Texture compensation with multi-scale dilated residual blocks for image denoising

Deep convolutional neural networks have achieved great success for image denoising recently. However, increasing the depth of the neural network cannot significantly boost the performance of the algorithms for image denoising. It is still a challenging research topic to recover structural information and fine details of the images. In this paper, we put forward a convolutional neural network for single image denoising, mainly consisting of a noise mapping block (NMB), a texture compensation block (TCB), and a composition block (CB). The NMB borrows a series of standard residual blocks to learn the noise mapping. Specifically, we employ the TCB to enhance the details via multi-scale dilated residual blocks (MDRBs) that hold the characteristics of fusing multi-scale contexture information with dilated convolution. Finally, the CB does an element-wise addition to composite the output. Besides, we have conducted extensive experiments on gray as well as color image datasets. Both quantitative and qualitative evaluations demonstrate the superior performance of our approach in comparison with the state-of-the-art methods.

Automated summary

In this paper, a convolutional neural network for image denoising is proposed, which includes noise mapping block, texture compensation block, and composition block. The model achieves superior performance by learning noise mapping, enhancing details, and synthesizing outputs.