謝文涌,柴琴琴,甘勇輝,陳舒迪,張 勛,王 武
基于多特征提取和Stacking集成學(xué)習(xí)的金線蓮品系分類
謝文涌1,2,柴琴琴1,2※,甘勇輝3,陳舒迪1,2,張勛4,王武1,2
(1. 福州大學(xué)電氣工程與自動(dòng)化學(xué)院,福州 350108;2. 福建省醫(yī)療器械和醫(yī)藥技術(shù)重點(diǎn)實(shí)驗(yàn)室,福州 350108;3.漳州職業(yè)技術(shù)學(xué)院食品工程學(xué)院,漳州 363000;4. 福建中醫(yī)藥大學(xué)藥學(xué)院,福州 350122)
針對(duì)傳統(tǒng)中藥鑒定、分子鑒定、生物技術(shù)鑒定及光譜檢測(cè)技術(shù)的主觀性強(qiáng)、耗時(shí)、操作復(fù)雜等不足,以及金線蓮整個(gè)葉片形態(tài)區(qū)分度小、單一分類器鑒別精度不高的問(wèn)題,該研究提出了基于機(jī)器視覺的葉片子區(qū)間多特征提取方法和基于多模型融合的Stacking集成學(xué)習(xí)算法實(shí)現(xiàn)金線蓮的品系分類。試驗(yàn)采集6個(gè)品系的金線蓮葉片圖像數(shù)據(jù),進(jìn)行圖像預(yù)處理后提取葉片子區(qū)間內(nèi)紋理、顏色共114個(gè)特征,基于這些特征,構(gòu)建堆疊式兩階段集成學(xué)習(xí)框架,以邏輯回歸、K最近鄰、隨機(jī)森林和梯度提升決策樹(Gradient Boosting Decision Tree,GBDT)作為基分類器,GBDT作為元分類器進(jìn)行學(xué)習(xí)。試驗(yàn)結(jié)果表明,Stacking集成學(xué)習(xí)模型的整體識(shí)別綜合評(píng)價(jià)指標(biāo)值達(dá)93.91%,分類正確率達(dá)94.49%,分別比邏輯回歸、K最近鄰、隨機(jī)森林和GBDT這4個(gè)單一分類模型高出4.40、11.87、11.01、12.94個(gè)百分點(diǎn)和5.36、11.34、6.93、12.13個(gè)百分點(diǎn)。因此,該研究能夠有效識(shí)別金線蓮品系,為形狀大小相似、形狀特征難以利用的植物葉片識(shí)別提供參考。
機(jī)器視覺;模型;金線蓮;子區(qū)間分割;特征提?。籗tacking集成學(xué)習(xí);植物葉片
金線蓮,又名金絲線、金耳環(huán)、烏人參、金錢草等,是中國(guó)的二級(jí)保護(hù)植物[1],被譽(yù)為“藥中之王”,能夠用于預(yù)防和治療高血壓,糖尿病,高脂血癥,肝炎和腫瘤等疾病[2-3]。由于野生金線蓮的自然繁殖率低、生長(zhǎng)條件受限制等原因?qū)е聰?shù)量有限,市面上出售的金線蓮大多為人工培育品種。目前,金線蓮品系繁多,不同品系的金線蓮?fù)庥^相似但藥效差異大,市場(chǎng)上以次充好的現(xiàn)象層出不窮,因此,如何準(zhǔn)確有效識(shí)別金線蓮的品系對(duì)保障藥方藥效、維護(hù)消費(fèi)者利益具有重要意義。
金線蓮的品質(zhì)鑒定通常依賴于化學(xué)分析方法,主要包括顯微鑒定法、高效液相色譜法、DNA分子鑒定法和近紅外光譜檢測(cè)技術(shù)等,然而這些方法需要在專業(yè)人士指導(dǎo)下進(jìn)行,存在主觀性強(qiáng),過(guò)程耗時(shí)且精度低,操作過(guò)程復(fù)雜且費(fèi)用昂貴等缺陷[4-5]。為了克服以上鑒別方法的缺陷,學(xué)者們將機(jī)器視覺技術(shù)引入到中藥材的鑒別中[6-7],主要集中在通過(guò)中藥材葉片的識(shí)別來(lái)判定藥材質(zhì)量,但尚未見有關(guān)基于機(jī)器視覺技術(shù)對(duì)金線蓮進(jìn)行品系分類的研究報(bào)道。在葉片的特征提取方面,通常使用形狀特征[8]、紋理特征[9]和顏色特征[10]來(lái)作為葉片的識(shí)別特征,為了更加充分表達(dá)葉片信息,進(jìn)一步提高葉片識(shí)別精度,特征融合在葉片識(shí)別中得到廣泛應(yīng)用[11]。然而上述研究主要是針對(duì)不同類別的植物葉片分類進(jìn)行,對(duì)于同種類別不同品系的植物葉片分類研究鮮有報(bào)道。
