李玉華，劉全程，李天華，吳彥強(qiáng)，牛子孺，侯加林

基于Jetson Nano處理器的大蒜鱗芽朝向調(diào)整裝置設(shè)計(jì)與試驗(yàn)

李玉華1,2,3，劉全程1，李天華1,2,3，吳彥強(qiáng)1,2,3，牛子孺1,2,3，侯加林1,2,3※

（1. 山東農(nóng)業(yè)大學(xué)機(jī)械與電子工程學(xué)院，泰安 271018；2. 山東省農(nóng)業(yè)裝備智能化工程實(shí)驗(yàn)室，泰安 271018；3. 山東省園藝機(jī)械與裝備重點(diǎn)實(shí)驗(yàn)室，泰安 271018）

為滿足大蒜定向播種的農(nóng)藝要求，針對(duì)現(xiàn)有大蒜鱗芽調(diào)整方法對(duì)雜交蒜適應(yīng)性差的問(wèn)題，該研究設(shè)計(jì)了一種基于Jetson Nano處理器的大蒜鱗芽朝向自動(dòng)調(diào)整裝置。采用雙卷積神經(jīng)網(wǎng)絡(luò)模型結(jié)構(gòu)，其中一個(gè)神經(jīng)網(wǎng)絡(luò)模型對(duì)大蒜是否被喂入進(jìn)行實(shí)時(shí)監(jiān)測(cè)，檢測(cè)到大蒜喂入調(diào)整裝置后，一個(gè)ResNet-18網(wǎng)絡(luò)模型對(duì)蒜種鱗芽朝向進(jìn)行判斷，當(dāng)鱗芽朝上時(shí)大蒜鱗芽調(diào)整機(jī)構(gòu)打開Y型料斗使大蒜以鱗芽朝上的姿態(tài)直接落下，當(dāng)鱗芽朝下時(shí)大蒜鱗芽調(diào)整機(jī)構(gòu)翻轉(zhuǎn)180°帶動(dòng)大蒜一起翻轉(zhuǎn)后以鱗芽朝上的姿態(tài)落下，實(shí)現(xiàn)大蒜鱗芽朝向?qū)崟r(shí)調(diào)整。神經(jīng)網(wǎng)絡(luò)模型推理及舵機(jī)控制采用英偉達(dá)邊緣計(jì)算處理器Jetson Nano進(jìn)行處理。利用離散元分析軟件EDEM結(jié)合正交試驗(yàn)方法對(duì)調(diào)整裝置的關(guān)鍵結(jié)構(gòu)參數(shù)進(jìn)行優(yōu)化，并以雜交大蒜為試驗(yàn)對(duì)象進(jìn)行臺(tái)架試驗(yàn)，試驗(yàn)結(jié)果表明：大蒜鱗芽調(diào)整成功率為96.25%，模型推理時(shí)間0.045 s，平均每粒大蒜調(diào)整時(shí)間為0.785 s，滿足大蒜播種機(jī)播種要求。該文研究結(jié)果可為解決雜交大蒜直立播種問(wèn)題及邊緣計(jì)算在精密播種設(shè)備中的應(yīng)用提供有益參考。

機(jī)器視覺(jué)；深度學(xué)習(xí)；邊緣計(jì)算；Jetson Nano處理器；大蒜；鱗芽朝向

0 引 言

定向播種[1]對(duì)大蒜產(chǎn)量、品質(zhì)及出苗一致性有極其重要的影響[1-3]，并且便于對(duì)薄膜覆蓋大蒜進(jìn)行破膜。單粒定向播種是大蒜種植的基本要求，大蒜播種密度大，人工勞動(dòng)強(qiáng)度極大，可靠的自動(dòng)化大蒜播種機(jī)是當(dāng)前大蒜種植戶的迫切需求，而大蒜鱗芽自動(dòng)調(diào)整是機(jī)械化播種推廣的關(guān)鍵技術(shù)[4]。

近年眾多學(xué)者及科研人員對(duì)大蒜鱗芽朝向自動(dòng)調(diào)整進(jìn)行了研究。其中典型機(jī)械式鱗芽朝向調(diào)整裝置有雙鴨嘴式大蒜鱗芽調(diào)整裝置[5]、三級(jí)錐形料斗鱗芽調(diào)整裝置[6]、錐形螺旋導(dǎo)管自動(dòng)定向裝置[7]和旋轉(zhuǎn)式蒜瓣單粒定向取種裝置[8]。這些鱗芽朝向調(diào)整裝置主要利用大蒜重心靠近底部或鱗芽狹長(zhǎng)的物理特點(diǎn)進(jìn)行鱗芽朝向調(diào)整，對(duì)于形狀規(guī)則的薹蒜等品種具有良好的效果，但對(duì)雜交大蒜等形狀不規(guī)則的大蒜品種則效果不理想。

近年隨著機(jī)器視覺(jué)[9-10]及深度學(xué)習(xí)[11-12]應(yīng)用的發(fā)展，在雜草識(shí)別[13-14]、害蟲檢測(cè)[15-17]、水果采摘[18]及分揀等農(nóng)業(yè)領(lǐng)域已有眾多研究及應(yīng)用，技術(shù)已相對(duì)成熟。已有學(xué)者利用機(jī)器視覺(jué)對(duì)大蒜鱗芽識(shí)別和朝向調(diào)整進(jìn)行了研究，楊清明等[19]利用數(shù)字圖像形態(tài)學(xué)對(duì)大蒜圖像進(jìn)行處理，進(jìn)而判斷大蒜鱗芽位置；方春等[20]利用CNN的局部感知特點(diǎn)，提出了基于深度學(xué)習(xí)的大蒜鱗芽朝向識(shí)別算法；吳獻(xiàn)等[21]通過(guò)對(duì)蒜瓣樣本圖像進(jìn)行形態(tài)學(xué)處理，采用種觀測(cè)窗的方法識(shí)別定位蒜瓣的尖角位置；趙麗清等[22]通過(guò)對(duì)大蒜圖像形態(tài)學(xué)處理后運(yùn)用質(zhì)心中心判別法判定大蒜的朝向并進(jìn)行定位。

