Orateur
Description
Noise bias is a persistent challenge in shear measurements, particularly for galaxies with low signal-to-noise ratios (SNR), which are critical for Euclid’s cosmic shear analysis. To address this, we present a novel, unsupervised, deep learning denoising technique (ElFNet) to improve galaxy shape measurements. Using a convolutional neural network (CNN), we predict individual galaxy shears directly from simulated postage-stamp galaxy images. Our method combines adaptive moments with best-fit Gaussian PSF corrections and incorporates Metacalibration to calibrate the shear. Compared to GalSim, our approach demonstrates significant improvements in reducing noise bias.