Research And Optimization Of Image Denoising Algorithm Based On Deep Neural Network

As a basic research in the field of computer vision and image processing,image denoising has extremely important significance in various practical applications such as security monitoring,medical diagnosis,and automatic driving.Traditional image denoising algorithms usually involve complex optimization problems and require manual setting of multiple parameters.Therefore,the denoising process is timeconsuming and has poor results.With the rapid development of artificial intelligence in recent years,image denoising algorithms based on deep neural networks have received widespread attention.This paper conducts detailed research and optimization of image denoising algorithms based on deep neural networks.The main tasks are as follows:(1)Aiming at the problems of insufficient utilization of image feature information and unsatisfactory denoising effect of existing algorithms,an image denoising algorithm based on multi-scale deep residual network is proposed.Among them,a multi-scale feature extraction module is proposed based on the Inception structure to extract rich feature information of different scales;a structure combining global jump connections and local jump connections is proposed to make full use of shallow image feature information to alleviate problems such as network degradation and gradient disappearance;based on attention The force mechanism introduces the channel attention module to increase the model's attention to important feature channels;the parameter adaptive activation function PRe LU is used to replace the Re LU function.The experimental results show that the image denoising algorithm based on the multiscale deep residual network has better image visual effects than other algorithms after denoising.The noise level ?=15,25,50,75 on the test set BSD68 is compared with Dn CNN The PSNR is improved by 0.05 d B,0.09 d B,0.15 d B,0.27 d B respectively,and the denoising performance is even better.(2)Aiming at the problems of large amount of parameters and high complexity of the existing algorithm models,an image denoising algorithm based on the hole convolution deep residual network is proposed.Among them,the multi-scale feature extraction module is optimized by the introduction of the hole convolution;the residual module of the hole convolution is proposed to alleviate the grid effect caused by the stacking of the hole convolution.The experimental results show that the image denoising algorithm based on the hole convolution deep residual network is equivalent to Dn CNN in subjective and objective evaluation,but the parameter amount is only 3/5 of Dn CNN,and the running time on CPU and GPU Are lower than other algorithms.