基于提取的特征,構(gòu)造合適的分類器是通過(guò)葉片識(shí)別解決品系鑒別問(wèn)題需研究的另一個(gè)重點(diǎn),常用于葉片識(shí)別領(lǐng)域的機(jī)器學(xué)習(xí)方法有K最近鄰(K Nearest Neighbor,KNN)[12]、支持向量機(jī)[13]、邏輯回歸(Logistic Regression,LR)[14]、極限學(xué)習(xí)機(jī)[9]等。然而這些識(shí)別方法均要求測(cè)試數(shù)據(jù)集與訓(xùn)練數(shù)據(jù)集的樣本概率分布一致,要在眾多假設(shè)函數(shù)構(gòu)成的空間中確定一個(gè)與實(shí)際情況最相符合的特征分布函數(shù)作為分類器。但是由于金線蓮種植基地分布廣泛、復(fù)雜的生長(zhǎng)環(huán)境會(huì)造成葉片形態(tài)異常,如存在如脈紋、葉形異常等情況,實(shí)際測(cè)試集數(shù)據(jù)分布存在不確定性,找到一個(gè)與實(shí)際情況最相符合的分布函數(shù)十分困難。因此,單一分類器往往會(huì)存在泛化能力不佳的問(wèn)題。為克服由于樣本不確定性帶來(lái)的分布函數(shù)難以精確估計(jì)的問(wèn)題,學(xué)者們提出了集成學(xué)習(xí)方法,通過(guò)將多個(gè)弱分類器集成為強(qiáng)分類器完成高精度的分類任務(wù)。目前,集成學(xué)習(xí)已成為機(jī)器學(xué)習(xí)的熱門研究方向之一[15],常用的集成學(xué)習(xí)方法有并行化集成Bagging[16],序列化集成的Boosting[17]和多層分類器組合的Stacking[18]。不同于Bagging和Boosting集成方式,Stacking使用高級(jí)別的元分類器來(lái)綜合低級(jí)別的基分類器的輸出特征以此來(lái)增強(qiáng)泛化能力,得到更高的預(yù)測(cè)精度,降低模型過(guò)擬合風(fēng)險(xiǎn)。Stacking集成學(xué)習(xí)在教育、醫(yī)學(xué)、社會(huì)科學(xué)等領(lǐng)域得到廣泛應(yīng)用[19-21],然而在農(nóng)業(yè)方面的應(yīng)用甚少[22]。
綜上所述,本文提出了基于多模型融合的Stacking集成學(xué)習(xí)算法實(shí)現(xiàn)不同品系金線蓮葉片的分類。首先針對(duì)利用整個(gè)葉片形狀難以區(qū)分不同品系的金線蓮的問(wèn)題,提出了基于子區(qū)間分割的顏色、紋理等特征提取和融合方法;其次以提取的特征為基礎(chǔ),在Stacking集成框架下設(shè)計(jì)LR、KNN、隨機(jī)森林(Random Forest,RF)[23]和梯度提升決策樹(Gradient Boosting Decision Tree,GBDT)[24]作為基學(xué)習(xí)器,并以GBDT作為元學(xué)習(xí)器輸出分類結(jié)果;最后,通過(guò)算法對(duì)比驗(yàn)證所提方法的有效性。
1.1.1 圖像獲取
金線蓮含有的多糖和黃酮成分是其藥效的有效成分,不同品系的金線蓮含量具有顯著差異,影響金線蓮的售價(jià)。目前,市場(chǎng)上主要售賣的金線蓮品系分為小圓葉、大圓葉、尖葉、臺(tái)灣金線蓮和雜交金線蓮。因此,本研究采集市場(chǎng)上廣泛流通的小圓葉、大圓葉、尖葉、臺(tái)灣金線蓮等主要品系以及雜交品系(紅霞、一株圓葉)作為研究對(duì)象,以此來(lái)驗(yàn)證品系分類方法的有效性。
試驗(yàn)所用全部金線蓮樣本均來(lái)源于福建葛園生物科技有限公司,并經(jīng)福建中醫(yī)藥大學(xué)教師專業(yè)認(rèn)證。采集人工套袋培養(yǎng)的6個(gè)品系金線蓮葉片,分別包含臺(tái)灣金線蓮58幅、紅霞64幅、小圓葉45幅、尖葉46幅、一株圓葉60幅、大圓葉44幅,總共317個(gè)樣本。為了使采集的樣本更具有代表性,將樣本平鋪于拍攝箱中的白色紙板上,箱內(nèi)燈光均勻照射在葉片,相機(jī)鏡頭與目標(biāo)樣本的距離為30 cm左右。采集設(shè)備為尼康單反數(shù)碼相機(jī),型號(hào)為NIKON D7100,圖像分辨率為6 000×4 000像素。采集的代表樣本圖片如圖1所示。
圖1 金線蓮不同品系葉片代表圖
1.1.2 圖像預(yù)處理與子區(qū)間劃分