上述關(guān)于大蒜鱗芽調(diào)整的研究只是針對(duì)鱗芽識(shí)別算法層面，還未對(duì)機(jī)械和控制相關(guān)的應(yīng)用進(jìn)行具體研究。此外，侯加林等[23]設(shè)計(jì)了一款大蒜鱗芽調(diào)整試驗(yàn)裝置，通過(guò)樹莓派對(duì)大蒜圖像進(jìn)行處理并判斷鱗芽朝向。Li等[24]也提出了一種基于機(jī)器視覺(jué)的大蒜鱗芽朝向調(diào)整裝置，利用LabVIEW[25-26]實(shí)現(xiàn)圖像采集、預(yù)處理、鱗芽識(shí)別及調(diào)整機(jī)構(gòu)控制。但樹莓派的運(yùn)算能力相對(duì)較弱，無(wú)法滿足深度學(xué)習(xí)網(wǎng)絡(luò)對(duì)運(yùn)算能力的需求；LabVIEW是基于PC的開發(fā)環(huán)境，體積大，成本高，無(wú)法在大蒜播種機(jī)上應(yīng)用。

本文針對(duì)上述大蒜鱗芽調(diào)整裝置存在的問(wèn)題，以邊緣計(jì)算處理器Jetson Nano[27-28]為平臺(tái)，以深度學(xué)習(xí)網(wǎng)絡(luò)為框架設(shè)計(jì)了一種大蒜鱗芽朝向自動(dòng)調(diào)整裝置，并進(jìn)行了臺(tái)架試驗(yàn)，以期為大蒜智能化定向播種機(jī)的設(shè)計(jì)提供參考。

1 總體結(jié)構(gòu)與工作原理

1.1 總體結(jié)構(gòu)

大蒜鱗芽朝向調(diào)整裝置主要由蒜種喂入通道、攝像頭、攝像頭支架、鱗芽調(diào)整機(jī)構(gòu)、翻轉(zhuǎn)舵機(jī)和支架等組成。其中鱗芽調(diào)整機(jī)構(gòu)由翻轉(zhuǎn)架、復(fù)位彈簧、料斗開合舵機(jī)、Y型料斗Ⅰ和Y型料斗Ⅱ等組成，如圖1所示。翻轉(zhuǎn)舵機(jī)驅(qū)動(dòng)整個(gè)鱗芽調(diào)整裝置翻轉(zhuǎn)180°，料斗開合舵機(jī)驅(qū)動(dòng)2個(gè)Y型料斗張開或閉合，復(fù)位彈簧兩端分別與2個(gè)Y型料斗連接并處于拉伸狀態(tài)確保2個(gè)Y型料斗打開后的自動(dòng)復(fù)位。

Y型料斗Ⅰ和Y型料斗Ⅱ采用非對(duì)稱結(jié)構(gòu)，通過(guò)優(yōu)化2個(gè)料斗的結(jié)構(gòu)參數(shù)保證大蒜進(jìn)入料斗后處于直立狀態(tài)。Y型料斗Ⅰ設(shè)立蒜種滑落引導(dǎo)槽，防止鱗芽調(diào)整機(jī)構(gòu)翻轉(zhuǎn)時(shí)大蒜自旋轉(zhuǎn)而導(dǎo)致鱗芽調(diào)整失效。

1.蒜種喂入通道 2.攝像頭 3.攝像頭支架 4.鱗芽調(diào)整機(jī)構(gòu) 5.翻轉(zhuǎn)舵機(jī) 6.支架 7.翻轉(zhuǎn)架 8.復(fù)位彈簧 9.Y型料斗Ⅰ 10.料斗開合舵機(jī) 11.Y型料斗Ⅱ

1.Garlic seeds feeding channel 2.Camera 3.Camera bracket 4.Garlic bulbil orientation adjustment mechanism 5.Servos for turning 6.Bracket 7.Tilting frame 8.Reset springs 9.Y-type hopper Ⅰ 10.Servos for hopper opening and closing 11.Y-type hopper Ⅱ

圖1 大蒜鱗芽調(diào)整裝置結(jié)構(gòu)示意圖

Fig.1 Structure diagram of garlic bulbil orientation adjustment device

1.2 工作原理

邊緣計(jì)算處理器Jetson Nano通過(guò)大蒜檢測(cè)深度學(xué)習(xí)網(wǎng)絡(luò)實(shí)時(shí)處理攝像頭抓取的圖像并判斷是否有大蒜從蒜種喂入通道進(jìn)入Y型料斗。當(dāng)檢測(cè)到大蒜進(jìn)入料斗后，立即通過(guò)鱗芽朝向判斷深度學(xué)習(xí)網(wǎng)絡(luò)對(duì)攝像頭抓取的圖像進(jìn)行大蒜鱗芽朝向識(shí)別。當(dāng)大蒜鱗芽朝上時(shí)，Jetson Nano處理器通過(guò)PWM信號(hào)控制料斗開合舵機(jī)旋轉(zhuǎn)一定角度，使Y型料斗Ⅰ和Y型料斗Ⅱ下端打開，大蒜以鱗芽朝上的狀態(tài)直接落入插播裝置；當(dāng)大蒜鱗芽朝下時(shí)，Jetson Nano處理器通過(guò)PWM信號(hào)控制翻轉(zhuǎn)舵機(jī)旋轉(zhuǎn)180°，帶動(dòng)鱗芽調(diào)整機(jī)構(gòu)一起翻轉(zhuǎn)，此時(shí)大蒜將沿著Y型料斗Ⅰ的引導(dǎo)槽滑落并翻轉(zhuǎn)180°，實(shí)現(xiàn)大蒜鱗芽朝向的調(diào)整。

大蒜檢測(cè)深度學(xué)習(xí)網(wǎng)絡(luò)通過(guò)對(duì)大蒜和背景圖像進(jìn)行實(shí)時(shí)分割進(jìn)行大蒜檢測(cè)，判斷是否有大蒜喂入鱗芽調(diào)整裝置。鱗芽朝向判斷深度學(xué)習(xí)網(wǎng)絡(luò)對(duì)大蒜鱗芽的朝向進(jìn)行實(shí)時(shí)判斷，從而為鱗芽調(diào)整裝置的控制提供依據(jù)。大蒜檢測(cè)和鱗芽朝向判斷采用大蒜檢測(cè)和大蒜鱗芽朝向識(shí)別2個(gè)網(wǎng)絡(luò)模型，以提高識(shí)別準(zhǔn)確率，便于訓(xùn)練，提高系統(tǒng)的實(shí)時(shí)性。

