Image Denoising Algorithm Based On Enhanced Depth Residual Network

Image denoising is an essential issue in computer vision field.Its main objective is to avoid the effect of noise.Traditional denoising methods are generally based on non-convex models,which means that users have to manually set adjustment parameters.Also,during system testing,the complex optimization strategy is adopted to improve the test results,which brings some problems,such as large computational complexity,poor visual effect of the generated noise reduction image and unsatisfactory details recovery.In order to overcome the shortcomings of traditional denoising methods,a high-performance single color image denoising algorithm is proposed and studied.The main results of this thesis can be given as follows.In this thesis,the image super-resolution frontier technology is applied to image denoising research.To improve the denoising performance and image detail texture restoration ability of traditional image denoising algorithm,an enhanced depth residual network combined with the generated image of the anti-network is proposed and a perceptual loss function for image denoising is presented.The developed denoising optimization scheme is divided into two phases.The first phase is the enhanced depth residual network denoising algorithm.The enhanced depth residual network used in image super-resolution is improved and used to solve the image denoising problem.The proposed algorithm is characterized by removing the batch normalization layer in the traditional residual block,which can effectively reduce the calculation amount and improve the efficiency.The jump connection layer is used to directly transmit the input information to the output,so that the main network can learn the residual between the input and output information.When in deep network,the network model training difficulty can be reduced,and the neural network can fully learn the nonlinear mapping between the noisy image and the original image such that the performance of the denoising algorithm can be improved.The second stage is the optimization of the image denoising algorithm.For the high-frequency detail loss of the image in the first stage is proposed,a denoising image detail recovery scheme based on the generated anti-network is proposed.The first stage denoising network is used as a generator,and the weighted sum of the loss-tolerant and content loss functions is used to construct a perceptual loss function.Through the confrontation training between the generator and the discriminator,the denoising network parameters are iteratively updated,thus the optimized denoising algorithm model is obtained.The image denoising scheme designed in this thesis can successfully restore the noisy single color image to a clear and high quality noise reduction image.The testing results in the open source data set demonstrate effectiveness of the enhanced depth residual proposed in this thesis.By comparison with some existed denoising algorithms,it is observed that the peak signal-to-noise ratio of the developed algorithm is 4.7% higher than that of other algorithms,and the structural similarity is increased by 4.3%.The denoising image is optimized by using the generated confrontation network,and the comparison experiments are carried out by various loss functions.It is verified that the optimized algorithm has better recovery ability for the denoised image detail texture.