Real-Time Weed Detection on Low-Cost Embedded Devices Using Quantized YOLO

Abstract

This research addresses industrial-scale data science challenges in precision agriculture through the development of an affordable embedded system for real-time weed detection. By applying advanced deep learning optimization techniques to YOLOv5 neural networks deployed on edge computing platforms, we demonstrate how industrial machine vision can be democratized for smaller operations. Our methodology combines post-training quantization (16-bit floating point, 8bit integer) and ONNX format conversion with systematic performance benchmarking on the Raspberry Pi 4B platform, representing a generalizable approach to edge AI deployment in industrial settings. Comprehensive evaluation on agricultural image data revealed two optimal configurations for different industrial use cases: YOLOv51 with int8 quantization achieved highest detection accuracy (mAP @ 0.5: 0.815) for stationary applications, while YOLOv5n with ONNX conversion delivered real-time performance (2.66 FPS) with acceptable accuracy (mAP@0.5: 0.727), enabling tractor-mounted operations at speeds up to 10 km / h. Beyond agriculture, this quantization framework offers valuable insights for industrial IoT applications requiring efficient AI deployment in resource-constrained environments. The system represents a significant advancement in affordable industrial machine vision, potentially reducing agricultural chemical usage by 60-80% through site-specific application-demonstrating how edge computing and deep learning optimization can deliver substantial economic and environmental benefits across industrial sectors.