Image denoising using convolutional neural network

Abstract

Image noise degrades the performance of various imaging applications including medical imaging, astronomy imaging and microscopy. Thus, image denoising is extremely important, especially when the data requires further processing. Several discriminative learning models have been developed recently to produce high denoising performance. The proposed denoising convolutional neural network (DnCNN) incorporates residual learning method and batch normalization to improve denoising performance as well as increase the computational efficiency. DnCNN model manages to deal with additive white Gaussian blind denoising while existing discriminative models are usually designed for a specific noise level. In this paper, we attempt to fine-tune the network parameters to further optimize the performance. With deeper network and larger patch size, DnCNN is able to extract more context information to produce better denoising performance. Moreover, experiments on different types of image noise, namely Poisson, Salt-and-Pepper noise will be conducted to evaluate the extensibility of DnCNN model.