為了避免除了葉片本身其他因素干擾,需對(duì)拍攝的原圖進(jìn)行預(yù)處理,流程如圖2所示。首先將圖像分辨率縮小至800×800像素,先使用均值濾波對(duì)彩色圖像降噪,其次將圖像灰度化,對(duì)灰度化后的圖像使用高斯濾波去除噪聲點(diǎn);然后用最大類間方差法(OTSU Method)[25]進(jìn)行自動(dòng)閾值分割得到二值化圖像,再用數(shù)學(xué)形態(tài)學(xué)方法中的閉運(yùn)算消除葉片存在的內(nèi)部孔洞,開運(yùn)算去除葉柄;最后,針對(duì)不同品系的金線蓮形狀大小極其相似,葉片形狀特征難以利用的問(wèn)題,提出對(duì)葉片進(jìn)行子區(qū)間分割來(lái)避免葉片形狀特征的影響,在二值圖像的基礎(chǔ)上繪制原圖的葉片輪廓,計(jì)算葉片的質(zhì)心,選擇以質(zhì)心為中心,邊長(zhǎng)為150個(gè)像素點(diǎn)的正方形區(qū)域作為葉片的子區(qū)間,從完整葉片中分割出相同部位和大小的子區(qū)間葉片。
圖2 圖像預(yù)處理與子區(qū)間劃分流程圖
1.2.1 紋理特征
本文使用基于統(tǒng)計(jì)的分析方法局部二進(jìn)制模式(Local Binary Pattern,LBP)[26]和灰度共生矩陣(Gray Level Co-occurrence Matrix,GLCM)[27]來(lái)提取紋理特征。為了更充分地表示紋理特征,使用時(shí)域和頻域相結(jié)合的Gabor濾波[28]進(jìn)行精細(xì)比例分析,并對(duì)紋理進(jìn)行多分辨率表示,將這三類特征融合作為金線蓮葉片識(shí)別的紋理特征。
1)LBP
隨著采樣點(diǎn)數(shù)的增加,二進(jìn)制模式的數(shù)量急劇增加,LBP算子計(jì)算量也隨之增大。為了解決這個(gè)問(wèn)題,提高統(tǒng)計(jì)性,采用等價(jià)的LBP模式進(jìn)行降維[29],這種方法既減少了二進(jìn)制模式數(shù)量又不丟失信息。對(duì)于有8個(gè)領(lǐng)域點(diǎn)的LBP算子,二進(jìn)制模式的數(shù)量由256種減少至59種。
2)GLCM
3)Gabor濾波
Gabor濾波器能捕獲到與圖像局部結(jié)構(gòu)信息相對(duì)應(yīng)的不同空間頻率、空間位置和方向的特征,被廣泛應(yīng)用于圖像中提取紋理特征。在空間域中,二維Gabor濾波器是由正弦平面波調(diào)制的高斯核函數(shù),具有表示正交方向的實(shí)部和虛部。Gabor濾波器的函數(shù)數(shù)學(xué)表達(dá)式如下
實(shí)數(shù)部分為
虛數(shù)部分為
在進(jìn)行特征提取之前,首先要將原圖像與Gabor函數(shù)進(jìn)行卷積運(yùn)算來(lái)生成與原圖像大小一致的目標(biāo)圖像。本文利用4個(gè)方向,6個(gè)尺度的Gabor濾波器生成24個(gè)卷積模板。將所提取的Gabor子帶(共4×6個(gè))的均值特征組合起來(lái)形成一個(gè)24維的特征向量。
1.2.2 顏色特征
顏色是圖像識(shí)別中最簡(jiǎn)單直接的特征,具有良好的魯棒性。由于顏色低階矩[30]中含有豐富的顏色分布信息,一階矩通過(guò)計(jì)算顏色的均值反映圖像的明暗信息,二階矩用來(lái)描述顏色的標(biāo)準(zhǔn)差反映圖像的顏色分布范圍,三階矩重點(diǎn)突出顏色的偏移性反映顏色的分布對(duì)稱性。所以,本文采用低階矩來(lái)提取色彩特征。通過(guò)測(cè)量整個(gè)圖像的顏色分布,針對(duì)顏色模型HSV和RGB進(jìn)行計(jì)算。然而HSV模型中分量與色彩無(wú)關(guān),所以提取顏色模型中的、、、、共5個(gè)顏色分量。最后,統(tǒng)計(jì)得到15個(gè)顏色特征。具體的計(jì)算公式如下:
對(duì)于Stacking集成學(xué)習(xí)而言,基分類器和元分類器的設(shè)計(jì)是關(guān)鍵之處。LR、KNN因?yàn)槔碚摮墒?、?jiǎn)單高效等優(yōu)點(diǎn),在很多領(lǐng)域有著很好的應(yīng)用效果[31-32]。RF、GBDT分別是基于Bagging和Boosting的思想,RF能夠高度并行化訓(xùn)練,大大提高了計(jì)算效率,而GBDT通過(guò)構(gòu)造弱分類器,使每個(gè)模型輸出結(jié)果與殘差之和盡可能的與預(yù)測(cè)值接近。從偏差和方差的角度分析,RF主要降低誤差的方差項(xiàng),GBDT既能降低偏差也能降低方差,這兩個(gè)算法相互組合通過(guò)不同的機(jī)制能確保結(jié)果的有效性[33-34]。所以,本文采用了經(jīng)典機(jī)器學(xué)習(xí)算法LR、KNN、RF、GBDT作為基分類器,元分類器使用泛化能力強(qiáng)的GBDT算法來(lái)糾正多個(gè)算法對(duì)訓(xùn)練集的偏置情況。
