Abstract:Most existing deep learning-based image denoisers heavily rely on supervised learning , where a denoising network is trained with massive noisy–clean training pairs. However, the lack of training pairs severely limits the applicability of this kind of denoisers in practice. Recently, an unsupervised denoiser, the so-called deep image prior (DIP), which can accomplish the task of noise removal only with the given noisy image itself, provides a great flexibility for those scenarios where acquiring a training image set is difficult, and has attracted extensive attention due to its practical significance and broad applications. However, the absence of external information dramatically reduces the upper bound of the denoising performance of the DIP method. In this paper, we propose a simple yet effective multi-channel approximation mechanism to boost the denoising performance of the DIP, where several representative denoisers are exploited to provide additional priors to ensure that the DIP model converges to a more reasonable position in the solution space during the iterative optimization process. Specifically, for a given noisy image, we first use two representative denoisers belonging to the internal and external prior-based denoising methods to denoise it, and the resulting denoised images are called initial denoised images that will be used as target images to constrain the search space. Then, the last layer of the original DIP model is expanded from single channel to multi-channel correspondingly, which allows the DIP model to produce multiple denoised images, called intermediate denoised images, in a single run. Thereby, the total execution time is considerably decreased. Meanwhile, we adopt a novel sharpness-based pseudo-reference image quality metric to automatically stop the iterative optimization process and ensure the image quality of the intermediate denoised images. Finally, the intermediate denoised images with good complementarity are fused to yield the final denoised image with structural patch decomposition-based fusion method. The extensive experiments on synthetic and real-world images show that the proposed method has superior performances over existing non-learning-based and learning-based methods.

Online Deep Equilibrium Learning for Regularization by Denoising

Deep Equilibrium Learning of Explicit Regularizers for Imaging Inverse Problems

Regularization by Denoising via Fixed-Point Projection (RED-PRO)

Monotonically Convergent Regularization by Denoising

RARE: Image Reconstruction Using Deep Priors Learned Without Groundtruth

Online Regularization by Denoising with Applications to Phase Retrieval

The Little Engine that Could: Regularization by Denoising (RED)

A Deep Optimization Approach for Image Deconvolution.

Averaged Deep Denoisers for Image Regularization

Async-RED: A Provably Convergent Asynchronous Block Parallel Stochastic Method using Deep Denoising Priors

Deep Equilibrium Object Detection

Simultaneous Nonlocal Low-Rank and Deep Priors for Poisson Denoising

A Triple Deep Image Prior Model for Image Denoising Based on Mixed Priors and Noise Learning

Multi-channel deep image prior for image denoising

DLRP: Learning Deep Low-Rank Prior for Remotely Sensed Image Denoising

RED-PSM: Regularization by Denoising of Factorized Low Rank Models for Dynamic Imaging

An Interpretation of Regularization by Denoising and its Application with the Back-Projected Fidelity Term

Learning Deep CNN Denoiser Prior for Image Restoration

Dynamic PET Image Denoising Using Deep Image Prior Combined with Regularization by Denoising