Uso de autoencoders para análise de imagens no experimento CYGNO

Abstract

The CYGNO experiment focuses on the direct detection of dark matter using a Time Projection Chamber (TPC) with an optical readout based on high-resolution sCMOS sensors. While this technology offers excellent directional sensitivity, it generates a massive data stream in which the vast majority of images consist strictly of electronic and optical noise—referred to as the "pedestal." To avoid the impractical storage of irrelevant data, it is essential to develop fast, online trigger algorithms capable of filtering out background noise and isolating Regions of Interest (ROIs) in real time. This work proposes and validates a computational filter based on self-supervised machine learning, utilizing a Denoising Autoencoder (DAE) architecture trained entirely from scratch. The model learns the stochastic characteristics of the noise by being fed images corrupted with synthetic anomalies; it is then tasked with reconstructing only the clean pedestal, guided by a loss function that weights both the Mean Squared Error (MSE) and the Structural Similarity Index (SSIM). During the inference phase, a subtraction operation reveals the actual physical signatures, which are highlighted in the resulting residual map. Tests conducted using Monte Carlo simulations (GEANT4) demonstrated that the network preserves the morphological integrity of particle tracks and ensures a signal-to-noise ratio gain greater than 1, even when operating at the detector's sensitivity limit in the sub-keV regime (0.25 keV). Furthermore, the spatial suppression step typically retained less than 10% of the sensor array area in frames containing valid events, while reducing the storage of false positives to zero in samples consisting purely of noise. The results demonstrate the method's high discriminative and compression capacity, with future work to include a systematic analysis of execution time on hardware, as well as physical validation using real-world signals collected by the LIME prototype.