在Stacking集成學(xué)習(xí)訓(xùn)練過(guò)程中如果直接用第一層模型的訓(xùn)練數(shù)據(jù)當(dāng)做第二層元學(xué)習(xí)器的輸入,可能會(huì)有過(guò)擬合的風(fēng)險(xiǎn),因此需要對(duì)訓(xùn)練過(guò)程進(jìn)行重新設(shè)計(jì),為此針對(duì)金線蓮品系識(shí)別提出了如圖3所示的模型訓(xùn)練過(guò)程。主要步驟如下:
圖3 Stacking框架下的金線蓮品系識(shí)別
由于元分類器的訓(xùn)練集和測(cè)試集均未參與到各個(gè)基分類器的訓(xùn)練過(guò)程,所以有效防止了過(guò)擬合的發(fā)生,并且元分類器綜合了各個(gè)基分類器的輸出特征提升了分類準(zhǔn)確率。
3.2.1 Stacking集成學(xué)習(xí)模型鑒別結(jié)果分析
子區(qū)間分割之前,對(duì)葉片進(jìn)行形狀特征提取以此來(lái)驗(yàn)證金線蓮葉片形態(tài)區(qū)分度小這一特點(diǎn)。本次試驗(yàn)選擇Hu不變矩特征[35]和6個(gè)幾何特征[36](狹長(zhǎng)度、矩形度、圓形度、周長(zhǎng)長(zhǎng)寬比、周長(zhǎng)直徑比、偏心率)一起作為Stacking模型的輸入特征,試驗(yàn)結(jié)果表明葉片形狀特征的分類正確率僅為49.61%,故形狀特征不能作為葉片有效識(shí)別特征,從而采用子區(qū)間分割技術(shù)來(lái)提高識(shí)別率。
子區(qū)間分割后,采用混淆矩陣來(lái)評(píng)估Stacking集成學(xué)習(xí)的性能,具體結(jié)果如圖4所示。矩陣的每一行代表樣本的實(shí)際類別,每一列代表預(yù)測(cè)類別結(jié)果,矩陣對(duì)角線上的數(shù)字代表每個(gè)品系被正確識(shí)別的樣本數(shù)目??梢钥闯觯_(tái)灣金線蓮、紅霞、一株圓葉、大圓葉這4個(gè)不同品系全部被正確識(shí)別,然而小圓葉品系中有1個(gè)樣本被錯(cuò)誤識(shí)別成尖葉,尖葉中有6個(gè)樣本被誤識(shí)別成一株圓葉,說(shuō)明尖葉在紋理、顏色方面與人工自主培育品種一株圓葉有很大的相似性。
注:坐標(biāo)軸數(shù)字代表1.臺(tái)灣金線蓮 2.紅霞 3.小圓葉 4.尖葉 5.一株圓葉 6.大圓葉
此外,還使用了常見分類性能度量指標(biāo)來(lái)評(píng)價(jià)分類模型,這些指標(biāo)包括準(zhǔn)確率(Accuracy)、精確率(Precision)、召回率(Recall)、綜合評(píng)價(jià)指標(biāo)值(F1-Score),公式如下:
根據(jù)公式(13),Stacking模型下金線蓮品系識(shí)別正確率(Accuracy)達(dá)到94.49%,而LR、KNN、RF、GBDT的正確率分別為89.13%、83.15%、87.56%、82.36%,Stacking集成學(xué)習(xí)在一定程度上提升了分類準(zhǔn)確性。
3.2.2 不同模型分類結(jié)果對(duì)比
為了充分驗(yàn)證構(gòu)建的Stacking集成學(xué)習(xí)模型的有效性,根據(jù)式(14)~(16),針對(duì)每個(gè)不同品系將集成學(xué)習(xí)模型與單一分類模型分別在精確率、召回率和綜合評(píng)價(jià)指標(biāo)值方面進(jìn)行對(duì)比分析,并歸納出分類器的整體識(shí)別能力。對(duì)比具體結(jié)果如表1~4所示。
表1 不同分類模型精確率比較
表2 不同分類模型的召回率比較
表3 不同分類模型的綜合評(píng)價(jià)指標(biāo)比較
表4 不同分類模型整體平均識(shí)別能力比較
從表1~4可以看出,Stacking表現(xiàn)出良好的分類性能,在類別1、類別2、類別6品系識(shí)別中精確率、召回率、綜合評(píng)價(jià)指標(biāo)值均達(dá)100%。與單一分類模型相比,Stacking集成模型在金線蓮各個(gè)品系識(shí)別中均擁有最高的精確率,整體平均精確率達(dá)95.39%,比單一分類模型中最高的LR模型高7.46個(gè)百分點(diǎn)。在召回率的比較中,Stacking模型在類別3的召回率為94.44%低于LR的100%和KNN的96.66%,在類別4上為66.67%低于KNN模型的76.67%,但在整體上達(dá)到93.52%優(yōu)于其中任何一個(gè)分類器。值作為綜合評(píng)價(jià)指標(biāo)表示分類器的整體性能,在每個(gè)品系識(shí)別中Stacking集成模型的值均最高,且整體平均值達(dá)93.91%,分別比LR、KNN、RF、GBDT這4個(gè)單一模型高出4.40、11.87、11.01、12.94個(gè)百分點(diǎn),表現(xiàn)出該模型良好的綜合性能。
