Fast reconstruction-based ROI triggering via anomaly detection in the CYGNO optical TPC

TL;DR

The paper addresses real-time data reduction for optical-readout TPCs by introducing an unsupervised, reconstruction-based anomaly-detection pipeline trained on pedestal frames to model detector noise. Residual maps from a pedestal-trained convolutional autoencoder enable ROI extraction that preserves the majority of signal while discarding most background, achieving $93.0 \pm 0.2\%$ signal coverage and a $97.8 \pm 0.1\%$ area reduction, with about $25\ \mathrm{ms}$ per-frame latency. A key insight is that the training objective critically shapes ROI localization, with a refined objective that suppresses reconstruction of faint structured deviations outperforming a baseline hybrid loss and a simple Gaussian baseline. The approach is detector-agnostic, calibration-light, and suitable for online trigger pipelines, offering a practical baseline for ML-assisted data reduction in future CYGNO detectors and other optical TPCs.

Abstract

Optical-readout Time Projection Chambers (TPCs) produce megapixel-scale images whose fine-grained topological information is essential for rare-event searches, but whose size challenges real-time data selection. We present an unsupervised, reconstruction-based anomaly-detection strategy for fast Region-of-Interest (ROI) extraction that operates directly on minimally processed camera frames. A convolutional autoencoder trained exclusively on pedestal images learns the detector noise morphology without labels, simulation, or fine-grained calibration. Applied to standard data-taking frames, localized reconstruction residuals identify particle-induced structures, from which compact ROIs are extracted via thresholding and spatial clustering. Using real data from the CYGNO optical TPC prototype, we compare two pedestal-trained autoencoder configurations that differ only in their training objective, enabling a controlled study of its impact. The best configuration retains (93.0 +/- 0.2)% of reconstructed signal intensity while discarding (97.8 +/- 0.1)% of the image area, with an inference time of approximately 25 ms per frame on a consumer GPU. The results demonstrate that careful design of the training objective is critical for effective reconstruction-based anomaly detection and that pedestal-trained autoencoders provide a transparent and detector-agnostic baseline for online data reduction in optical TPCs.

Figures (6)

• Figure 1: Schematic representation of the convolutional autoencoder architecture.
• Figure 2: Example of synthetic perturbations injected during training of the refined autoencoder. From left to right: (a) clean pedestal frame used as reconstruction target; (b) corrupted input obtained by injecting synthetic curved strokes and Gaussian blobs with varying amplitude; (c) absolute difference between input and target (shown for visualization); (d) binary injection mask m used to up-weight the reconstruction loss in perturbed regions. The injected structures are generic and detector-agnostic, and serve solely to regularize the reconstruction objective
• Figure 3: Representative Regions of Interest (ROIs) returned by the anomaly-detection framework. Each row shows: (a) the fiducialized camera image; (b) the anomaly map, where track-like structures appear as localized high-residual regions; (c) the final ROI mask after spatial aggregation. The ROIs reliably enclose particle-induced structures while excluding noise-dominated background regions.
• Figure 4: Trade-off between mean signal-intensity coverage and mean area cut for the three anomaly-scoring approaches, obtained by sweeping the residual threshold $\tau$. All methods share the same ROI-extraction pipeline and are evaluated on the same event sample. Curves closer to the top-right indicate stronger compression at fixed signal retention.
• Figure 5: Signal-intensity coverage as a function of reconstructed event energy for the refined-training autoencoder. The method maintains high coverage across the full energy range. A small number of outliers at very low coverage are discussed separately in Section \ref{['sec:lowcases']}.