Enhancing Image Quality Through Photon Based Generative Denoising with Hyper Parameter Tuning

Poisson shot noise is a problem in low-light imaging because photon detection is random, and it frequently arises when fluorescence microscopy aims to minimise light damage while maintaining image quality. Although deep learning techniques such as Content-Aware Image Restoration (CARE) have demonstrated potential for noise reduction, they necessitate paired noisy-clean images, which are challenging to acquire. We suggest using the Generative Accumulation of Photons (GAP) framework to address this, a self-supervised method reinterpreting image formation as photon accumulation over time. GAP predicts the distribution of future photon arrivals, enabling effective denoising and generating diverse clean images from a single noisy input. Additionally, GAP serves as a fully generative model capable of creating images from scratch via iterative photon prediction. To support benchmarking, we introduce the shot noise-corrupted conv pc dataset, providing a valuable resource for advancing noise reduction in microscopy.