In the past decade, supervised denoising models trained on large datasets have demonstrated impressive performance in image denoising due to their superior denoising effect. However, these models lack flexibility and exhibit varying degrees of degradation in denoising performance in practical applications, particularly when the noise distribution of the given noisy images does not match the training images. Our preliminary experiments suggest that even under ideal conditions, the denoised images obtained using these supervised denoising models (also known as preprocessed images) are already very similar to the ground truth images in terms of pixel intensities. Adding low-level noise to a preprocessed image can approximate the intensities of some pixels to their original values, but not all pixels. Based on this observation, we propose a novel two-stage approach to enhance the denoising effect of existing supervised denoisers using a low-noise injection strategy. In the first stage, we use a state-of-the-art supervised denoiser to denoise the given noisy image and obtain a preprocessed image. Then, we repeatedly inject different random low-level Gaussian noises to further improve certain pixels of the preprocessed image. The generated images are used as target images, and we obtain corresponding fine-tuned images within the framework of the unsupervised deep image prior (DIP) method by fully utilizing its flexibility. As a result, we obtain several denoised fine-tuned images that, respectively, approximate the ground truth image at specific pixels and complement each other. In the second stage, these fine-tuned images are fed to an unsupervised fusion network, which fully leverages the complementarity among the sample images to generate a fused image as the final denoised result. Experimental results demonstrate that the proposed method significantly improves the denoising effectiveness of synthetic noisy images, especially far surpassed the state-of-the-art methods in dealing with real noisy images.
