Deployment of YOLOv8n for chrysanthemum growth stage detection on an edge AI device

Abstract

This paper presents the deployment and evaluation of the YOLOv8n model for detecting the growth stages of Chrysanthemum indicum L. (CIL) on the Infineon PSOC™ Edge E84 AI Kit, a microcontroller-class Edge AI platform integrated with an Arm Ethos-U55 neural processing unit (NPU). A field image dataset consisting of 9517 images was collected from the Nghia Trai medicinal herb village in Hung Yen Province, Vietnam, and annotated based on the knowledge of agricultural experts. The YOLOv8n model was post-training quantized to INT8, reducing the model size by 73.4% compared with the FP32 TFLite version. A memory placement strategy was applied by storing the read-only model section in Flash while allocating execution memory in system SRAM (SoCMEM). The INT8 version of the model achieved an mAP50 of 0.831 on the test set. After deployment on the PSOC™ Edge E84 device, the system achieved a processing latency of 233 ms per image, approximately 3.39 FPS; an average power consumption of 0.97 W, an energy consumption of 0.226 J per detection, including both inference and post-processing, and memory requirements of approximately 3.27 MB RAM and 2.68 MB Flash. The achieved performance is suitable for periodic monitoring of medicinal chrysanthemum growth stages, where growth-related changes occur slowly and continuous high-frame-rate processing is not required.