Channel and Space Attention Neural Network for Image Denoising

Yi Wang,Xiao Song,Chen Kai

DOI: https://doi.org/10.1109/LSP.2021.3057544

2021-01-01

IEEE Signal Processing Letters

Abstract:Recently, convolutional neural networks (CNN) have been widely used in image denoising. But with most CNN denoising methods, all the channels are treated equally and the relationship between spatial locations are neglected. In the letter, we propose a novel channel and space attention neural network (CSANN) for image denoising. In CSANN, we concatenate the noise level with the average and maximum values of each channel as the input and propose a convolutional network to learn the relationship between channels. Meanwhile, we combine the noise level map with the average and maximum values of each spatial locations as the input and use a convolutional network to learn the relationship between spatial locations. Moreover, we combine them as an attention network and introduce it into the main CNN and symmetric skip connections, which makes channels related to attention network play different roles in the subsequent convolution and offsets the performance degradation caused by using a single convolution kernel in spatial locations. In addition, the use of symmetric skip connections and resnet blocks avoid the vanishing gradient problem and the loss of shallow features. Experimental results show that, compared with some state-of-the-art denoising algorithms, the experimental results of CSANN have better visual effects and higher peak signal-to-noise ratio (PSNR) values.