A lightweight and effective deep learning model for Gaussian noise removal

Presently, Gaussian noise removal is a very attractive research direction, because Gaussian noise can effectively simulate the real world noise when the real noise sources are complex and diverse. Existing traditional denoising methods have high complexity and low efficiency. Means of image Gaussian noise denoising will become more efficient with the development of deep learning. In this paper a novel deep learning model for Gaussian noise removal is proposed. The proposed method combines dilated convolution with skip connection of residual learning, which is trained by our proposed mixed loss function during back propagation. Not only compared with existing traditional methods, but also compared with representative deep learning methods, our proposed model has a better performance. In addition, our proposed model has less layers and training parameters than representative deep learning methods. Through several comparative experiments on test data sets, we can draw a conclusion that our proposed deep learning model can remove Gaussian noise more effectively than some state-of-art methods with quantitative and visual qualitative analyses.