We propose a novel approach for unpaired learning of deep image denoising that does not require paired clean-noisy training data or knowledge of the noise model. Our method uses a Dilated Blind-Spot Network (D-BSN) trained in a self-supervised manner followed by knowledge distillation to produce a practical denoiser. The D-BSN learns to predict each pixel from its surrounding context without seeing the pixel itself, and the distilled student network produces clean outputs directly. This achieves competitive results with supervised methods on real-world noisy images.

Image denoising has been dominated by supervised learning methods that require paired clean-noisy image datasets. In practice, obtaining perfectly aligned clean-noisy pairs is extremely difficult or impossible for real-world noise. Existing unsupervised methods either assume a known noise model (e.g., Gaussian) or use noise-to-noise training that still requires multiple noisy observations of the same scene. We propose a fully unpaired approach that works with only a collection of noisy images, without any clean references or noise model assumptions.

The key theoretical foundation is that if noise is pixel-wise independent, a network that predicts a pixel without accessing it will converge to the conditional expectation of the clean signal. This is the same principle behind Noise2Self and blind-spot networks. However, existing blind-spot architectures produce suboptimal results due to the receptive field constraint. Our two-stage approach elegantly resolves this limitation.

Our approach consists of two stages. In Stage 1, we train a Dilated Blind-Spot Network (D-BSN) that predicts each pixel from its surrounding context using dilated convolutions that skip the center pixel. Under the assumption that noise is pixel-wise independent, the network can learn to estimate the clean signal without ever seeing clean images. The D-BSN output is a denoised estimate with some residual artifacts due to the blind-spot constraint.

In Stage 2, we use the trained D-BSN as a teacher to generate pseudo-clean targets from noisy images. A standard U-Net student without the blind-spot constraint is then trained to map noisy inputs to these pseudo-clean targets. This distillation step removes the artifacts caused by the blind-spot constraint and produces a practical denoiser that can process pixels using their full local context.

On the DND benchmark for real-world denoising, our method achieves 39.28 dB PSNR, compared to 39.75 dB for supervised CBDNet and 37.50 dB for BM3D. On SIDD benchmark, we achieve 38.92 dB. Remarkably, our unpaired method comes within 0.5 dB of fully supervised approaches while requiring no paired training data. The distilled student outperforms the D-BSN teacher by 0.8 dB, confirming the value of the distillation stage.

We also compare different blind-spot architectures. Our D-BSN with dilated convolutions outperforms the standard blind-spot network by +0.6 dB, as dilated convolutions provide a larger effective receptive field while maintaining the blind-spot property. We further show that the method works across different noise types including Gaussian, Poisson, and real sensor noise.

We have presented a fully unpaired approach to deep image denoising that combines self-supervised blind-spot learning with knowledge distillation. Our method achieves near-supervised performance without any paired data or noise model assumptions. This work demonstrates that self-supervised learning can largely close the gap with supervised methods in low-level vision tasks.