Résumé: In this work, by analyzing the modeling capability of the unsupervised deep image prior (DIP) network and its uncertainty in recovering the lost details, we aim to significantly boost its denoising effect by jointly exploiting generative and fusion strategies, resulting into a highly effective unsupervised three-stage recovery process. More specifically, for a given noisy image, we first apply two representative image denoisers that, respectively, belong to the internal and external prior-based denoising methods to produce corresponding two initial denoised images. Based on the two initial denoised images, we can randomly generate enough target images with a novel spatially random mixer. Then, we follow the standard DIP denoising routine but with different random inputs and target images to generate multiple complementary samples at separate runs. For more randomness and stability, some of generated samples are dropped out. Finally, the remaining samples are fused with weight maps generated by an unsupervised generative network in a pixel-wise manner, obtaining a final denoised image whose image quality is significantly improved. Extensive experiments demonstrate that, with our boosting strategy, the proposed method remarkably outperforms the original DIP and previous leading unsupervised networks with comparable peak signal-to-noise ratio and structural similarity, and is competitive with state-of-the-art supervised ones, on synthetic and real-world noisy image denoising.

Langue: Anglais