Speaker
Description
Detecting faint, linear satellite streaks in astronomical images is challenging due to high sensor dynamic range, stellar clutter, and atmospheric noise. Standard global normalization techniques often cause information loss by drowning these weak features in the background noise. To compensate for this, we propose a feature engineering pipeline that transforms raw FITS data into a three-channel composite image. The pipeline integrates multi-percentile contrast optimization to handle high dynamic range, a 2D Gabor filter bank to extract linear features across multiple orientations and spatial frequencies, and Contrast Limited Adaptive Histogram Equalization (CLAHE) for enhanced local contrast and noise suppression. These engineered channels directly feed geometric and intensity priors into a YOLOv8 object detection network, yielding consistent results throughout different confidence thresholds.