Convolutional Neural Network Based Image De-Noising

Abstract

De-noising of an image is a cardinal pre-processing step for image scrutiny and computer vision such as visual tracking, image registration, image classification, image segmentation and image restoration. The reasons of using Convolution Neural Network (CNN) based image de-noising are CNN with very broad design is successful in improving the versatility and efficiency to leverage image resources. Second, substantial progress has been made on regularizing and practicing strategies for CNN instruction, including the Rectifier Linear Unit (ReLU), batch normalization, and residual practicing. In CNN such approaches may be implemented to simplify the training cycle and boost the de-noising efficiency. Usually, all of those strategies suffer from two big pitfalls. First, in the testing level, such approaches usually have complicated question of optimization, which would make the procedure time consuming. Therefore, most approaches without losing numerical efficiency can hardly attain high results. Furthermore, the models are usually non-convex and require many parameters selected by hand, CNN centered simple discriminative learning approach is used to solve such issues, in which noise is isolated from a discrete picture by feed-forward convolutionary neural networks. Discriminative learning based image de-noising has been attracting considerable attention due to its fast extrapolation and good performance. Wavelet decomposition enhance the performance and minimize the computational complexity of this process. It provides us best compression ratio without degrading the quality of image. Convolutional neural network based image de-noising use batch normalization and residual learning which help to speed up the training process and accelerate the convergence of the network. Latent clean image extracted using residual learning from hidden layers. De-noising Convolution neural network (DnCNNs) model able to handle multiple Gaussian noise level as well as blind noise level. Several experiments are performed and compared with Avant-grade de-noising method to evaluate the de-noising performance and complexity of the network. The result shows that neural network based image de-noising is effective and efficient for practical image de-noising applications.