2 機(jī)械系統(tǒng)優(yōu)化設(shè)計(jì)

前期理論分析及單因素試驗(yàn)表明，大蒜鱗芽朝向調(diào)整裝置的結(jié)構(gòu)參數(shù)對(duì)鱗芽調(diào)整合格率及效率具有較大影響。尤其是2個(gè)Y型料斗的結(jié)構(gòu)參數(shù)。理想的Y型料斗結(jié)構(gòu)應(yīng)使大蒜喂入后處于直立狀態(tài)，且結(jié)構(gòu)緊湊，便于在整機(jī)中的布置。大蒜在料斗內(nèi)處于直立狀態(tài)（鱗芽朝上或朝下），可以保證大蒜鱗芽調(diào)整的成功率，且大蒜直立狀態(tài)時(shí)從上部拍攝的鱗芽圖像特征更便于神經(jīng)網(wǎng)絡(luò)訓(xùn)練和提高識(shí)別準(zhǔn)確率。Y型料斗的高度尺寸應(yīng)盡量減小有利于減小調(diào)整機(jī)構(gòu)翻轉(zhuǎn)時(shí)所需要的空間，便于整體布置，縮短翻轉(zhuǎn)時(shí)大蒜滑落所有需要的時(shí)間，提高調(diào)整效率，并防止調(diào)整后的大蒜在下落過(guò)程中再次翻轉(zhuǎn)導(dǎo)致調(diào)整失敗。為找到最優(yōu)的結(jié)構(gòu)參數(shù)，根據(jù)對(duì)大蒜鱗芽調(diào)整裝置的結(jié)構(gòu)參數(shù)的要求，采用離散元仿真試驗(yàn)對(duì)Y型料斗I的結(jié)構(gòu)進(jìn)行優(yōu)化。

2.1 離散元仿真模型

為獲得Y型料斗的最優(yōu)結(jié)構(gòu)參數(shù)，本文采用離散元分析軟件EDEM對(duì)大蒜鱗芽調(diào)整裝置性能進(jìn)行仿真分析。以雜交大蒜為仿真分析對(duì)象，按體積分為大、中、小3級(jí)，每級(jí)大蒜籽粒的尺寸根據(jù)體積按正態(tài)分布進(jìn)行設(shè)置[27]。顆粒間及顆粒與設(shè)備間的接觸模型均采用Hertz-Mindlin模型，仿真模型及關(guān)鍵結(jié)構(gòu)參數(shù)如圖2所示。

注：為喂入通道角度，（°）；為Y型料斗I底部半徑，mm；為料斗高度差，mm。

Note:is the angle of the feeding channel, (°);is the radius of the Y-hopper I, mm;is the height difference of the hopper, mm.

圖2 EDEM仿真模型及關(guān)鍵結(jié)構(gòu)參數(shù)

Fig.2 EDEM simulation model and key structural parameters

2.2 參數(shù)優(yōu)化

通過(guò)實(shí)物模型前期試驗(yàn)及仿真試驗(yàn)發(fā)現(xiàn)，蒜種喂入通道角度、Y型料斗I底部半徑（后文簡(jiǎn)稱料斗半徑）和料斗高度差對(duì)蒜種喂入Y型料斗后的直立率影響最大。因此取這3個(gè)參數(shù)為仿真試驗(yàn)因素，以蒜種喂入Y型料斗后的直立率為試驗(yàn)響應(yīng)指標(biāo)。根據(jù)前期單因素仿真試驗(yàn)，綜合考慮結(jié)構(gòu)尺寸限制，確定各試驗(yàn)因素的取值范圍，如圖2所示。依據(jù)Box-Benhnken中心組合設(shè)計(jì)理論進(jìn)行試驗(yàn)設(shè)計(jì)，試驗(yàn)因素和水平如表1所示。

表1 試驗(yàn)因素編碼

仿真試驗(yàn)方案與結(jié)果如表2所示，運(yùn)用Design Expert 11.1.2.0數(shù)據(jù)分析軟件進(jìn)行多元回歸擬合分析，得到直立率的回歸方程為

任意2個(gè)因素的相互作用對(duì)直立率的影響如圖3所示。由圖3a可知，隨料斗半徑增加，直立率先升后降。由圖3b可知，隨高度差增加，直立率也會(huì)相應(yīng)升高。由圖3c可知，料斗半徑一定時(shí)，高度差增大直立率提高。

表2 試驗(yàn)設(shè)計(jì)方案及結(jié)果

注：1、2、3表示各因素編碼值

Note:1、2and3represent the coding values of each factor

為確定Y型料斗的最優(yōu)結(jié)構(gòu)參數(shù)，以直立率最高為目標(biāo)函數(shù)，以蒜種喂入通道角度、Y型料斗半徑和料斗高度差為約束條件，建立Y型料斗結(jié)構(gòu)參數(shù)優(yōu)化模型：

利用軟件Design-Expert軟件Optimization Numerical模塊進(jìn)行優(yōu)化，得到直立率最佳的參數(shù)組合為：喂入通道角度為80°，Y型料斗半徑為12.49 mm，料斗高度差為20 mm，此時(shí)模型預(yù)測(cè)的直立率為96.85%?？紤]結(jié)構(gòu)加工要求，對(duì)優(yōu)化參數(shù)進(jìn)行圓整，取Y型料斗半徑為12.50 mm，料斗高度差為20 mm，在相同試驗(yàn)條件下重復(fù)3次試驗(yàn)，結(jié)果取平均值，得到直立率均值為96.05%，試驗(yàn)結(jié)果與優(yōu)化結(jié)果基本一致，誤差小于5%。

3 控制系統(tǒng)

3.1 硬件設(shè)計(jì)

系統(tǒng)硬件由邊緣計(jì)算處理器Jetson Nano、電源模塊、攝像頭、PCA9685模塊、翻轉(zhuǎn)舵機(jī)和料斗開合舵機(jī)構(gòu)成，如圖4所示。

