Stochastic Degradation Coverage Training for Robust Small-UAV Detection in Grayscale Low-Altitude Surveillance Images

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Reliable detection of small unmanned aerial vehicles (UAVs) is difficult in grayscale imagery degraded by low illumination, reduced contrast, noise, and blur. We propose stochastic degradation coverage training (SDCT), an offline strategy that broadens training data without changing the detector. A single-channel YOLOv8n model, denoted YOLOv8n-Gray, is trained using clean images and five degraded replicas per image. On the DUT Anti-UAV-derived benchmark, three-seed mean mAP50 on the seen degraded suite increases from 0.7369 with exposure-matched Clean Repeat to 0.8419. On the DUT-Adv target-domain test package, mAP50 is 0.8089 for Clean Repeat, 0.8519 for DART, and 0.8481 for SDCT. On FRED’s 47-sequence canonical test split, pooled mAP50 is 0.0062, 0.0052, and 0.0059, respectively, with Clean Repeat reporting the highest value. On the 13-sequence challenging intersection, SDCT exceeds Clean Repeat (0.0080 vs. 0.0073), while DART is highest (0.0088). Thus, FRED indicates a conditional gain over clean repetition rather than consistent SDCT superiority. Under a synthetic held-out stress test, SDCT achieves mean mAP50 of 0.6939. TensorRT FP16 inference on Jetson AGX Orin achieves 4.23 ms latency and 236.36 FPS. External test sets were used for final evaluation.