Research and Experiments on Adaptive Root Cutting Using a Garlic Harvester Based on a Convolutional Neural Network

Abstract

Aimed at the problems of a high leakage rate, a high cutting injury rate, and uneven root cutting in the existing combined garlic harvesting and root-cutting technology, we researched the key technologies used in a garlic harvester for adaptive root cutting based on machine vision. Firstly, research was carried out on the conveyor alignment and assembly of the garlic harvester to realize the adjustment of the garlic plant position and the alignment of the bulb’s upper surface before the roots were cut, to establish the parameter equations and to modify the structure of the conveyor to form the adaptive garlic root-cutting system. Then, a root-cutting test using the double-knife disk-type cutting device was carried out to examine the root-cutting ability of the cutting device. Finally, a bulb detector trained with the IRM-YOLO model was deployed on the Jetson Nano device (NVIDIA, Jetson Nano(4GB), Santa Clara, CA, USA) to conduct a harvester field trial study. The pass rate for the root cutting was 82.8%, and the cutting injury rate was 2.7%, which tested the root cutting performance of the adaptive root cutting system and its field environment adaptability, providing a reference for research into combined garlic harvesting technology.