Poisson2Poisson-Sparse: Unsupervised Poisson noise image denoising based on sparse modeling

In fields such as low-light photography, astronomical imaging, and low-dose computed tomography scanning, Poisson noise severely degrades image quality due to extremely low photon counts (averaging below one) and their Poisson-distributed statistical characteristics. A recent self-supervised Poisson denoising method uses only a single noisy image to improve image quality. However, it struggles under high Poisson noise due to its denoising model based on Gaussian distribution and suffers from long inference times. To address these issues, we propose an unsupervised Poisson denoising method based on sparse representation. Specifically, we first establish a more accurate sparse representation model based on Poisson distribution to enhance denoising performance. Given the difficulty of solving this model directly, we develop an iterative optimization algorithm using convolutional sparse coding and the alternating direction method of multipliers. Inspired by the unfolding technique, we further reduce computational cost by unfolding the iterative process into a finite-cycle learning network. To overcome the reliance on paired datasets and accelerate inference times, we employ a Poisson loss function, a Neighbor2Neighbor training strategy, and incorporate total variation loss, which together enable unsupervised learning. Experimental results demonstrate that our proposed method significantly outperforms existing unsupervised Poisson denoising methods and achieves high computational efficiency.