Fast and Dense Denoising Convolutional Neural Network

Abstract

Deep neural networks show us their superior image denoising capability due to the powerful fitting ability. However, they suffer from the following drawbacks: (i) too deep neural networks often imply a very large number of parameters and considerable computational overhead; (ii) neural networks that are too deep are difficult to converge by training and may lead to degradation. In this study, we propose a novel denoising network called the fast and dense denoising convolutional neural network(FDDCNN). In particular, the depthwise separable convolutions in the fast module and the homogeneous cascade structure in the dense module can efficiently solve the above problem. Extensive experiments with publicly available datasets have shown that this model can have the same excellent denoising power as existing methods with fewer parameters and less computational overhead.