Enhancing deep image prior with roughly clean pairs and spatially random sampling

In recent years, significant advancements in image denoising have been achieved using large-scale datasets and strong supervision. However, a key challenge is obtaining well-aligned pairs of noisy and clean training images for specific scenarios. Existing unsupervised methods can perform denoising without ground-truth images but often rely on impractical conditions like paired noisy images, leading to suboptimal performance. This work aims to enhance the deep image prior (DIP) method’s effectiveness and efficiency by introducing a hybrid strategy that combines supervised and unsupervised approaches, termed EDIP (enhancing deep image prior). Our method incorporates roughly clean image pairs and employs two advanced supervised denoisers on noisy images to generate seed images. A spatially random sampler creates multiple subtly varied sampled images for stable training. Additionally, we use the average of the seed images as a secondary target alongside the noisy input. The network is then trained using the standard DIP unsupervised approach to produce the denoised image. Notably, both input and output consist of nearly clean images, which limits the search space and simplifies the mapping task for the network. Consequently, the proposed method yields high-quality denoised images while improving execution efficiency. Extensive experiments on public real-world datasets demonstrate that EDIP outperforms the original DIP method and state-of-the-art supervised denoisers in denoising effectiveness, without requiring labeling.