Research On Key Technology Of Online Root Cutting Of Garlic Based On Machine Vision

As an important spice vegetable,garlic is loved by consumers for its unique flavor and is a necessary seasoning for Chinese people.China is a major country in garlic production,consumption and export,with an annual harvesting area of over 800,000 hm~2 and an output of over 20 million tons,both of which rank first in the world.The mechanization of the whole process of garlic seedling starting,seedling cutting,root cutting and harvesting will be necessary and inevitable for the production.However,due to the diverse varieties of garlic,complicated patterns,large differences in bulbs and strong flexibility of garlic roots,there is still no garlic harvester with satisfying root cutting function in domestic and international,and garlic root cutting is still mainly done manually.In this study,the key technology research of online root cutting of garlic based on machine vision was carried out to address the problems of combined harvesting and root cutting technology of garlic;we combined pressure sensor sensing technology,established a dynamic force analysis model of the cutting root,optimized the key structure and parameters of the cutting device,and researched methods to improve the cutting root capacity and root neatness of the cutting device;a comparative study of the detection performance of the bulb target was conducted by comparing three features in garlic images,determining the bulb feature as the detection target,and improving the YOLOv2 model using a Res Net50-based feature extraction network;with lightweight and high performance as the optimization goal,we propose a quantitative analysis method for image brightness,optimize the training parameters of the Mobile Net V2-based IRM-YOLOv2 model,propose the concept of predicted cut line reliability,and compare and study the detection performance of different algorithms;establishing the calculation equations of the speed ratio of the clamping conveyor chain and the finger chain to analyze the kinematics of the garlic plant during the online root cutting and harvesting process;a set of equations for the motion relationship between the rotation angle of the dual-axis stepper motor and the height variation of the cutting device is derived,and a machine vision-based online garlic root cutting control algorithm is investigated;The Jetson Nano-based online root cutting system for garlic was tested in bench and field trials to verify the technical feasibility and field applicability of the online root cutting system.The main research is summarized as follows:(1)Study of a double-disc garlic root cutting device.Based on the statistical data of bulb dimensions,a mathematical model of the key structure of the double-disc knife was established and the main structural parameters of the cutting device were determined.Combining the cutting model and energy conservation analysis,the influence law of disc knife structure and parameters on root cutting quality was analyzed,and the structure and parameters of disc knife were optimized.On the basis of theoretical analysis,a double disc knife garlic root cutting test bench was built.Experimental studies on the residual root length and cutting force were carried out with the disc knife rotation direction and structure as the test factors.The change of force on garlic roots during root cutting was quantitatively analyzed.The root cutting mechanism of the double disc knife was revealed.And through the comparison test,the tooth-shaped disc knife root cutting device was preferably selected,and the root cutting effect was effectively improved.(2)Development and experimentation of a convolutional neural network-based target detection algorithm for garlic.The camera aberration mechanism is analyzed by the central perspective imaging model,and the camera aberration is corrected.An image acquisition test bench was built and a Matlab-based human-computer interface was developed for the test bench to obtain garlic images of different brightness.With the help of Matlab deep learning platform,the YOLOv2 model was improved by introducing a residual network based on the Res Net50 feature extraction network.Three features in garlic images were compared to identify bulbs as detection targets,and cut lines were proposed based on the target detection algorithm.The improved YOLOv2 model is compared and analyzed with 10 models based on 5 algorithms to verify the effectiveness of the algorithm improvements.(3)Target detection algorithm optimization and experimentation.In order to better compare the data information,a quantitative analysis method of image brightness is proposed,and the means of image brightness description is optimized.The K-means clustering algorithm was used to cluster the training data and analyze the relationship between the number of anchor frames and the average IOU.With the research direction of lightweight and high performance,and aiming at reducing the amount of operations,we propose the IRM-YOLOv2 detection model based on Mobile Net V2 by decomposing the standard convolution into deep convolution and point-by-point convolution.Borrowing from SOTA detectors,a warm-up mechanism was introduced into the training process.From the need to evaluate the accuracy of the detection model in predicting the position of the cut line,the reliability evaluation method of the target detection model in predicting the cut line was innovatively proposed.Based on the clarification of the relationship between each key training parameter and detector performance,an experimental study is conducted with the small batch size and training period of three lightweight feature extraction networks as the test factors,and the confidence score,accuracy,reliability,detection time and training time as the indexes.The feature extraction visualization reveals the reasons for the differences in detection performance of different models.A further comparative study of the detection performance of different algorithms verifies the reliability of the IRM-YOLOv2 detection model.(4)A study on garlic root cutting method based on convolutional neural network.Based on further comparison of YOLOv2,YOLOv3 and YOLOv4 algorithms,it was verified that IRM-YOLOv2 can be effectively applied to bulb target detection.The structural characteristics and operating principles of the conveyor alignment assembly were clarified,and the parametric equations for the speed relationship between the clamping conveyor chain and the finger paddle chain were established.The feasibility of implementing online root cutting on garlic combine harvesters was explored.A deep learning based garlic root cutting test bench was built and the decimal digital command code used to communicate between the upper and lower UARTs was programmed.A testbed-based root-cutting control algorithm was developed,and the machine-vision-based garlic root-cutting process was explored.(5)Garlic online root cutting system design and experiment.A calculation equation was established with the angle between the garlic plant and the clamping conveyor chain as the parameter.A calculation equation was established with the angle between the garlic plant and the clamping conveyor chain as the parameter,and analyzed the effect of the speed ratio between the clamping conveyor chain speed and the finger chain speed on the position adjustment of the garlic plant and the alignment of the upper surface of the bulb.The algorithm of online root-cutting control was developed by analyzing the motion of the parallel four-bar mechanism and deriving a set of equations for the motion relationship between the rotation angle of the double-output stepper motor and the height change of the cutting device.Constructed garlic online root cutting system,designed online root cutting conveyor alignment assembly and online root cutting control system hardware part.The software part of the online root-cutting control system containing the online root-cutting control algorithm was developed using Matlab,GPU Coder,STM32Cube MX and Keil5 software.And the Jetson Nano is the upper computer and the STM32F103development board is the lower computer.Root cutting performance verification tests were conducted through bench tests and field trials.The results of field trials showed that,1)the online root cutting system effectively reduced the cutting injury rate of bulbs by dynamically adjusting the cutting height under normal delivery of garlic plants;2)the technical feasibility and field applicability of the online root cutting system were verified.