영상 데이터 감쇄에 따른 보행자·차량 객체 탐지AI 신뢰도의 비선형적 성능 저하 메커니즘과 정량적 변곡점 규명: COCO 데이터 세트 기반 YOLOv8 사례연구

초록

This study quantitatively analyzes, as a case study, the impact of image resolution degradation and JPEG compression on AI object detection reliability, focusing on the pedestrian (Person) and vehicle (Car) classes relevant to autonomous driving and smart-city surveillance, using the COCO 2017 dataset and YOLOv8 (Nano). Recognition confidence does not decrease linearly with degradation; it maintains performance up to a threshold before a non-linear step-drop, with confidence falling sharply near a 0.5 resolution-reduction ratio and a JPEG quality of 15. This observation, specific to the present experimental setting, provides preliminary evidence for a minimum-reliability safety lower bound (average confidence 0.38). A comparison with prior studies shows the non-linear collapse pattern is broadly reproducible, though exact thresholds vary by dataset and model. These findings serve as a preliminary reference, pending further validation, for balancing bandwidth efficiency and AI reliability in system design.

키워드