Image Denoising Based On Deep Convolutional Neural Networks

Image denoising is one of the most classic problems in the field of computer vision.Removing the noise from the noisy image can greatly improve the quality of the image and bring better visual performance.And the processed image will also contain more features.Therefore,Image denoising has become a hot topic in recent years,and it plays a great role in many fields such as camera imaging,criminal investigation,medical image processing,video surveillance imaging,and satellite remote sensing image.With the development of the era of big data,image denoising based on deep learning has become one of the current mainstream methods due to its unique superiority.The main contributions of this thesis are as follows:1.A brief overview of the research background,significance and development status of image denoising technology.The current image denoising algorithms are classified and summarized,and several common methods in this field are analyzed.After that,the related theoretical knowledge of convolutional neural networks is elaborated,and the applications and development of deep learning in image denoising are introduced.2.Improving the network structure based on the original DnCNN algorithm.a multi-scale feature extraction residual network combined with perceptual loss is proposed for image denoising.The network first extracts the shallow features of different scales by combining the feature extraction layers of three different size convolution kernels,to increase the adaptability of the network to multi-scale features.Then use a series of residual units to learn the residual map to speed up the network convergence.Finally,in order to address the problem of over-blurring in reconstructed images,this thesis combines perceptual loss with traditional pixel-by-pixel loss to define the joint loss function,so that the network training can not only be compared at the pixel level,but also can be learned at a higher level of semantic features to get clear images.3.Introducing dilated convolution in the network,improving the denoising performance by expanding the denoising neural network perception field.Aiming at the characteristics of pixel-level prediction regression of image restoration,a hybrid convolutional layer with a dilated convolution kernel and a standard convolution kernel are proposed to expand the receptive field while avoiding the appearance of blind spots in the sensory domain.Simultaneously,the blind denoising model training is performed to realize a single network to process multiple levels of noise at the same time.4.At the end of the thesis,the image denoising technique is applied to satellite remote sensing image,and a high frequency layer network which is suitable for remote sensing image denoising is proposed.Firstly,the high-frequency layer decomposition of the initial noise image is performed,and the low-frequency background information that is not affected by the noise point is ignored,and only the high-frequency layer information is processed,so that more ground information can be protected from being destroyed.Then,through a symmetric codec convolutional neural network,a series of sampling operations are performed to realize end-to-end remote sensing image denoising in multiple scale spaces.