Robust deep ensemble method for real-world image denoising

Recently, deep learning-based image denoising methods have achieved promising performance. However, when the distribution of real-world noisy images is unknown, performance still limited due to domain gap between training set and testing set. Nonetheless, unknown noise usually can be modeled as proper combination existing distributions. In this paper, we propose a simple yet effective Bayesian ensemble (BDE) method for denoising, where several representative denoisers pre-trained with various data settings fused improve robustness. The foundation BDE that noises are highly signal-dependent, heterogeneous in separately handled by different denoisers. particular, take well-trained CBDNet, NBNet, HINet, Uformer GMSNet into denoiser pool, U-Net adopted predict pixel-wise weighting maps fuse these Instead solely learning maps, strategy introduced uncertainty well map, which prediction variance improving robustness on images. Extensive experiments shown better removed fusing instead big expensive cost. Furthermore, our extended other restoration tasks, i.e. deblurring, deraining single super-resolution, also achieves significance gain benchmark datasets.