DRFENet: Toward a residual network of dilated convolution for image denoising

Abstract Deep learning technology dominates current research in image denoising. However, denoising performance is limited by target noise feature loss from information propagation in association with the depth of the network. This paper proposes a Dense Residual Feature Extraction Network (DRFENet) combined with Dense Enhancement Block (DEB), Residual Dilated Block (RDB), Feature Enhancement Block (FEB), and Simultaneous Iterative Reconstruction Block (SIRB). DEB enhances the extraction of noise features in the initial stage of the network. RDB increases the effect of the concatenated dilated convolution by using skip connections to broaden the receptive field of the model. FEB connects the incoming noisy images to enhance local feature information. SIRB uses the attention block to learn the noise distribution while using residual learning (RL) technology to reconstruct a denoised image. We examined the performance of DRFENet in gray image denoising on datasets BSD68 and SET12. Then, we evaluated its color image denoising performance on datasets McMaster, Kodak24, and CBSD68. The experimental results show that the denoising accuracy of DRFENet is better than the most existing image denoising methods. The code will be posted on Github