王 震,褚桂坤,張宏建,劉雙喜,黃信誠(chéng),高發(fā)瑞,張春慶,王金星
?
基于無(wú)人機(jī)可見(jiàn)光圖像Haar-like特征的水稻病害白穂識(shí)別
王 震1,2,褚桂坤1,張宏建1,劉雙喜1,2,黃信誠(chéng)3,高發(fā)瑞3,張春慶4,王金星1,2※
(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. 山東省園藝機(jī)械與裝備重點(diǎn)實(shí)驗(yàn)室,泰安 271018;3. 濟(jì)寧市農(nóng)業(yè)科學(xué)研究院,濟(jì)寧 273013;4. 山東農(nóng)業(yè)大學(xué)農(nóng)學(xué)院,泰安 271018)
實(shí)現(xiàn)稻田精準(zhǔn)植保的關(guān)鍵是自然環(huán)境下病變區(qū)域的準(zhǔn)確識(shí)別。為實(shí)現(xiàn)大面積稻田中白穗的精確識(shí)別,該文提出一種小型多旋翼無(wú)人機(jī)水稻病害白穂識(shí)別系統(tǒng),該系統(tǒng)以無(wú)人機(jī)平臺(tái)作為圖像采集、處理和識(shí)別的基礎(chǔ),首先對(duì)白穗圖像提取Haar-like特征,其次以Adaboost 算法進(jìn)行白穗訓(xùn)練識(shí)別。以4類Haar-like特征及其組合構(gòu)建弱分類器,用采集的稻田白穗和背景共700個(gè)樣本點(diǎn)訓(xùn)練生成強(qiáng)分類器。所得強(qiáng)分類器對(duì)測(cè)試集中65幅圖像中的423個(gè)白穗樣本點(diǎn)進(jìn)行識(shí)別驗(yàn)證,結(jié)果表明:白穗識(shí)別率可達(dá)93.62%,誤識(shí)別率為5.44%,該方法可有效抑制一般的稻葉遮擋、稻穗黏連以及光照等復(fù)雜背景的影響,適合于自然環(huán)境下的稻田白穗現(xiàn)場(chǎng)識(shí)別。
無(wú)人機(jī);算法;病害;水稻白穗;Haar-like 特征
白穂是稻田中常見(jiàn)的一種影響稻米產(chǎn)量和品質(zhì)的病蟲(chóng)害特征,嚴(yán)重時(shí)可造成稻田大面積減產(chǎn),嚴(yán)重地區(qū)白穂率可達(dá)50%[1],在白穂形成早期對(duì)稻田進(jìn)行病蟲(chóng)害防治是預(yù)防白穂大面積發(fā)生的最佳時(shí)期。目前稻田白穂的識(shí)別還是以人工肉眼觀察為主,且白穂早期的呈現(xiàn)往往具有局簇性[2]。在大面積稻田種植中,人工觀測(cè)法很難準(zhǔn)確全面的識(shí)別到早期形成的白穂。小型多旋翼無(wú)人機(jī)是近幾年迅速發(fā)展起來(lái)的農(nóng)情觀測(cè)手段,尤其適用于田間環(huán)境不易于人工進(jìn)入的農(nóng)田區(qū)域。
白穂發(fā)生早期的識(shí)別,關(guān)鍵是對(duì)白穂和正常稻穗的準(zhǔn)確區(qū)分、識(shí)別以及白穂發(fā)生位置的定位。在農(nóng)田環(huán)境中,目標(biāo)識(shí)別一般基于顏色或者形態(tài)特征等算法進(jìn)行。目標(biāo)識(shí)別的顏色空間主要有RGB、HSI和La*b*;形態(tài)特征識(shí)別主要采用支持向量機(jī)(SVM)、神經(jīng)網(wǎng)絡(luò)算法、遺傳模糊神經(jīng)網(wǎng)絡(luò)算法等模式識(shí)別方法[3-7]。顏色模型方面,謝忠紅等[8]提出了一種基于改進(jìn)圓形隨機(jī) Hough 變換的快速類圓果實(shí)目標(biāo)檢測(cè)方法;蔡建榮等[9]應(yīng)用球體HSI顏色系統(tǒng)描述成熟西紅柿的顏色,利用Otsu算法自動(dòng)獲取分割閾值,提取目標(biāo)區(qū)域,分割效果顯著;詹文田等[10]利用顏色空間多個(gè)通道構(gòu)建不同的弱分類器,再通過(guò)樣本訓(xùn)練得到一個(gè)強(qiáng)分類器,對(duì)田間獼猴桃進(jìn)行識(shí)別,識(shí)別率高達(dá)96.7%。