Application Research Of Convolutional Neural Network In Image Denoising

About 75% of the external information obtained by humans comes from images.As an important information carrier,image is affected by various factors when it is acquired,which leads to noise.The purpose of image denoising is to remove noise from the image while preserving the original information in the image as much as possible.Traditional image denoising methods mostly involve manual selection of parameters,which leads to the problem that the global features of images can not be obtained.At the beginning of the 21 st century,deep learning has once again set off a research boom,which has been widely used in the fields of image processing and natural language processing.Convolutional neural network(CNN)is more suitable for image denoising as a deep learning method.The convolutional layer and the active layer in the CNN architecture can establish a nonlinear mapping relationship between input and output,and effectively learn the feature information in the noisy image.In this paper,multi-scale parallel convolutional neural network(MPCNN)and multi-scale parallel feature extraction convolutional neural network(MPFE-CNN)are designed for image denoising.The main work is as follows:MPCNN is mainly composed of the multi-scale feature extraction layer and the convolutional neural network parallel structure.The multi-scale feature extraction layer uses four convolution kernels of different sizes to extract different features from the original image,and fuses the different features into the parallel structure.The parallel structure consists of two forward transfer networks,deep channel and shallow channel.The deep channel is mainly responsible for extracting advanced features,while the shallow channel is responsible for extracting low-level features.Experiments show that under the test set BSD68,when the noise levels are σ=15,25,and50,the PSNR of MPCNN is increased by 0.05 dB,0.07 dB,and 0.10 dB,respectively,compared to DnCNN-S.MPFE-CNN is an improved structure for MPCNN.Five multi-scale parallel feature extraction modules are used in MPFE-CNN.Each moduleconstructs a parallel network using two different convolution kernels of 3×3and 5×5.Dense connections are added to the whole network to continuously exploit low-level features and accelerate network convergence.Experiments show that under the test set BSD68,when the noise levels are σ=15,25 and50,compared with MPCNN and DnCNN-S,the PSNR of MPFE-CNN is increased by 0.03 dB,0.03 dB,0.05 dB and 0.08 dB,0.10 dB,0.15 dB,respectively.Figure [42] table [8] reference [90].