A multiscale dilated residual network for image denoising

• Published in: Multimedia tools and applications Vol. 79; no. 45-46; pp. 34443 - 34458
• Main Authors: Li, Dongjie, Chen, Huaian, Jin, Guoqiang, Jin, Yi, Zhu, Changan, Chen, Enhong
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.12.2020; Springer Nature B.V
• Subjects: Competition; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Experiments; Image enhancement; Image quality; Learning; Methods; Multimedia; Multimedia Information Systems; Neural networks; Noise; Noise reduction; Parameters; Special Purpose and Application-Based Systems; Deep learning; Dilated convolution; Residual learning; Image denoising
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-020-09113-z

In this paper, a more effective Gaussian denoiser is designed to enhance the resulting image quality. We propose a novel image denoising method using a multiscale dilated residual network, named MDRNet. The proposed method is based on two main strategies. First, we adopt dilated convolutions in our network to enlarge the receptive field while requiring fewer parameters. The hybrid dilation rate pattern (HDP) is implemented such that each pixel in the pattern contributes similarly to the receptive field, allowing our network to learn the image details equally. Second, we employ a contextualized structure to take advantage of the low-level features which are mainly concentrated in the first two layers. Our method achieves competitive denoising performance and requires fewer parameters compared to existing denoising methods that using convolutional network. Through comprehensive experiments, we show that the denoising performance of our method is competitive with the state-of-the-art methods in terms of both quantitative and qualitative evaluation.