模式識(shí)別方面,宋懷波等[11]利用-means聚類算法將果樹(shù)圖像分為樹(shù)葉、枝條和果實(shí)3個(gè)類別,然后利用形態(tài)學(xué)方法對(duì)果實(shí)目標(biāo)進(jìn)行處理,得到目標(biāo)邊緣并進(jìn)行輪廓跟蹤,有效識(shí)別到遮擋的蘋(píng)果,平均定位誤差為4.28%;李昕等[12]提出的人工免疫網(wǎng)絡(luò)識(shí)別的多特征融合識(shí)別方法,利用偏好免疫算法進(jìn)行多特征有效融合,使油茶果的識(shí)別率達(dá)到了93.9%;趙源深等[13]提出了一種基于非顏色編碼的西紅柿識(shí)別算法,使成熟西紅柿的識(shí)別率達(dá)到了93.3%。在稻穗識(shí)別方面,劉占宇等[14]提出了一種基于學(xué)習(xí)矢量量化神經(jīng)網(wǎng)絡(luò)的水稻白穗和正常穗的高光譜識(shí)別方法,在稻穗離體識(shí)別試驗(yàn)中,識(shí)別精度可達(dá)100%;劉占宇等[15]還通過(guò)測(cè)定稻穗室內(nèi)高光譜反射率,對(duì)稻穗的健康狀態(tài)進(jìn)行了分類,分類精度也達(dá)到了理想效果,但該方法不適于稻田實(shí)時(shí)識(shí)別;白曉東[16]提出了一種基于稻穗顏色特征檢測(cè)、梯度直方圖檢測(cè)以及卷積神經(jīng)網(wǎng)絡(luò)提升的水稻抽穗期自動(dòng)檢測(cè)方法,該方法可在稻田中識(shí)別出新生稻穗,滿足農(nóng)氣觀測(cè)需要,對(duì)于病害稻穗未做相關(guān)研究。
綜上所述,從農(nóng)田目標(biāo)識(shí)別的研究來(lái)看,不管是基于顏色特征,還是形態(tài)特征的農(nóng)田目標(biāo)識(shí)別算法,都得到了較高的目標(biāo)識(shí)別率,但是各算法之間也存在著特征能力描述和識(shí)別速度之間互斥矛盾的不足之處[17-18]。從識(shí)別手段來(lái)看,多數(shù)目標(biāo)識(shí)別可做到無(wú)損識(shí)別,圖像采集多以地面隨機(jī)多點(diǎn)方式采集,圖像質(zhì)量較高卻難以覆蓋大面積農(nóng)田目標(biāo),高光譜采集雖可進(jìn)行大面積信息獲取,卻難以滿足高精度識(shí)別要求,存在識(shí)別覆蓋度與識(shí)別精度之間的矛盾。
為平衡農(nóng)田目標(biāo)識(shí)別過(guò)程中,特征描述能力與識(shí)別速度之間、識(shí)別覆蓋度與識(shí)別精度之間的矛盾,本文提出一種利用小型多旋翼無(wú)人機(jī)機(jī)載高分辨率相機(jī)采集稻田圖像,基于Haar-like特征和AdaBoost學(xué)習(xí)算法[19-20]的稻田白穂識(shí)別方法,并通過(guò)試驗(yàn)研究Haar-like特征描述能力與AdaBoost學(xué)習(xí)過(guò)程中訓(xùn)練次數(shù)對(duì)識(shí)別算法性能的影響,以期達(dá)到快速準(zhǔn)確識(shí)別稻田白穂的效果。迄今為止,尚未看到有關(guān)利用小型多旋翼無(wú)人機(jī)圖像進(jìn)行水稻白穂識(shí)別的研究報(bào)道,本文提出的基于小型多旋翼無(wú)人機(jī)圖像和Haar-like特征的稻田水稻白穂識(shí)別,將為稻田無(wú)人機(jī)精準(zhǔn)施藥提供參考。
采集山東省濟(jì)寧市陳莊農(nóng)林科技試驗(yàn)站“圣稻19”水稻齊穗期至成熟期圖像供試。2015年至2017年期間,每年8月-10月采集稻田白穂圖像建立試驗(yàn)樣本圖像庫(kù)。以多旋翼無(wú)人機(jī)SPREADING WINGS S900平臺(tái)(如圖1所示)為采集設(shè)備,設(shè)備具體參數(shù)見(jiàn)表1。
1.多旋翼無(wú)人機(jī) 2.相機(jī)云臺(tái) 3.工業(yè)CDD數(shù)字相機(jī)4.GPS和指南針
表1 試驗(yàn)設(shè)備參數(shù)
