Densely Connected Dilated Residual Network for Image Denoising: DDR-Net

Acar V., Ekşioğlu E. M.

Neural Processing Letters, 2022 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Publication Date: 2022
  • Doi Number: 10.1007/s11063-022-11100-4
  • Journal Name: Neural Processing Letters
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, PASCAL, Applied Science & Technology Source, Compendex, Computer & Applied Sciences, Information Science and Technology Abstracts, INSPEC, zbMATH, DIALNET
  • Keywords: Image denoising, Deep learning, Dense connections, Dilated convolution
  • Istanbul Technical University Affiliated: Yes


Image denoising is an important ill-posed problem of image processing. The main goal in image denoising is to suppress noise while protecting textures of the image. A plethora of methods and different approaches have targeted image denoising problem. In this manuscript, we propose a novel method which brings dense and residual blocks together with dilated convolutions in image denoising applications. The proposed method introduces new strategies for proper combination of dense residual blocks and dilated convolution layers. The resulting approach is called as Densely connected Dilated Residual Network (DDR-Net). The proposed DDR-Net extracts multi-scale information by employing dilated convolutions. Use of dilated convolutions leads to improved receptive field performance while keeping the complexity of the network in check. Residual and dense connections on the other hand prevent loss of information along the DDR-Net pipeline. Simulations are performed in both color and grayscale image denoising settings. Quantitative and qualitative results indicate that the proposed network architecture leads to improved results when compared to multiple high performance convolutional denoising networks from the recent literature.