Research On Image Denoising Algorithm Based On Convolutional Neural Network

With the rapid development of camera and smartphone,digital image processing technology is prevalent.Digital images exist in every aspect of our lives,but they are influenced by factors such as the imaging system,the storage capacity and network bandwidth,and then digital images are often corrupted by noise.The restoration of noise image is a research hotspot in in both academia and industry.With the tremendous progress of convolutional neural network(CNN),recent years have witnessed a research upsurge of applying CNN to solve image denoising.The main research contents and contributions are summarized as follows:1.Since downsampling operation can increase the receptive field of the network effectively,multi-level network architecture uses the downsampling operation and upsampling operation,so based on novel multi-level convolutional neural network methods have achieved great success in image denoising task.Unfortunately,the increase in the level or depth of multi-level CNN does not result in a linear increase in denoising performance.To solve this problem,we propose a multi-scale&multi-level shuffle-CNN via multi-level attention(DnM3Net),which mainly adopts multi-path network parallel mechanism and plugs the multi-scale feature extraction,fine-to-coarse feature fusion strategy and multi-level attention module into the dual-Branch parallel sub-network architecture.DnM3Net utilizes dual-Branch parallel sub-network to further improve learning ability and enhance the ability of feature extraction.Dn3Net gets impressive performance because the better trade-off between denoising and detail preservation by use of coarse-to-fine feature fusion.The proposed novel network architecture is validated by applying on synthetic Gaussian noise gray and RGB images.Experimental results show that the DnM3Net effectively improves the quantitative metrics and visual quality compared to the state-of-the-art denoising methods.2.The existing denoising network adopts the traditional training mode of single supervision,and the network carries on the training with the complete network from beginning to end.We propose hierarchical training strategy,and the corresponding denoising network is called Hierarchical Trained Multi-level Wavelet-Resnet(HTMWResnet).HT-MWResnet plugs wavelet sub-bands of noised and clean image at different decomposition levels into different levels sub-network of network to better capture the dependence between wavelet sub-bands.The model is trained in a progressive growing fashion which trains the bottom level of the network at the beginning and higher levels are fused into the trained network gradually.Compared with the traditional training mode,the hierarchical training strategy makes deep network easier to train and achieves better denoising performance.Experiments on synthetic Gaussian noisy images demonstrate that our HT-MWResnet achieves better accuracy and visual improvements against other state-of-the-art methods.3.Existing CNN denoisers,which heavily rely on synthetic Gaussian noisy images as inputs and clean images as targets,usually come across the generalization problem when being applied in real noisy images with more sophisticated acquisition process.The main reason is that there is a big difference between real noise image and Gaussian noise image.Existing real noise image denoising methods use real noise image and clean image pairs as training data,but those methods ignore the use of noise estimation or use less efficient noise estimation methods.In order to solve the above problems,we introduce a efficient noise estimation method with standard deviation under the sub-block image,the corresponding network is called Cascade Blind Multi-level Wavelet-Resnet(CBMWResnet).CBMWResnet is composed of a cascade of noise estimation subnetwork and non-blinded denoising subnetwork.CBMWResnet achieves the state-of-the-art performance on the realistic Darmstadt Noise Dataset(DND)and Smartphone Image Denoising Dataset(SIDD).