Learning Deep Image Priors for Blind Image Denoising

Image denoising is the process of removing noise from noisy images, which is an image domain transferring task, i.e., from a single or several noise level domains to a photo-realistic domain. In this paper, we propose an effective image denoising method by learning two image priors from the perspective of domain alignment. We tackle the domain alignment on two levels. 1) the feature-level prior is to learn domain-invariant features for corrupted images with different level noise; 2) the pixel-level prior is used to push the denoised images to the natural image manifold. The two image priors are based on H-divergence theory and implemented by learning classifiers in adversarial training manners. We evaluate our approach on multiple datasets. The results demonstrate the effectiveness of our approach for robust image denoising on both synthetic and real-world noisy images. Furthermore, we show that the feature-level prior is capable of alleviating the discrepancy between different level noise. It can be used to improve the blind denoising performance in terms of distortion measures (PSNR and SSIM), while pixel-level prior can effectively improve the perceptual quality to ensure the realistic outputs, which is further validated by subjective evaluation.

  1. State of the art model-based methods such as BM3D usually involve a complex and time-consuming optimization process in the testing stage. Second, the image priors employed in most of these approaches are hand-crafted, such as nonlocal self-similarity and gradients, which are mainly based on the internal information of the input image without any external information. On the other hand, existing discriminative learning methods are usually designed to a specific noise level with limited flexibility. They still cannot generalize well to real-world noisy images. Therefore, it is of great interest to develop general image priors which can help handle image denoising with a wide range of noise levels and generalize well for real-world noisy images.


  1. we propose a new image denoising model, referred to Deep Image Prior Network (DIPNet), based on data-driven image priors. In particular, we consider image denoising as a domain transferring problem, e., from noise domain to photo-realistic domain. Inspired by this, we propose two image priors: 1) the feature-level prior which is designed to help decrease domain discrepancy between corrupted images with different noise levels for robust image denoising (Figure 1); 2) the pixel-level prior which is used to push the denoised image to photo-realistic domain for perceptual improvement (Figure 2). In particular, we model both priors as discriminator networks, which are trained by an adversarial training strategy to minimize the H-divergence between different image domains.

                                                                                                               Figure 1 feature-level prior

                                                                                                                Figure 2 Image-level prior


  1. We conduct experiments on synthetic images for additive white Gaussian noise removal with either known or unknown noise levels as well as real–world noisy image denoising. 

3.1AWGN noise Removal
We compare our method with CBM3D and FFDNet. Table 1 reports denoising results on different datasets. We can see that our method can achieve state of the art results and outperform other methods. Moreover, the improvement generalizes well across different datasets as well as different noise levels. In Figure 3, we compare the visual results of different methods for an image in CBSD68 corrupted with noise level σ = 50.
world noisy image denoising.

                                                                                 Table 1. Non-blind denoising results of different methods

                                                      Figure 3 Denoising comparisons of an image from CBSD68 dataset with noise level σ = 50

3.2 Noise Level Sensitivity

We compare the denoising performances of several different non-blind and blind DIPNet models with different input noise levels. Figure 4 shows the noise level sensitivity curves of different DIPNet models. Specifically, we consider the 5 non-blind DIPNets trained with known noise levels, e.g., “DIPNet-15” represents DIP- Net trained with the fixed noise level σ = 15. We also com- pare the results of DIPNet-BF and DIPNet-BP. It is clear that the non-blind DIPNet-S models with specific noise levels are more sensitive to input noise, especially higher levels. DIPNet-BF demonstrates much stable performance in terms of a wide range of noise levels. DIPNet-BP is more sensitive to higher level noise.


                                                       Figure 4 Noise level sensitivity curves of DIPNet models trained with different noise levels.

3.3 Real Noise Removal

Furthermore, we evaluate our blind models on two real noisy image datasets with different methods. Figure 5 shows denoising results on a real noisy image with different method.  We can observe that using feature-level prior can effectively improve PSNR but still produce over-smooth textures. The pixel-level prior can help produce sharper appearance but with lower PSNR and SSIM. This is due to the appearance of high-frequency artifacts.


  1. Reference

Xianxu Hou, Hongming Luo, Bozhi Liu, Jingxin Liu, Bolei Xu, Ke Sun, Yuanhao Gong, Bozhi Liu,  Guoping Qiu. “Learning Deep Image Priors for Image Denoising”. CVPRW 2019.