邊緣計(jì)算處理器Jetson Nano是一款低成本高性能的AI嵌入式處理器，可應(yīng)用于圖像分類、目標(biāo)檢測(cè)、分割和語(yǔ)音處理等領(lǐng)域，具體規(guī)格參數(shù)見(jiàn)表3。其中軟件開發(fā)工具包NVIDIA JetPack SDK擁有用于深度學(xué)習(xí)、計(jì)算機(jī)視覺(jué)和加速計(jì)算的CUDA-X加速庫(kù)與應(yīng)用程序接口（API），可極大提高人工智能（AI）應(yīng)用開發(fā)的速度，PCA9685模塊可同時(shí)輸出16路PWM信號(hào)用于對(duì)翻轉(zhuǎn)舵機(jī)和料斗開合舵機(jī)進(jìn)行驅(qū)動(dòng)。

表3 Jetson Nano處理器規(guī)格參數(shù)

3.2 軟件設(shè)計(jì)

3.2.1 系統(tǒng)工作流程

根據(jù)系統(tǒng)的功能要求，設(shè)計(jì)系統(tǒng)控制流程為：初始化后控制攝像頭對(duì)鱗芽調(diào)整機(jī)構(gòu)內(nèi)部圖像進(jìn)行采集，然后利用大蒜檢測(cè)網(wǎng)絡(luò)模型對(duì)采集的圖像進(jìn)行處理，判斷是背景圖像還是大蒜圖像，如果是背景圖像則進(jìn)行下一幀圖像采集并繼續(xù)判斷；如果是大蒜圖像，則將圖像輸入到鱗芽判斷網(wǎng)絡(luò)模型進(jìn)行鱗芽的朝向判斷，如果鱗芽朝上控制料斗開合舵機(jī)打開2個(gè)Y型料斗使大蒜保持直立狀態(tài)落入插播裝置，如果鱗芽朝下則控制翻轉(zhuǎn)舵機(jī)將鱗芽調(diào)整機(jī)構(gòu)翻轉(zhuǎn)180°，帶動(dòng)大蒜一起翻轉(zhuǎn)180°并落下，實(shí)現(xiàn)鱗芽朝向調(diào)整，系統(tǒng)控制流程圖如圖5所示。

3.2.2 大蒜檢測(cè)及鱗芽識(shí)別方法

為了簡(jiǎn)化模型的復(fù)雜度，降低訓(xùn)練難度，提高識(shí)別準(zhǔn)確率，對(duì)大蒜檢測(cè)和鱗芽朝向識(shí)別采用不同的神經(jīng)網(wǎng)絡(luò)模型，網(wǎng)絡(luò)模型結(jié)構(gòu)如圖6所示。

大蒜檢測(cè)模型主要對(duì)采集的大蒜圖像進(jìn)行實(shí)時(shí)處理，判斷是否有蒜種喂入，由于2種圖像特征差異顯著，采用5層卷積神經(jīng)網(wǎng)絡(luò)模型，前4層為4個(gè)由卷積層、ReLU層和池化層構(gòu)成的卷積塊，第5層為全連接層。大蒜鱗芽朝向判斷模型對(duì)大蒜鱗芽朝上或朝下進(jìn)行判斷，采用ResNet-18網(wǎng)絡(luò)結(jié)構(gòu)，通過(guò)遷移學(xué)習(xí)對(duì)網(wǎng)絡(luò)參數(shù)進(jìn)行訓(xùn)練。

大蒜檢測(cè)模型訓(xùn)練用圖像為8 000張RGB圖像，其中4 000張背景圖像，4 000張大蒜圖像，圖像大小為112×112。訓(xùn)練過(guò)程中采用adam優(yōu)化器對(duì)模型進(jìn)行優(yōu)化，設(shè)置初始學(xué)習(xí)率為0.001，批大?。╞achsize）為32。訓(xùn)練結(jié)果如圖7a所示，訓(xùn)練后模型在測(cè)試集的識(shí)別準(zhǔn)確率為99.3%。

大蒜鱗芽朝向判斷模型訓(xùn)練用圖像為4 000張RGB圖像，其中鱗芽朝上的圖像2 000張，大蒜鱗芽朝下圖像2 000張，圖像大小為224×224。為加快訓(xùn)練過(guò)程提高訓(xùn)練準(zhǔn)確率，采用基于ResNet-18網(wǎng)絡(luò)遷移學(xué)習(xí)訓(xùn)練，訓(xùn)練結(jié)果如圖7b所示，訓(xùn)練后模型在測(cè)試集的識(shí)別準(zhǔn)確率為99.4%。

4 臺(tái)架試驗(yàn)

4.1 試驗(yàn)材料與設(shè)備

為驗(yàn)證基于Jetson Nano處理器的大蒜鱗芽朝向調(diào)整裝置的作業(yè)效果，于2020年11月在山東農(nóng)業(yè)大學(xué)107實(shí)驗(yàn)室開展了臺(tái)架試驗(yàn)。試驗(yàn)材料選用國(guó)內(nèi)種植面積最大品種金鄉(xiāng)雜交大蒜為樣本，隨機(jī)選取200粒飽滿的蒜種進(jìn)行試驗(yàn)。

試驗(yàn)設(shè)備為自行搭建的基于邊緣計(jì)算處理器Jetson Nano的大蒜鱗芽調(diào)整裝置試驗(yàn)臺(tái)，結(jié)構(gòu)如圖8所示，主要包括支架、Jetson Nano、PCA9685模塊、蒜種喂入通道、攝像頭、攝像頭支架、復(fù)位彈簧、翻轉(zhuǎn)舵機(jī)、料斗開合舵機(jī)和鱗芽調(diào)整機(jī)構(gòu)等。其中，PCA9685驅(qū)動(dòng)模塊頻率范圍為40-1000 Hz，通道數(shù)為16，電壓為DC 5-10 V；攝像頭具有IMX219的感光芯片，800 W分辨率和77°的對(duì)角視場(chǎng)角。

1.蒜種喂入通道 2.攝像頭 3.支架 4.翻轉(zhuǎn)舵機(jī) 5.鱗芽調(diào)整機(jī)構(gòu) 6.邊緣計(jì)算處理器 7.舵機(jī)控制芯片 8.電源 9.料斗開合舵機(jī)

1.Garlic seed feeding channels 2.Camera 3.Bracket 4.Servos for turning 5.Garlic bulbil orientation adjustment mechanism 6. Edge computing processor 7.Servo control chip 8.Power 9.Servos for hopper opening and closing

