Research On Image Denoising Method Based On Convolutional Neural Network

As one of the important and effective carriers of information transmission,images have become an important source for people to obtain information.However,noise may occur during the process of image acquisition,processing,transmission,and reception.Noise images hinder people's recognition and acquisition of the required information in the images,and also affect the subsequent processing of the images.Therefore,the work of image denoising is extreme significant and it has become one of the considerable topics of people's research.In recent years,due to the flexible connection of network architecture and the strong learning ability of data model,deep learning has been increasingly applied in the field of image denoising,especially the application of deep convolutional neural network with great modeling ability in the field of image denoising.In this paper,the research on image denoising is based on the convolutional neural network technology in deep learning,and proposes a multi-scale parallel convolutional neural network image denoising method.The specific research content is as follows:The multi-scale parallel convolutional neural network designed in this paper is mainly composed of two modules,namely module one and module two.Module one make use of a multi-feature extraction method.First of all,three different sizes of convolution kernel filters are applied to learn the noise feature distribution information of the input noise image data.Since each convolution kernel is different in size,the receptive field in the image where it is located is also different.The extracted feature information of pixels is also different.By extracting pixel feature information of different sizes of the same pixel location,the network can learn more rich noise feature information at the input end;then,the extracted feature distribution information is separately trained in each size of the network channel;finally,the multi-feature serial splicing technology is applied to concatenate the information extracted from different network channels.Module two is a relatively shallow network structure.It mainly trains the multi-feature information extracted from module one;then the network output is obtained by subtracting the residual image information from the input noise image through the residual learning technology,and finally the denoised image is obtained.In this paper,batch normalization and residual learning techniques are applied in the network.Batch normalization is conducive to solve the problem of internal co variate migration in the training phase of the network.Residual learning can alleviate the gradient disappearance and explosion problems of deep networks during training,at the same time,it also reduces the loss of information in network training.By comparing the image denoising algorithm in this paper with other image denoising algorithms,experimental data show that:in terms of objective evaluation indicators such as peak signal to noise ratio(PSNR)and structural similarity(SSIM),the method in this paper can obtain better value;in terms of subjective observation,the method in this paper has better ability to protect the texture and details in the image than other methods,making the denoised image appear more clear and natural.Therefore,it can be shown that the method in this paper has a good capability of denoising performance.