A multi-scale adaptive feature enhancement network for image denoising

Recently, deep learning has been widely used in image denoising. However, most of the existing deep learning-based methods are not adequate in blind denoising for additive white Gaussian noise (AWGN) images and real-world noisy images, which are still noisy or blurred. The difficulty is how to handle different noise levels and different types of noise with only one pre-trained model. In this paper, we propose a multi-scale adaptive feature enhancement network (MFENet) to improve the performance on blind image denoising. The MFENet is based on residual learning and batch normalization to speed up the network convergence. In the MFENet, dilated convolution and deformable convolution can expand the receptive field to obtain rich information from different scales. The deformable convolution is also able to adjust the sampling position to fit different shapes of objects. Spatial attention is used to enhance important features in the large amount of information. The experimental results show that the proposed method for blind denoising outperforms the state-of-the-art methods on both synthetic and real-world noisy images.