Research On Convolutional Neural Network For Image Denoising

With the popularization of digital cameras and smart phones,digital images have appeared in all aspects of people’s lives.Due to the influence of the environment and equipment,digital images are easily disturbed by noise during the process of capture,storage and transmission,which seriously affects the visual quality of the image.Image denoising,as a research hotspot in low-level visual tasks,can not only enhance the visual quality of images,but also improve the accuracy of high-level visual tasks.Nowadays,image denoising methods are dominated by discriminative learning methods based on convolutional neural networks.Since model-based denoising methods require numerous hyperparameters and complex optimization algorithms,the denoising performance is uncertain and the denoising speed is slow.Compared with model-based denoising methods,discriminative learning methods not only do not have a large number of hyperparameters,but also perform well in denoising performance and testing speed.In recent years,as the performance of the denoising network based on discriminative learning gradually improves,the complexity of these networks has also increased,resulting in a dramatic increase in the number of network parameters and a decrease in test speed.For this reason,this paper designs fast and effective denoising networks from the perspective of balancing the performance and complexity of the denoising network.The main work and innovations of this paper are as follows:(1)Aiming at the inflexibility of learning a single model to handle images with specific noise level and the inability to deal with images with spatial variation of noise level,a fast and effective U-shaped convolutional neural network(FEUNet)is proposed for image denoising.The network takes the noisy image and the noise level map as its input.Since the noise level map is changeable or non-uniform,the network can deal with a wide range of noise levels,including spatially varying and unvarying noise.In order to trade-off the denoising performance and speed,the proposed FEUNet works on the downsampled sub-images to improve the denoising speed.However,the downsampling operations will lead to the loss of image information,and skip connections are used to features superposing to address this problem.Compared with the existing image denoising methods,the proposed network has achieved an excellent balance in efficient and effectiveness,making it still have practical application value.(2)To solve the drawbacks of network accuracy and denoising performance degradation and large number of parameters caused by the deep denoising network,a dual denoising convolutional neural network is proposed for image denoising.The network contains two different 12-layer sub-networks,which can provide complementary feature information to improve the denoising performance of the network.Further,noisy images and noise level maps are used as the input of the network to enhance its flexibility.In addition,downsampling and dilated convolution are introduced to increase the receptive field to extract more context information and promote the denoising performance of the network;skip connections are utilized to obtain more image details and avoid gradient vanishing or exploding during backpropagation.Compared with other denoising methods,the proposed network performs well overall in terms of running time,number of parameters,FLOPs and effectiveness,making it provides an additional option on the image denoising task.(3)To address the weaknesses for the complex,deep,and large number of parameters of the network currently used for real image denoising,a dual convolutional blind denoising network with attention is proposed.In the network,the noise estimation sub-network is used for noise estimation of the image,and the estimated value is utilized as the input of the attention module together with the noisy image.The attention module is applied to suppress useless feature information and let the features with more useful information pass,thereby improving the feature representation ability of the denoising network.In addition,the dual convolutional denoising network uses downsampling and dilated convolution to increase the receptive field so that it can extract more contextual information and enhance the denoising performance.Through the analysis of the experimental results,it is found that compared with the more complex network,the denoising network still shows good denoising performance.