Data-driven point spread function parameterization with generative neural networks. Application to high-contrast VLT/SPHERE observations

Abstract

The impulse response, or point spread function (PSF), of an optical system plays a central role in many image analysis problems, including deconvolution and detection. Obtaining accurate and noise-free PSF estimates remains challenging, especially for ground-based observations, where the PSF depends on both instrumental effects and time varying atmospheric turbulence. Here, we propose a data-driven low dimensional parameterization of telescope PSFs based on a generative neural network trained on experimental observations. The generator was trained on a dataset of measured PSFs from a given instrument, enabling it to model the diversity of the instrument’s PSF morphologies. In this way, the input space of the generator can be used as a constrained and low-dimensional parameter space for PSF estimation and reconstruction tasks. To enable the training procedure on experimental data, we introduced a dedicated framework combining flux normalization, signal equalization, and noise simulation during training. This made it possible for the model to learn to generate near-noise-free and flux-independent PSFs from noisy measurements. As a proof of concept, we trained the generator on PSFs acquired with the VLT/SPHERE instrument, which is dedicated to the study of the circumstellar environment of nearby stars. The proposed model successfully reproduces the diversity of observed PSFs through a few input ( latent) parameters and generates realistic synthetic PSFs free of noise artifacts. We estimate that approximately 500 to 1000 of the measured PSFs are sufficient to capture the main variability of the instrument response. Finally, we illustrate PSF denoising with the proposed approach in the context of object reconstruction. Replacing the measured PSF with its denoised counterpart improves the reconstruction of circumstellar environments and enables the recovery of finer disk structures from VLT/SPHERE observations.