Research On Gaussian And Embedded Image Noise Removal Method Based On Deep Learning

Digital images are an important carrier of human perception and dissemination of information in the era of big data.However,during the process of image collection,transmission and storage,images are affected by transmission channel equipment or are artificially destroyed,resulting in noise interference and loss of effective image information,and affect subsequent image analysis and processing,such as image segmentation,target recognition,edge extraction,etc.On the other hand,if both the carrier and the hidden information are images during information hiding,when the carrier image is destroyed by noise interference,the embedded image will also be destroyed,the noise of the embedded image is also a special random noise,and its noise intensity is much greater than that of the carrier image.Image denoising is one of the most basic problems in the field of image processing,and removing random noise has always been a hot and difficult research point,so it is of great significance to study image denoising.For random noise,this thesis uses deep learning knowledge to design two denoising networks to restore images damaged by noise interference.The specific work is as follows:(1)Aiming at Gaussian random noise,a dual-path multi-feature fusion Gaussian denoising convolutional neural network is designed.The network adopts a dual-path mode to capture richer feature maps through dual-path.In order to increase the difference of the learning results of the two paths,dilated convolution is introduced into one of the paths.The feature fusion module is used to effectively fuse the feature maps obtained from the middle and end training of the network dual path,so that the low-level and high-level feature maps can play a role in denoising.The experimental results of the dual-path multi-feature fusion Gaussian denoising network on multiple grayscale and color public datasets show that the network can effectively remove Gaussian noise and retain most of the image details and texture features.(2)For the embedded image random noise,a UNet-based sliding window self-attention mechanism embedded image noise removal network is designed.The image is scaled through the UNet network structure,and the image is denoised globally by convolution.A sliding window self-attention mechanism is introduced,and its multi-window perception mechanism is used to locally denoise the image.By adding segmentation and double special sampling operations to the network,high-definition and high-resolution images can be obtained while solving the image checkerboard effect problem.Aiming at the problems of long iteration time and data redundancy,skip connections are replaced by residual connections.The denoising experiment results of the network on the embedded image noise dataset show that the network can remove the embedded image noise well,and the image texture details can be restored to a large extent,and a good denoising effect is achieved.In conclusion,this thesis analyzes the distribution characteristics of two types of noise and applies deep learning knowledge to design two corresponding image denoising neural networks,which can better restore images damaged by noise.