Image Denoising And Super-Resolution Reconstruction Applications Based On Convolutional Neural Network

Enhancing image quality has become a hot issue in the field of image processing,such as removing noise from the image,enhancing or suppressing certain parts of the image,or changing the brightness of the image,etc.,thereby improving the quality of the image.With the wide application of deep learning methods in image processing,convolutional neural networks provide new ideas for image restoration and reconstruction.The combination of the popular convolutional neural network and traditional image processing application algorithms has brought more innovation and applied research directions.This work focuses on image denoising and super-resolution reconstruction tasks based on convolutional neural networks.The research includes:(1)A Symmetric model of image denoising based on residual unit and wavelet transform is proposed.The model is a symmetric structure model based on the residual learning algorithm and multi-level wavelet transform.It combines the powerful representation capabilities of convolutional neural networks and the advantages of wavelet decomposition algorithms,using pre-activation layer in the residual unit and optimizing the complex redundant structure of the deep networks to prevent the problem the gradient disappearance of the model and avoids the use of downsampling to cause the loss of image details.In addition,the method uses two training samples of different sizes for network training,which proves that the network can effectively improve the model performance by using larger training data samples.This method improves the denoising performance of the network model in a Gaussian noise environment to a certain extent,and obtains a higher evaluation value of image restoration quality and better visual effects.(2)A deep learning color image super-resolution reconstruction model fusing different color spaces is proposed.Images in different color spaces represents different types of calculation data from the perspective of image processing.This method introduces color space image data in different formats for model training,which can provide rich image feature information to the model.From the model structure,the network can be divided into RGB and XYZ branch networks.This method embeds a custom residual block structure and wavelet decomposition algorithm in the model.In the model training,the image features extracted from the XYZ branch are mapped to the RGB branch model for feature fusion,and finally deploys an inverse wavelet transform to reconstruct a high-resolution feature map.In the experiment,compared with other image super-resolution reconstruction methods,this method combines different color space information to obtain better image restoration effect in the super-resolution reconstruction task.In summary,this work mainly focuses on the research of natural image restoration tasks base on convolutional neural networks,wavelet transform decomposition and image color space characteristics.Experiments prove that the proposed method performs well in image denoising and super-resolution reconstruction.