Chee-Kheng Chng, Adelaide University; Timothy Payne, Lockheed Martin Corporation; Tat-Jun Chin, Adelaide University
https://doi.org/10.64861/JCFM1316
Keywords: Streak detection, Radon Transform, Deep Learning, SDA
Abstract:
Maintaining a comprehensive resident space object (RSO) catalogue is a cornerstone of Space Domain Awareness (SDA) and requires reliable detection of RSOs in optical imagery. When a sensor is inertially fixed, a moving RSO appears as a streak in long-exposure images, often at low contrast-to-noise ratio (CNR). Our goal is to improve sensitivity to faint streaks while keeping false alarms low enough for operational screening. Existing detectors largely fall into two categories. Image domain methods are typically conservative and achieve high precision, but can miss faint streaks when their intensity approaches the noise floor. Radon transform (RT) methods accumulate signal and therefore achieve high recall, but they often generate many false positives because stars, hot pixels, and processing artifacts can produce strong RT responses that are difficult to disambiguate. As a result, current pipelines frequently face a tradeoff between sensitivity and manageable candidate volume._x000D_
_x000D_
Motivated by this complementarity, we propose a dual-stream network that jointly reasons over the image and RT domains. The upper stream extracts RT-domain features, while the lower stream predicts a streak-probability map from the input image. A cross-domain alignment module transforms this probability map into the RT space, where it modulates the RT-domain features through element-wise multiplication. By associating RT features with their corresponding image-domain evidence, the aligned representation helps the network distinguish genuine streak responses from those induced by artifacts. We train the network on semi-synthetic data generated by inserting synthetic streaks into real background images and evaluate it on both unseen semi-synthetic data and real observations. Compared with the RT-only and image-only baselines, the proposed model reduces false negatives by 40% and 60%, respectively, while reducing false alarms by more than 55% and approximately 27%. These results demonstrate that explicitly aligning and jointly reasoning across the image and RT domains enables more reliable faint-streak detection. This capability lowers screening workload and supports more complete catalogue maintenance in SDA pipelines._x000D_
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications