无噪图像之间像素级的映射。但是在处理高强度噪声时,会产生过于粗糙的结果,并且会损失原始图像的纹理细节。为提高在高噪声下的去噪效果,本文提出了一种神经网络和平稳小波变换相结合的图像去噪方法。不是直接预测无噪图像的像素,而是先预测无噪图像的小波系数,然后通过预测的小波系数重建去噪图像。本文去噪模型由平稳小波变换、特征提取网络、小波系数预测网络和去噪图像重建四部分组成,含噪图像直接作为模型输入,在经过平稳小波变换后会产生不同的子带系数,每个子带系数都有一个与之对应的由卷积神经网络组成的小波系数预测网络,用来预测无噪图像在不同频段内的小波系数系数。实验结果表明,本文方法在低噪情况下,PSNR和SSIM基本上可以接近目前优秀的去噪算法;在高斯白噪声标准差高于25时,本文方法可以超越他们,并且可以很好地维持原始图像的结构,可以取得更好的视觉效果。更多还原

【Abstract】 With the popularity of various digital instruments and digital products,images and video have become the most commonly used information carriers in human life.The image contains a lot of information in our daily life and becomes the main way for people to get the original information from the outside world.While in the process of acquisition,transmission and preservation of image singnal it will inevitably brings in various noises which can degrading the image.This noise not only affects people’s visual experience,but also greatly affects subsequent image processing,such as medical imaging,satellite imaging,and so on.Image denoising is still very challenging in image processing.Many researchers are thinking about how to remove noise more thoroughly and preserve the original image information to the utmost extent?This paper mainly studies the methods of image denoising based on convolutional neural networks.The main contributions of the thesis includes:（1）A denoising model based on deep convolutional neural network is proposed.In order to extract the deeper features of the image and solve the problem of poor learning efficiency caused by the increase of network depth,a residual network is introduced.Most of the denoising methods are to train different models to deal with different intensities of noise,so you need to estimate the noise intensity first,and then use the corresponding model to denoise.We used a trained MLP^[1]model to test"Lena"with a standard deviationσof 10,25,50 and 70 Gaussian white noise the corresponding PSNR is 28.87 dB,31.28 dB,17.66 dB and 13.82dB respectively.From this result we can get that the model trained at a single noise intensity can not handle other intensity noise very well.A convolutional neural network model designed in this paper and it can handle different intensities of noise.（2）We have combined the denoising model with stationary wavelet transform in order to better preserve the original image information while denoising.Many current deep learning image denoising methods exploit the strong nonlinear fitting ability of convolutional neural networks to learn pixel-level mapping from noisy images to clean images,but sometimes they will produce smoother results when dealing with high-intensity noise,and will lose some texture details of the original image.In order to improve the performance under high-level noise,this paper proposes an image denoising method which combined with neural network and stationary wavelet transform.Instead of directly predicting the pixels of the clean image,we first predict the wavelet coefficients of the clean image and then reconstruct the denoised image.The denoising model used in this paper is composed of four parts:stationary wavelet transform,feature extraction network,wavelet coefficients prediction network and the reconstruction of denoised image.We directly take noisy image as input,after the stationary wavelet transform,different subband coefficients are generated through correspoding subnetwork.The coefficients prediction network which composed of convolutional neural networks is used to extract the features of each subbands and predict the subband coefficients of the clean images.The experimental results show that the PSNR and SSIM of our method can basically approach the current excellent denoising algorithm in the low-intensity noise.When the gaussian noise’s standard deviation is higher than 25,we can surpass those excellent algorithm and can achieve good visual effects.更多还原