圖8 大蒜鱗芽朝向調(diào)整試驗(yàn)平臺(tái)

Fig.8 Test platform of garlic bulbil orientation adjustment

4.2 試驗(yàn)方法

鱗芽朝向調(diào)整要求參照農(nóng)藝及農(nóng)戶實(shí)際播種需求，以地面垂直線為基準(zhǔn)，蒜種鱗芽朝上且傾斜角小于30°視為朝上[29-30]，調(diào)整后鱗芽朝向在此范圍為合格。大蒜播種機(jī)大多采用錐形鴨嘴插播機(jī)構(gòu)，蒜種落入錐形鴨嘴中時(shí)受到鴨嘴結(jié)構(gòu)尺寸的限制限制，大蒜鱗芽處于朝上且傾斜角小于30°或朝下且傾斜角小于30°狀態(tài)。因此試驗(yàn)采用錐形鴨嘴結(jié)構(gòu)承接調(diào)整后的蒜種，鱗芽朝向調(diào)整合格個(gè)數(shù)通過(guò)試驗(yàn)觀察統(tǒng)計(jì)得到，鱗芽朝向調(diào)整成功率計(jì)算公式如式（3）。

式中1為鱗芽朝向調(diào)整成功率，%；0為試驗(yàn)蒜種個(gè)數(shù)；1為鱗芽朝向調(diào)整合格的蒜種個(gè)數(shù)。

識(shí)別成功率是指大蒜鱗芽朝向識(shí)別準(zhǔn)確的概率，計(jì)算公式如式（4）。

式中2為鱗芽朝向識(shí)別成功率，%；2為鱗芽朝向識(shí)別正確的蒜種個(gè)數(shù)。

模型推理時(shí)間為采集圖像到鱗芽識(shí)別完成所需的時(shí)間；調(diào)整時(shí)間為圖像采集完成到大蒜鱗芽朝向調(diào)整完成所需要的總時(shí)間，是圖像采集時(shí)間、模型推理時(shí)間和舵機(jī)驅(qū)動(dòng)鱗芽調(diào)整機(jī)構(gòu)運(yùn)動(dòng)所需要的時(shí)間總和。

4.3 結(jié)果與分析

試驗(yàn)結(jié)果如表4。由表4可知，鱗芽識(shí)別成功率均值為97.25%，鱗芽朝向調(diào)整成功率均值為96.25%，鱗芽朝向調(diào)整成功率低于識(shí)別成功率的原因是識(shí)別正確的大蒜在鱗芽調(diào)整過(guò)程中由于大蒜重心位置不規(guī)則引起大蒜下落時(shí)翻轉(zhuǎn)，最終導(dǎo)致調(diào)整失敗。模型推理時(shí)間均值為0.045 s，調(diào)整時(shí)間均值為0.785 s，其中鱗芽上時(shí)的調(diào)整時(shí)間均值為0.59 s，鱗芽朝下時(shí)均值為0.98 s，鱗芽朝上與鱗芽朝下時(shí)的調(diào)整時(shí)間相差較大，是因?yàn)閷?duì)鱗芽朝下的大蒜進(jìn)行鱗芽朝向調(diào)整需翻轉(zhuǎn)調(diào)整機(jī)構(gòu)180°，然后將翻轉(zhuǎn)機(jī)構(gòu)復(fù)位，而對(duì)于鱗芽朝上的大蒜進(jìn)行鱗芽朝向調(diào)整時(shí)只需將2個(gè)Y型料斗下端打開即可，開合舵機(jī)旋轉(zhuǎn)角度較小，并且2個(gè)舵機(jī)的轉(zhuǎn)動(dòng)角度速度一致，因此鱗芽朝下時(shí)的調(diào)整時(shí)間比鱗芽朝上的長(zhǎng)。

表4 大蒜鱗芽朝向識(shí)別與調(diào)整試驗(yàn)結(jié)果

5 討 論

本文試驗(yàn)結(jié)果的鱗芽朝向調(diào)整時(shí)間均值為0.785 s，其中模型推理時(shí)間均為0.045 s，調(diào)整時(shí)間約為網(wǎng)絡(luò)模型推理時(shí)間的17倍，主要原因是所用舵機(jī)的旋轉(zhuǎn)角速度相對(duì)較小。因此縮短鱗芽調(diào)整時(shí)間，提高鱗芽調(diào)整效率的主要措施是提高舵機(jī)的轉(zhuǎn)速。已知現(xiàn)有的大蒜種植農(nóng)藝為株距0.1～0.12 m，行距0.18～0.20 m。大蒜播種效率通過(guò)公式（5）進(jìn)行計(jì)算：

式中為播種效率，hm2/h；為1次播種行數(shù)；為鱗芽朝向調(diào)整時(shí)間，s。

假設(shè)大蒜大播種機(jī)一次播種行數(shù)為12，當(dāng)鱗芽朝向調(diào)整時(shí)間為0.59 s時(shí)，通過(guò)公式（5）計(jì)算大蒜播種效率為0.132～0.176 hm2/h；當(dāng)鱗芽朝向調(diào)整時(shí)間為0.785 s時(shí)，大蒜播種效率為0.099～0.132 hm2/h；當(dāng)鱗芽朝向調(diào)整時(shí)間為0.98 s時(shí)，大蒜播種效率為0.079～0.106 hm2/h?，F(xiàn)有機(jī)械式鱗芽調(diào)整方法的大蒜播種工作效率在0.05～0.1 hm2/h范圍內(nèi)[6,28]。通過(guò)對(duì)比分析可以看出，本文設(shè)計(jì)的大蒜鱗芽調(diào)整裝置的鱗芽朝向調(diào)整時(shí)間可以滿足播種機(jī)對(duì)播種效率的要求。

6 結(jié) 論

1）設(shè)計(jì)了一種基于Jetson Nano處理器的大蒜鱗芽朝向調(diào)整裝置，以Jetson Nano處理器對(duì)深度學(xué)習(xí)模型進(jìn)行推理，實(shí)現(xiàn)大蒜檢測(cè)和鱗芽朝向識(shí)別。根據(jù)大蒜鱗芽朝向控制調(diào)整機(jī)構(gòu)對(duì)大蒜鱗芽朝向進(jìn)行調(diào)整，使大蒜最終均能以鱗芽朝上的姿態(tài)落入插播裝置，解決了難以采用純機(jī)械機(jī)構(gòu)對(duì)雜交大蒜鱗芽朝向調(diào)整的問(wèn)題。

