Edge AI rice disease detection: A hardware performance comparison

Abstract

This study addresses the challenge of real-time rice disease detection under resource-constrained conditions by developing an edge AI
system for rice leaf disease identification without cloud dependency. A lightweight YOLOv5n object detection model was trained on a rice disease dataset and optimized using post-training quantization. The quantized model was deployed on an ultra-low-power microcontroller (MCU) integrated with an Arm Ethos-U55 Neural Processing Unit (NPU), and its performance was compared against a Raspberry Pi 4 and a workstation with a high-performance CPU. Results show that the quantized model maintains high detection accuracy (mAP > 90%) and achieves real-time inference on the microcontroller (around 16 FPS) at only 1.53W, roughly 10 times faster and with 54% lower power consumption compared to the Raspberry Pi. While the CPU performance reached the fastest inference (9.5 ms), its energy consumption was significantly higher. In conclusion, the research demonstrates the feasibility of deploying quantized vision models on low-power edge devices for smart agriculture. The findings highlight the trade-offs between performance and energy efficiency, marking a successful implementation of quantized YOLOv5n on a microcontroller NPU in the smart agricultural sector.