Target-Adaptive Neural Architecture Search in YOLOv9 Machine Vision Models

초록

Recent advances in real-time object detection have been driven by YOLO models, which effectively balance accuracy and speed. However, architectures optimized under a fixed search configuration often show limited adaptability when applied to diverse deployment targets. To address this limitation, this paper introduces a target-adaptive YOLOv9 neural architecture search (NAS) algorithm that applies NAS to the neck block of YOLOv9 through a once-for-all supernet. By defining multiple search scopes with varying exploration ranges, the framework enables the automatic generation of sub-networks tailored to different target requirements without additional retraining. Experimental results demonstrate that the small configuration achieves 73.4% precision and 53.2% mAP50, confirming an effective balance between accuracy and model efficiency. The proposed approach facilitates scalable deployment across diverse target settings by flexibly adjusting the architectural search range.

키워드