2）大蒜檢測(cè)和鱗芽朝向識(shí)別采用雙神經(jīng)網(wǎng)絡(luò)架構(gòu)，大蒜檢測(cè)神經(jīng)網(wǎng)絡(luò)實(shí)現(xiàn)對(duì)大蒜是否被喂入調(diào)整機(jī)構(gòu)進(jìn)行實(shí)時(shí)檢測(cè)，當(dāng)檢測(cè)到大蒜時(shí)鱗芽朝向識(shí)別神經(jīng)網(wǎng)絡(luò)對(duì)大蒜鱗芽朝向進(jìn)行實(shí)時(shí)判斷，利用判斷結(jié)果控制調(diào)整機(jī)構(gòu)對(duì)大蒜鱗芽朝向進(jìn)行調(diào)整。

3）通過(guò)離散元仿真軟件EDEM對(duì)調(diào)整機(jī)構(gòu)的關(guān)鍵參數(shù)進(jìn)行優(yōu)化，并通過(guò)臺(tái)架試驗(yàn)進(jìn)行了試驗(yàn)驗(yàn)證。試驗(yàn)結(jié)果，大蒜鱗芽朝向調(diào)整成功率為96.25%，模型推理時(shí)間0.045 s，平均每粒大蒜調(diào)整時(shí)間為0.785 s，大蒜理論播種效率可達(dá)到0.099～0.132 hm2/h。

[1] 王僑，陳兵旗，朱德利，等. 基于機(jī)器視覺(jué)的定向播種用玉米種粒精選裝置研究[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào)，2017，48(2)：27-37.

Wang Qiao, Chen Bingqi, Zhu Deli, et al. Vision-based selection machine of corn seed used for directional seeding[J]. Transactions of the Chinese Society for Agricultural Machinery, 2017, 48(2): 27-37. (in Chinese with English abstract)

[2] 金誠(chéng)謙，袁文勝，吳崇友，等. 大蒜播種時(shí)鱗芽朝向?qū)Υ笏馍L(zhǎng)發(fā)育影響的試驗(yàn)研究[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2008，24(4)：155-158.

Jin Chengqian, Yuan Wensheng, Wu Chongyou, et al. Experimental study on effects of the bulbil direction on garlic growth[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2008, 24(4): 155-158. (in Chinese with English abstract)

[3] 劉靜. 不同大蒜品種及鱗芽播種朝向?qū)ιL(zhǎng)特性與品質(zhì)的影響[D]. 泰安：山東農(nóng)業(yè)大學(xué)，2018.

Liu Jing. Effects of Different Garlic Varieties and Bu1bil Seeding Directions on Growth Characteristic and Quality[D]. Tai’an: Shandong Agricultural University, 2018. (in Chinese with English abstract)

[4] 崔榮江，黃嘉寶，張振河，等. 大蒜機(jī)械化播種技術(shù)研究現(xiàn)狀[J]. 農(nóng)業(yè)裝備與車輛工程，2018(6)：54-56.

Cui Rongjiang, Huang Jiabao, Zhang Zhenhe，et al. Research status of garlic mechanized sowing technology[J]. Agricultural Equipment & Vehicle Engineering, 2018(6): 54-56. (in Chinese with English abstract)

[5] 侯加林，黃圣海，牛子孺，等. 雙鴨嘴式大蒜正頭裝置調(diào)頭機(jī)理分析與試驗(yàn)[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào)，2018，49(11)：87-96.

Hou Jialin, Huang Shenghai, Niu Ziru, et al. Mechanism analysis and test of adjusting garlics upwards using two duckbill devices[J]. Transactions of the Chinese Society for Agricultural Machinery, 2018, 49(11): 87-96. (in Chinese with English abstract)

[6] 耿愛(ài)軍，栗曉宇，侯加林，等. 自動(dòng)定向大蒜播種機(jī)的設(shè)計(jì)與試驗(yàn)[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2018，34(11)：17-25.

Geng Aijun, Li Xiaoyu, Hou Jialin, et al. Design and experiment of automatic directing garlic planter[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(11): 17-25. (in Chinese with English abstract)

[7] 薦世春，劉云東. 大蒜播種機(jī)蒜瓣自動(dòng)定向控制裝置的試驗(yàn)研究[J]. 農(nóng)業(yè)裝備與車輛工程，2009(10)：28-29，37.

Jian Shichun, Liu Yundong. Experimental research on the garlic clove automatic orientation control device of garlic planting machine[J]. Transactions of the Agricultural Equipment and Vehicle Engineering, 2009(10): 28-29, 37. (in Chinese with English abstract)

[8] 薦世春，趙峰，李青，等. 旋轉(zhuǎn)式蒜瓣單粒定向取種器的研究設(shè)計(jì)[J]. 農(nóng)業(yè)裝備與車輛工程，2009(2)：18-20.

Jian Shichun, Zhao Feng, Li Qing, et al. Research and design on rotary garlic single-clove directional seed metering device[J]. Transactions of the Agricultural Equipment and Vehicle Engineering, 2009(2): 18-20. (in Chinese with English abstract)

[9] Liu H, Chahl J S. A multispectral machine vision system for invertebrate detection on green leaves[J]. Computers & Electronics in Agriculture, 2018, 150: 279-288.

[10] He Y, Wang H, Zhu S, et al. Method for grade identification of tobacco based on machine vision[J]. Transactions of the ASABE, 2018, 61(5): 1487-1495.

[11] 謝忠紅，徐煥良，黃秋桂，等. 基于高光譜圖像和深度學(xué)習(xí)的菠菜新鮮度檢測(cè)[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2019，35(13)：277-284.

Xie Zhonghong, Xu Huanliang, Huang Qiugui, et al. Spinach freshness detection based on hyperspectral image and deep learning method[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(13): 277-284. (in Chinese with English abstract)

[12] 孫鈺，周焱，袁明帥，等. 基于深度學(xué)習(xí)的森林蟲害無(wú)人機(jī)實(shí)時(shí)監(jiān)測(cè)方法[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2018，34(21)：74-81.