以上結(jié)果表明,Stacking集成模型綜合了其他4種弱分類器分類性能,整體上擁有比單一分類器更好的表現(xiàn)性能,能夠更加有效識(shí)別不同品系的金線蓮。這是因?yàn)閱我环诸惼髟谟?xùn)練過(guò)程中往往可能陷入局部最優(yōu)點(diǎn),而局部最優(yōu)對(duì)應(yīng)的模型泛化性能不佳,Stacking集成學(xué)習(xí)通過(guò)結(jié)合基分類器來(lái)有效減少陷入局部最優(yōu)點(diǎn)的風(fēng)險(xiǎn)。然而,Stacking集成學(xué)習(xí)在類別3與類別4的召回率表現(xiàn)上有所欠缺,這是因?yàn)镾tacking模型受到基分類器的精度與多樣性影響的原因,其基分類器之間的相關(guān)性大,在數(shù)據(jù)空間中獲得的分類假設(shè)函數(shù)相似,無(wú)法在最大程度上體現(xiàn)不同基分類器的優(yōu)勢(shì)。
本文提出了多模型融合的Stacking集成學(xué)習(xí)方法,對(duì)6種不同品系的金線蓮葉片進(jìn)行分類識(shí)別。首先對(duì)原始圖像進(jìn)行灰度化、濾波降噪、閾值分割、形態(tài)學(xué)處理等預(yù)處理,得到葉片二值圖像,在二值圖像的基礎(chǔ)上提出葉片子區(qū)間分割方法,得到不同葉片相同部位的子區(qū)間目標(biāo)圖像。然后使用局部二進(jìn)制模式(Local Binary Pattern,LBP)、灰度共生矩陣(Gray Level Co-occurrence Matrix,GLCM)、Gabor濾波等方法提取圖像紋理特征,使用顏色低階矩提取顏色特征,這些特征相互融合作為金線蓮的識(shí)別特征。為了充分學(xué)習(xí)數(shù)據(jù)特征,構(gòu)建了以邏輯回歸(Logistic Regression,LR)、K最近鄰(K Nearest Neighbor,KNN)、隨機(jī)森林(Random Forest,RF)和梯度提升決策樹(Gradient Boosting Decision Tree,GBDT)為基分類器的Stacking分類模型,以此來(lái)提升分類性能。通過(guò)試驗(yàn)研究比較,Stacking模型在分類準(zhǔn)確率、精確率、召回率和綜合評(píng)價(jià)指標(biāo)值等指標(biāo)上分別達(dá)94.49%、95.39%、93.52%和93.91%,優(yōu)于LR、KNN、RF和GBDT模型。在今后的研究中,將進(jìn)一步考慮Stacking配置選擇方法,減少計(jì)算時(shí)間的情況下提高學(xué)習(xí)效率,同時(shí)將進(jìn)一步擴(kuò)充圖像數(shù)據(jù)集,在更加復(fù)雜的背景下實(shí)現(xiàn)不同品系的植物識(shí)別。
[1] 邵清松,葉申怡,周愛存,等. 金線蓮種苗繁育及栽培模式研究現(xiàn)狀與展望[J]. 中國(guó)中藥雜志,2016,41(2):160-166. Shao Qingsong, Ye Shenyi, Zhou Aicun, et al. Current researches and prospects of seedling propagation and cultivation modes of Jinxianlian[J]. China Journal of Chinese Materia Medica, 2016, 41(2): 160-166. (in Chinese with English abstract)
[2] Ye S, Shao Q, Zhang A. Anoectochilus roxburghii: A review of its phytochemistry, pharmacology, and clinical applications[J]. Journal of Ethnopharmacology, 2017, 209: 184-202.
