A prior-guided deep network for real image denoising and its applications

Abstract

Recently, the popularity of deep learning in computer vision has brought about a rapid development of deep convolutional neural network (CNN)-based image denoising algorithms. Most of the existing methods have shown their superiority in removing additive white Gaussian noise (AWGN) with a specific noise level, but have limited performance in handling much more complicated real noise. To address this problem, we develop a novel prior-guided dynamic tunable network (PDTNet) for real image denoising. Firstly, we break up the image denoising optimization problem into noise estimation and image reconstruction sub-problems, and employ the inference process to guide the architecture design of PDTNet. Then, we exploit an internal and external dual modulation (IEDM) scheme to achieve real image denoising. Specifically, a designed global spatial and channel attention (GSCA) is embedded in the external estimator and internal stacked dynamic residual blocks (DRBs) to extract global features from the noise prior and iterative image features, respectively. Next, a dynamic weight generator block (DWGB) is leveraged to adaptively combine external and internal features at each DRB. In addition, we also analyze a realistic real noise model from the physical perspective to generate synthetic noisy images for model training. Experimental results show the superiority of PDTNet over state-of-the-arts both quantitatively and visually and its applications on different networks and tasks.