Sun Yu, Zhou Yan, Yuan Mingshuai, et al. UAV real-time monitoring for forest pest based on deep learning[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(21): 74-81. (in Chinese with English abstract)

[13] 苗榮慧，楊華，武錦龍，等. 基于圖像分塊及重構(gòu)的菠菜重疊葉片與雜草識(shí)別[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2020，36(4)：178-184.

Miao Ronghui, Yang Hua, Wu Jinlong, et al. Weed identification of overlapping spinach leaves based on imagesub-block and reconstruction[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(4): 178-184. (in Chinese with English abstract)

[14] 孫俊，譚文軍，武小紅，等. 多通道深度可分離卷積模型實(shí)時(shí)識(shí)別復(fù)雜背景下甜菜與雜草[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2019，35(12)：184-190.

Sun Jun, Tan Wenjun, Wu Xiaohong, et al. Real-time recognition of sugar beet and weeds in complex backgrounds using multi-channel depth-wise separable convolution model[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(12): 184-190. (in Chinese with English abstract)

[15] Tetila E C, Machado B B, Menezes G V, et al. A deep-learning approach for automatic counting of soybean insect pests[J]. IEEE Geoence and Remote Sensing Letters, 2019, PP(99): 1-5.

[16] 張博，張苗輝，陳運(yùn)忠. 基于空間金字塔池化和深度卷積神經(jīng)網(wǎng)絡(luò)的作物害蟲識(shí)別[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2019，35(19)：209-215.

Zhang Bo, Zhang Miaohui, Chen Yunzhong. Crop pest identification based on spatial pyramid pooling and deep convolution neural network[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(19): 209-215. (in Chinese with English abstract)

[17] Thenmozhi K, Reddy U S. Crop pest classification based on deep convolutional neural network and transfer learning[J]. Computers and Electronics in Agriculture, 2019, 164: 104906.

[18] 趙德安，吳任迪，劉曉洋，等. 基于YOLO深度卷積神經(jīng)網(wǎng)絡(luò)的復(fù)雜背景下機(jī)器人采摘蘋果定位[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2019，35(3)：164-173.

Zhao Dean, Wu Rendi, Liu Xiaoyang, et al. Apple positioning based on YOLO deep convolutional neural network for picking robot in complex background[J] Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(3): 164-173. (in Chinese with English abstract)

[19] 楊清明，李娟玲，何瑞銀. 基于圖像處理的大蒜蒜瓣朝向識(shí)別[J] . 浙江農(nóng)業(yè)學(xué)報(bào)，2010，22(1)：119-123.

Yang Qingming, Li Juanling, He Ruiyin. Direction identification of garlic seeds based on image processing[J]. Acta Agriculturae Zhejiangensis, 2010, 22(1): 119-123. (in Chinese with English abstract)

[20] 方春，孫福振，任崇廣. 基于深度學(xué)習(xí)的大蒜鱗芽朝向識(shí)別研究[J]. 計(jì)算機(jī)應(yīng)用研究，2019，36(2)：598-600，610.

Fang Chun, Sun Fuzhen, Ren Chongguang. Identifying bulbil direction of garlic based on deep learning[J]. Application Research of Computers, 2019, 36(2): 598-600, 610. (in Chinese with English abstract)

[21] 吳獻(xiàn)，胡偉. 基于觀測(cè)窗的大蒜鱗芽朝向識(shí)別研究[J]. 測(cè)控技術(shù)，2016，35(7)：35-39.

Wu Xian, Hu Wei. Research on garlic clove orientation recognition based on observation window[J]. Measurement & Control Technology, 2016, 35(7): 35-39. (in Chinese with English abstract)

[22] 趙麗清，馬志勇. 大蒜播種機(jī)裝盤系統(tǒng)蒜瓣定向識(shí)別算法的研究[J]. 農(nóng)機(jī)化研究，2013，35(6)：163-166.

Zhao Liqing, Ma Zhiyong. The study of garlic machine installation system of the directional recognition algorithm[J]. Journal of Agricultural Mechanization Research, 2013, 35(6): 163-166.(in Chinese with English abstract)

[23] 侯加林，田林，李天華，等. 基于雙側(cè)圖像識(shí)別的大蒜正芽及排種試驗(yàn)臺(tái)設(shè)計(jì)與試驗(yàn)[J] . 農(nóng)業(yè)工程學(xué)報(bào)，2020，36(1)：50-58.

Hou Jialin, Tian Lin, Li Tianhua, et al. Design and experiment of test bench for garlic bulbil adjustment and seeding based on bilateral image identification[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(1): 50-58. (in Chinese with English abstract)

[24] Li Y H, Wu Y Q, Li T H, et al. Design and experiment of adjustment device based on machine vision for garlic clove direction[J]. Computers and Electronics in Agriculture, 2020, 174: 105513.

[25] 白士寶，滕光輝，杜曉冬，等. 基于LabVIEW平臺(tái)的蛋雞舍環(huán)境舒適度實(shí)時(shí)監(jiān)測(cè)系統(tǒng)設(shè)計(jì)與實(shí)現(xiàn)[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2017，33(15)：237-244.

Bai Shibao, Teng Guanghui, Du Xiaodong, et al. Design and implementation on real-time monitoring system of laying hens environmental comfort based on LabVIEW[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(15): 237-244. (in Chinese with English abstract)

[26] 許順，佟金，李默. 基于LabVIEW的蔬菜切碎機(jī)性能測(cè)試及工作參數(shù)優(yōu)化[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2017，33(3)：250-256.

Xu Shun, Tong Jin, Li Mo. Performance testing of vegetable chopping machine based on LabVIEW and operation parameter optimization[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(3): 250-256. (in Chinese with English abstract)

[27] Debauche O, Sad M, Mahmoudi S A, et al. Edge computing and artificial intelligence for real-time poultry monitoring[J]. Procedia Computer Science, 2020, 175: 534-541.

[28] Carvajal O, Melin P, Miramontes I, et al. Optimal design of a general type-2 fuzzy classifier for the pulse level and its hardware implementation[J]. Engineering Applications of Artificial Intelligence, 2021, 97(4): 104069.