[3] Tseng C C, Shang H F, Wang L F, et al. Antitumor and immunostimulating effects of Anoectochilus formosanus Hayata[J]. Phytomedicine, 2006, 13(5): 366-370.
[4] 陳瑩,任麗,嚴(yán)桂杰,等. 不同來(lái)源金線蓮的HPLC指紋圖譜[J]. 沈陽(yáng)藥科大學(xué)學(xué)報(bào),2019,36(9):794-804. Chen Ying, Ren Li, Yan Guijie, et al. Fingerprint analysis of Anoectochilus roxburghii from different sources by HPLC[J]. Journal of Shenyang Pharmaceutical University, 2019, 36(9): 794-804. (in Chinese with English abstract)
[5] Lv T, Teng R, Shao Q, et al. DNA barcodes for the identification of Anoectochilus roxburghii and its adulterants[J]. Planta, 2015, 242(5): 1167-1174.
[6] Tao O, Lin Z, Zhang X B, et al. Research on identification model of chinese herbal medicine by texture feature parameter of transverse section image[J]. World Science and Technology-Modernization of Traditional Chinese Medicine, 2014 (12): 2558-2562.
[7] 朱黎輝,李曉寧,張瑩,等. 基于形狀特征及紋理特征的中藥材檢索方法[J]. 計(jì)算機(jī)工程與設(shè)計(jì),2014,35(11):3903-3907. Zhu Lihui, Li Xiaoning, Zhang Ying, et al. Image retrieval method for Chinese herbal medicine based on shape features and texture features[J]. Computer Engineering and Design, 2014, 35(11): 3903-3907. (in Chinese with English abstract)
[8] 宋彥,謝漢壘,寧井銘,等. 基于機(jī)器視覺形狀特征參數(shù)的祁門紅茶等級(jí)識(shí)別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(23):279-286. Song Yan, Xie Hanlei, Ning Jingming, et al. Grading Keemun black tea based on shape feature parameters of machine vision[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(23): 279-286. (in Chinese with English abstract)
[9] Turkoglu M, Hanbay D. Leaf-based plant species recognition based on improved local binary pattern and extreme learning machine[J]. Physica A: Statistical Mechanics and its Applications, 2019, 527: 121297.
[10] 張凱兵,章愛群,李春生. 基于HSV空間顏色直方圖的油菜葉片缺素診斷[J]. 農(nóng)業(yè)工程學(xué)報(bào),2016,32(19):179-187. Zhang Kaibing, Zhang Aiqun, Li Chunsheng. Nutrient deficiency diagnosis method for rape leaves using color histogram on HSV space[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(19): 179-187. (in Chinese with English abstract)
[11] 鄧向武,齊龍,馬旭,等. 基于多特征融合和深度置信網(wǎng)絡(luò)的稻田苗期雜草識(shí)別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(14):165-172. Deng Xiangwu, Qi Long, Ma Xu, et al. Recognition of weeds at seedling stage in paddy fields using multi-feature fusion and deep belief networks[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(14): 165-172. (in Chinese with English abstract)
[12] Sharma P, Aggarwal A, Gupta A, et al. Leaf identification using HOG, KNN, and neural networks[C]//International Conference on Innovative Computing and Communications. Springer, Singapore, 2019: 83-91.
[13] 馬娜,李艷文,徐苗. 基于改進(jìn)SVM 算法的植物葉片分類研究[J]. 山西農(nóng)業(yè)大學(xué)學(xué)報(bào):自然科學(xué)版,2018,38(11):39-44. Ma Na, Li Yanwen, Xu Miao. Plant leaf classification using improved SVM algorithm[J]. Journal of Shanxi Agricultural University: Natural Science Edition, 2018, 38(11): 39-44. (in Chinese with English abstract)
[14] 劉立波,程曉龍,戴建國(guó),等. 基于邏輯回歸算法的復(fù)雜背景棉田冠層圖像自適應(yīng)閾值分割[J]. 農(nóng)業(yè)工程學(xué)報(bào),2017,33(12):201-208. Liu Libo, Cheng Xiaolong, Dai Jianguo, et al. Adaptive threshold segmentation for cotton canopy image in complex background based on logistic regression algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(12): 201-208. (in Chinese with English abstract)
[15] Dong X, Yu Z, Cao W, et al. A survey on ensemble learning[J]. Frontiers of Computer Science, 2020, 14(2): 241-258.
[16] Andiojaya A, Demirhan H. A bagging algorithm for the imputation of missing values in time series[J]. Expert Systems with Application, 2019, 129(9): 10-26.
