What is NEODetect and how it works

The temporal streak detection

Unlike conventional preprocessing approaches, which typically rely on a single image or a difference image focused on transient phenomena, the NEODetect system employs a preprocessing pipeline based on sequences of three images aligned via stellar registration. For both conceptual clarity and ease of visualization, the measurements are represented as three-channel, 16-bit BGR images, where the B, G, and R channels correspond to observations taken at epochs T-1, T0, and T+1, respectively. Beyond standard dark, flat, bias calibration and iterative median filtering of the background, no additional preprocessing steps are applied. The principal objective is to enable the detection of streaks at the lowest possible signal-to-noise ratios (SNR). To avoid introducing interpolation artifacts, image alignment is performed using only pixel shifts. Inevitably, subpixel displacements during acquisition or variations in atmospheric seeing introduce differences in the images of otherwise static sources (e.g., stars, galaxies). In difference imaging, such discrepancies may persist and cannot be fully eliminated through subtraction. Consequently, a neural network trained solely on difference images cannot account for the origin of these signals. In contrast, the NEODetect input data preserves all such information in its original form, thereby providing the neural network with the opportunity to learn and distinguish various phenomena arising during the imaging process (e.g., cosmic rays, CCD amplification artifacts, diffraction spikes, seeing variations).
Following semantic segmentation by the neural network, all regions across the three temporal frames that may correspond to NEO candidates are identified. The initial step is to merge candidate detections across the image plane axes. Due to the high-noise environment, detected streaks often appear as discontinuous fragments rather than contiguous features. These fragments are typically collinear and share a common orientation, as determined by line fitting. After merging, the system searches across the temporal axis for collinear, directionally consistent segments, taking into account both the acquisition times and the exposure fill factor. If collinear candidate fragments are identified, share a consistent direction, and exhibit a coherent temporal sequence with the T0 segment always centrally positioned, the NEO candidate is added to a provisional list.

The imposition of chronological alignment – as a constraint – results in the automatic exclusion of 99.2% of false positives: in the test dataset, 71,995,819 of 72,575,480 candidates were classified as pixel noise or cosmic ray artifacts. The remaining 579,661 candidates, however, still represent an excessive number of false positives. To further minimize this figure, an additional criterion called the temporal consequence filter is enforced: there must exist at least one subsequent triplet of images containing a consistent continuation of the candidate within the measurement sequence, i.e., an independent NEO candidate whose position, direction, and apparent speed are compatible with the preceding detection. Only detections that satisfy this condition are reported and stored in the database.