[29] 李玉華，張智龍，李天華，等. 輪勺式大蒜單粒取種裝置設(shè)計(jì)與試驗(yàn)[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào)，2020，51(3)：61-68.

Li Yuhua, Zhang Zhilong, Li Tianhua, et al Design and experiment of wheel-spoon type garlic precision seed-picking device[J]. Transactions of the Chinese Society for Agricultural Machinery, 2020, 51(3): 61-68. (in Chinese with English abstract)

[30] 文恩楊，吳彥強(qiáng)，李天華，等. 牽引式大蒜播種機(jī)的設(shè)計(jì)[J].農(nóng)機(jī)化研究，2020，42(1)：96-100.

Wen Enyang, Wu Yanqiang, Li Tianhua, et al. Design of traction garlic sowing machine[J]. Journal of Agricultural Mechanization Research, 2020, 42(1): 96-100. (in Chinese with English abstract)

Design and experiments of garlic bulbil orientation adjustment device using Jetson Nano processor

Li Yuhua1,2,3, Liu Quancheng1, Li Tianhua1,2,3, Wu Yanqiang1,2,3, Niu Ziru1,2,3, Hou Jialin1,2,3※

(1.,,271018,;2,271018,; 3,271018,)

Garlic cultivationis highly demanding for a single seedto maintain upright-directional sowing with the roots downwardvertically.However, current adjustment devicesfor the direction of garlic clovescannot besuitable for hybrid garlic varieties. In this study, an intelligent adjustment device was designedfor the direction of garlic cloveusing edge computing. The deviceconsistedof a feeding channelof garlic clove, a camera, a camera bracket, clove direction adjustment mechanism, turning servo and brackets. The adjustment mechanism of clove directionwas composed of flip frame, reset spring, hopper opening and closing servo, Y-type hopper I and II. A dual convolution neural network (CNN) structurewasadopted in thecontrol system,wherea custom deep learningCNNwas for garlic monitoringin real time, and a ResNet-18 network was for thedeterminationof garlic clove orientation. In monitoring, the garlic clove was distinguished from the background of images,therebydetermining whether the clove was fed to the adjustment. Areal-timedetection of orientation was to keep the pointy endof garlic clove facing upward,while the blunt end down into the soil. Asuitable control strategy was providedtopromptlyadjust the direction of garlic clove. Higher identification accuracyand real-time performance were achieved in two different networksfor separate detection and orientation of garlic clove.The specific procedure of orientation adjustment was as follows. Animage processing was performed to determine whether the garlic clove entered the Y-shaped hopper from the feeding channel. Once the garlic clove was detected to be in the hopper, an image was real-time captured by the camera. The captured image was processed immediately through the deep learning network of detection and orientation. When the scales (blunt end)of garlic cloveswere facingupward, the opening and closing servos of ahopper rotated at a certain angle to open the lower end of Y-type hopper I and II. As such, the garlic clove fell directly into the inserting with the scales facing upward. If the scales of garlic cloveswere facing downward, the turning servosand adjusting mechanismindividually rotated 180°, to accurately tailor the orientation of scaleswhen the garlic clove was slidingdown the guide slot of Y-type hopper I. Both theoretical and empirical data demonstrated that the structural parameters of the adjusting mechanismgreatly dominatedthe success rate of the adjusted scale bud. A discrete element method (DEM)was performed onacommercial software EDEM to simulate the working effect of the adjusting mechanism. An orthogonal test was also utilized to optimize the keyparameters of adjusting mechanism. An optimal combination of parameters was obtained,where the inclination angle of the garlic clove was 80°in the feeding channel,the radius of the Y-shaped hopper was 12.49 mm, and the height difference of the hopper was 20 mm.Finallya bench testwascarried out to verify the direction adjustment of garlic cloves.In scale bud,the success rate of identification was 97.25%, and the success rate of adjustment was 96.25%, where the success rate of adjustment was slightly lower than thatofrecognition.The reasonwas that the correctly identified garlic turnedover unexpectedly when falling, due to the irregular center of gravity inasingle seed. The mean inference timeof the model was 0.045s, indicating a small proportion of adjustment timefor the scale bud. The averageadjustment time was 0.785s,where the mean value was0.59s when the garlic cloves were facing up and0.98s when facing down. There was a relatively large difference inthe adjustment time when the garlic buds were faced up and down. This difference came into being because therewas inconsistent movement stroke of the adjustment mechanism in two cases, particularly where the rotation speed of the drive motor was the samewhen the scale buds were facingdown.Consequently, the adjustment timeof scale buds facing upwas shorterthan that of the roots downwardvertically in garlic planting.

machine vision; deep learning; edge computing; Jetson Nano processor; garlic; bulbil orientation

2020-12-13

2021-02-01

國(guó)家特色蔬菜產(chǎn)業(yè)技術(shù)體系項(xiàng)目(CARS-24-D-01)；山東省農(nóng)業(yè)重大應(yīng)用技術(shù)創(chuàng)新項(xiàng)目（SD2019NJ004）；山東省現(xiàn)代農(nóng)業(yè)產(chǎn)業(yè)技術(shù)體系蔬菜產(chǎn)業(yè)創(chuàng)新團(tuán)隊(duì)項(xiàng)目（SDAIT-05-11）

李玉華，講師，博士生，研究方向?yàn)橹悄苻r(nóng)業(yè)裝備研究。Email：liyuhua@sdau.edu.cn

侯加林，教授，博士生導(dǎo)師，研究方向?yàn)橹悄苻r(nóng)業(yè)裝備研究。Email：jlhou@sdau.edu.cn

10.11975/j.issn.1002-6819.2021.07.005

S223.2

A

1002-6819(2021)-07-0035-08

李玉華，劉全程，李天華，等. 基于Jetson Nano處理器的大蒜鱗芽朝向調(diào)整裝置設(shè)計(jì)與試驗(yàn)[J]. 農(nóng)業(yè)工程學(xué)報(bào)，2021，37(7)：35-42. doi：10.11975/j.issn.1002-6819.2021.07.005 http://www.tcsae.org

Li Yuhua, Liu Quancheng, Li Tianhua, et al. Design and experiments of garlic bulbil orientation adjustment device using Jetson Nano processor[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(7): 35-42. (in Chinese with English abstract) doi：10.11975/j.issn.1002-6819.2021.07.005 http://www.tcsae.org