[17] Wang B, Pineau J. Online bagging and boosting for imbalanced data streams[J]. IEEE Transactions on Knowledge and Data Engineering, 2016, 28(12): 3353-3366.
[18] Hui Y, Mei X, Jiang G, et al. Milling tool wear state recognition by vibration signal using a stacked generalization ensemble model[J]. Shock and Vibration, 2019(3): 1-16.
[19] Elayidom S, Idikkula S M, Alexander J. A hybrid stacking ensemble framwork for employment predicyion problems[J]. Advances in Computational Research, 2011, 3(1): 25-30
[20] Dinakar K, Weinstein E, Lieberman H, et al. Stacked generalization learning to analyze teenage distress[C]// 2014 8th International AAAI Conference on Weblogs and Social Media (ICWSM 2014), Ann Arbor, Michigan, USA, 2014: 81-90
[21] Haddad B M, Yang S, Karam L J, et al. Multifeature, sparse-based approach for defects detection and classification in semiconductor units[J]. IEEE Transactions on Automation Science and Engineering, 2016, 15(1): 1-15.
[22] 袁培森,楊承林,宋玉紅,等. 基于Stacking集成學(xué)習(xí)的水稻表型組學(xué)實(shí)體分類研究[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2019, 50(11):144-152. Yuan Peisen, Yang Chenglin, Song Yuhong, et al. Classification of rice phenomics entities based on stacking ensemble learning[J]. Transactions of the Chinese Society for Agricultural Machinery, 2019, 50(11): 144-152. (in Chinese with English abstract)
[23] 王麗愛,周旭東,朱新開,等. 基于HJ-CCD數(shù)據(jù)和隨機(jī)森林算法的小麥葉面積指數(shù)反演[J]. 農(nóng)業(yè)工程學(xué)報(bào),2016,32(3):149-154. Wang Liai, Zhou Xudong, Zhu Xinkai, et al. Inverting wheat leaf area index based on HJ-CCD remote sensing data and random forest algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(3): 149-154. (in Chinese with English abstract)
[24] 張會(huì)清,牛錚. 基于線性判別分析和梯度提升決策樹的WLAN室內(nèi)定位算法[J]. 儀器儀表學(xué)報(bào),2018,39(12):136-143. Zhang Huiqing, Niu Zheng. WLAN indoor positioning algorithm based on linear discriminant analysis and gradient boosting decision tree[J]. Chinese Journal of Scientific Instrument. 2018, 39(12): 136-143. (in Chinese with English abstract)
[25] 王見,周勤,尹愛軍. 改進(jìn)Otsu算法與ELM融合的自然場(chǎng)景棉桃自適應(yīng)分割方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(1):173-180. Wang Jian, Zhou Qin, Yin Aijun. Self-adaptive segmentation method of cotton in natural scene by combining improved Otsu with ELM algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(1): 173-180. (in Chinese with English abstract)
[26] Naresh Y G, Nagendraswamy H S. Classification of medicinal plants: An approach using modified LBP with symbolic representation[J]. Neurocomputing, 2016, 173: 1789-1797.
[27] Wu Q, Gan Y, Lin B, et al. An active contour model based on fused texture features for image segmentation[J]. Neurocomputing, 2015, 151: 1133-1141.
[28] VijayaLakshmi B, Mohan V. Kernel-based PSO and FRVM: An automatic plant leaf type detection using texture, shape, and color features[J]. Computers and Electronics in Agriculture, 2016, 125: 99-112.
[29] Yang H, Yin J, Jiang M. Perceptual image hashing using latent low-rank representation and uniform LBP[J]. Applied Sciences, 2018, 8(2): 317.
[30] Das S, Rudrapal D. Analysis of color moment as a low level feature in improvement of content based image retrieval[C]// Proceedings of the Fourth International Conference on Signal and Image Processing 2012 (ICSIP 2012). Springer, India, 2013: 387-397.
[31] Rymarczyk T, Kozowski E, Kosowski G, et al. Logistic regression for machine learning in process tomography[J]. Sensors, 2019, 19(15): 3400.
[32] Lee T R, Wood W T, Phrampus B J. A machine learning (KNN) approach to predicting global seafloor total organic carbon[J]. Global Biogeochemical Cycles, 2019, 33(1): 37-46
[33] Gui L, Xia Y, Li H, et al. Prediction of NOX emission from coal - fired boiler based on RF - GBDT[C]// 2017 6th International Conference on Energy and Environmental Protection (ICEEP 2017). 2017: 344-350.
[34] 徐兵,劉瀟,汪子揚(yáng),等. 采用梯度提升決策樹的車輛換道融合決策模型[J]. 浙江大學(xué)學(xué)報(bào):工學(xué)版,2019,53(6):158-168. Xu Bing, Liu Xiao, Wang Ziyang, et al. Fusion decision model for vehicle lane change with gradient boosting decision tree[J]. Journal of Zhejiang University: Engineering Science, 2019, 53(6): 158-168. (in Chinese with English abstract)
[35] Wang X F, Huang D S, Du J X, et al. Classification of plant leaf images with complicated background[J]. Applied Mathematics & Computation, 2008, 205(2): 916-926.
