Study On Image Denoising Algorithm And Noise Estimation

Digital images will inevitably generate noises during the period of acquisition,processing,compression,storage,transmission and reproduction,resulting in visual quality degradation of the images.In order to obtain high-quality images,various image denoising algorithms have been proposed,while they may be roughtly divided into two categories,namely model-based methods and learning-based methods.Traditional model-based image denoising algorithms have an abundant theoretical foundation,but often have complex optimization problems.On the other hand,with the booming of deep learning,the existing image denoising methods based on Convolutional Neural Networks(CNN)regard image denoising as a simple discriminative learning problem,bringing new ideas and advantages compared with the traditional image denoising methods.In this thesis,Additive Gaussian White Noise(AWGN)image is studied,the advantages of model-based and learning-based image denoising methods are used in image denoising algorithms and noise parameter estimation methods.The main research work of this thesis is as follows:(1)Research on wavelet packet noise parameter estimation methodIn order to effectively eliminate noise,it is necessary to accurately estimate the noise parameters at first.Traditional noise estimation algorithms always have the problem of overestimation or underestimation,which leads to inaccurate results of noise estimation.Inspired by the classic wavelet domain absolute median deviation method of estimating noise parameters,and the characteristic of wavelet packet analysis has more detailed multi-resolution than wavelet analysis,the absolute median deviation was introduced into wavelet packet transform for noise estimation,and the representative denoising algorithm in traditional image denoising methods---Block-matching and 3D Filtering(BM3D)denoising method was used to conduct experimental verification for noise estimation results.The experimental results show that the noise estimation method based on wavelet packet is more accurate than other comparison estimation methods.(2)Research on an improved deep convolutional neural network method for imagedenoisingThe loss function is a metric function used to evaluate the difference between the predicted value and the real value of network model output.In the regression tasks of deep convolutional neural networks,the mean square error is often used as the loss function of the network model,because it has the advantages of comparing differences from pixel to pixel and fast convergence,but it does not take human visual perception into consideration and does not conform to the visual information processing mechanism of human visual system(HVS).Therefore,on the basis of the classic deep convolutional neural network---Dn CNN network,the structural similarity index(SSIM)that conforms to human visual perception is used as a part of the loss function,and the mean square error together is used as the loss function of the whole convolutional neural network model in this thesis.Experimental results show that this method can not only better protect the detail of images,but also improve the quality of images.