[36] Ahmed F, Almamun H A, Bari A S, et al. Classification of crops and weeds from digital images: A support vector machine approach[J]. Crop Protection, 2012, 40: 98-108.
Strains classification ofusing multi-feature extraction and Stacking ensemble learning
Xie Wenyong1,2, Chai Qinqin1,2※, Gan Yonghui3, Chen Shudi1,2, Zhang Xun4, Wang Wu1,2
(1.,,350108,;2.,,350108,;3.,,363000,;4.,,350122,)
(A.) is a rare medicinal herb that mainly distributed in China. It is necessary to identify strains of A.for the guidance of clinical medication, due to different strains distinctly vary in medicinal values. However, similar leaf morphology has made difficult to discern different strains directly by naked eyes. In this study, a sub-interval segmentation method was proposed to identify the different strains of A., based on leaf identification methods. Firstly, 6 strains of A.were selected, including Taiwan, Hongxia, Xiaoyuanye, Jianye, Yizhu, Dayuanye. A total of 317 images with the resolution of 800×800 pixels were taken, while two filtering methods were used to remove noise. The maximum inner variance algorithm was used for automatic threshold segmentation, in order to obtain the binary image. In the binary image, the leaf contour was drawn, and the mass center of the leaf was calculated. The square area with 150 pixels centered on the mass center was selected as the sub-interval of the leaf, to obtain the target image with the same position and size. Secondly, a combination of texture and color features was applied for the target image, in which texture features were derived by local binary patterns (LBP), gray level co-occurrence matrix (GLCM) and gabor filters, whereas, the color feature was composed of the first, second and third moments. After that, 114 merged features were obtained. Thirdly, the stacking ensemble learning was proposed to improve the accuracy of traditional single classifier. The stacking framework consisted of a base classifier, and a meta-classifier. Logistic regression (LR), K nearest neighbor (KNN), random forest (RF), and gradient boosting decision tree (GBDT) were used as the base classifiers, whereas, GBDT was used as the meta-classifier for stacking. Finally, the cross-validation method different from conventional model was used to divide the data set. The original data was normalized and randomly segmented, where 60% for training and 40% for testing. The training data set was randomly divided into 5 training subsets, and then testing subset for training each base classifier. The prediction results of base classifiers were used as the input vectors of the GBDT. The final prediction result was output by GBDT. The experiment results showed that the average recognition accuracy of the stacking reached 94.49%, while that of LR, KNN, RF and GBDT was 89.13%, 83.15%, 87.56%, 82.36%, respectively. Moreover, the Precision, Recall, and1-Score of the stacking model for the identification of Taiwan, Hongxia, and Dayuanye were all 100%. The Recall performance of stacking model was better than any of the single classifiers for identification of the Xiaoyuanye, just slightly worse than that of the LR and KNN models. The1-Score of stacking model reached the maximum in each strain identification, showing the excellent overall performance of the model. Therefore, the proposed method can significantly improve the classification performances of A.with different strains. The findings can provide a promising application method to recognize leaves of different plants using shape features. A further research is still necessary to select proper configuration, in order to improve learning efficiency of stacking model.
computer vision; models;; sub-interval segmentation; feature extraction; Stacking ensemble learning; plant leaf
謝文涌,柴琴琴,甘勇輝,等. 基于多特征提取和Stacking集成學(xué)習(xí)的金線蓮品系分類[J]. 農(nóng)業(yè)工程學(xué)報(bào),2020,36(14):203-210.doi:10.11975/j.issn.1002-6819.2020.14.025 http://www.tcsae.org
Xie Wenyong, Chai Qinqin, Gan Yonghui, et al. Strains classification ofusing multi-feature extraction and Stacking ensemble learning[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(14): 203-210. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2020.14.025 http://www.tcsae.org
2020-04-03
2020-06-17
國(guó)家自然科學(xué)基金項(xiàng)目(61773124);福建省科技廳高校產(chǎn)學(xué)合作項(xiàng)目(No.2019Y4009);福建省食品藥品監(jiān)督管理局金線蓮質(zhì)量標(biāo)準(zhǔn)提升專項(xiàng)([3500]FJJF[DY]2018008)
謝文涌,主要從事圖像處理與機(jī)器學(xué)習(xí)研究。Email:1024396820@qq.com
柴琴琴,副教授,主要從事機(jī)器學(xué)習(xí)與模式識(shí)別研究。Email:qq.chai@fzu.edu.cn
10.11975/j.issn.1002-6819.2020.14.025
TP391
A
1002-6819(2020)-14-0203-08