稻田圖像采用對(duì)無(wú)人機(jī)航拍視頻流進(jìn)行幀提取的方式進(jìn)行采集。首先利用小型多旋翼無(wú)人機(jī)沿預(yù)定航線進(jìn)行視頻拍攝,然后通過(guò)各項(xiàng)參數(shù)計(jì)算提取圖像的間隔幀數(shù),最后通過(guò)premiere軟件提取有效圖像。提取到的有效圖像按照時(shí)序拼接處理可還原出整個(gè)稻田圖像[21],進(jìn)而建立試驗(yàn)樣本庫(kù),進(jìn)行白穗識(shí)別,實(shí)現(xiàn)無(wú)人機(jī)大面積稻田白穗的識(shí)別。另外,多旋翼無(wú)人機(jī)自帶GPS定位導(dǎo)航系統(tǒng),如圖1所示,從而可獲得無(wú)人機(jī)航拍時(shí)的坐標(biāo),計(jì)算可得提取樣本的位置信息,從而定位白穗的具體位置。稻田圖像提取計(jì)算公式如下
式中為單幀圖像實(shí)際拍攝寬度,m;為相機(jī)盲區(qū)視角,(°);為相機(jī)視角,(°);為無(wú)人機(jī)航拍水平飛行速度,m/s;為無(wú)人機(jī)距作物冠層垂直高度,m;為幀提取時(shí)間,s;0為航拍視頻流速率,幀/s;為2張有效圖像之間幀圖像數(shù)。視頻流采集示意圖如圖2所示。
注:α為相機(jī)盲區(qū)視角,(°);β為相機(jī)視角,(°);v為無(wú)人機(jī)飛行速度,m·s-1。
采樣過(guò)程中,設(shè)定無(wú)人機(jī)航拍視頻流速率0=25幀/s,無(wú)人機(jī)飛行速度=1 m/s,航拍飛行高度=1.5 m,相機(jī)視角=54°,視頻格式為MPEG-4,為規(guī)避無(wú)人機(jī)旋翼下旋氣流影響拍攝效果,拍攝過(guò)程中利用云臺(tái)遙控器固定相機(jī)中軸線與地面垂直夾角為45°進(jìn)行拍攝,即+0.5=45°,故=18°。以上數(shù)據(jù)帶入式(1)~式(3)計(jì)算可知,單幀圖像實(shí)際拍攝寬度=3.64 m,幀提取時(shí)間=3.64 s。所以每隔3.64×25=91幀圖像提取一幀作為試驗(yàn)樣本圖像。圖3為視頻流拆解的幀圖像,圖3為按照91幀間隔規(guī)則提取的第2 184、2 275、2 366、2 457幀圖像;圖4為圖3中4幀提取圖像的拼接效果圖,圖像方向?yàn)闊o(wú)人機(jī)行進(jìn)方向。因無(wú)人機(jī)風(fēng)力和擾動(dòng)影響,拼接圖像存在一定的誤差,但對(duì)于單幅提取的幀圖像進(jìn)行算法處理沒(méi)有太大影響。由航拍飛行高度=1.5 m,相機(jī)視角=54°,可計(jì)算出相機(jī)視角至地面的中垂線距離為1.5 m/cos45°=2.12 m,進(jìn)而可計(jì)算相機(jī)視角幅寬為2.12 m×tan27°×2=2.16 m,再由無(wú)人機(jī)飛行速度=1 m/s可知,無(wú)人機(jī)每小時(shí)可進(jìn)行識(shí)別作業(yè)面積為2.12 m× 1 m/s×3 600 s=7 632 m2,發(fā)揮了無(wú)人機(jī)大面積作業(yè)的快速、高效優(yōu)勢(shì)。
圖3 視頻流拆解的幀提取圖像
圖4 圖3中4幀提取圖像的拼接圖像
利用多旋翼無(wú)人機(jī)進(jìn)行稻田白穗圖像采集,較傳統(tǒng)人工采集和田間固定設(shè)備采集而言,具有效率高、采集圖像連續(xù)、應(yīng)用面積大等特點(diǎn)。圖像包含位置信息,可對(duì)目標(biāo)進(jìn)行精確定位,是無(wú)人機(jī)采集圖像與傳統(tǒng)方式采集圖像的最大區(qū)別和優(yōu)勢(shì)。因用于白穗識(shí)別的圖像是由固定幀間隔提取而來(lái),且圖像中含有位置坐標(biāo)信息,使得本文提出的算法可有效的檢測(cè)和定位大田中的白穗。雖然利用Haar-like特征提取和AdaBoost學(xué)習(xí)算法進(jìn)行目標(biāo)識(shí)別已成熟應(yīng)用于較多領(lǐng)域,但基于位置信息和大田作業(yè)來(lái)說(shuō),本文提出的算法只適用于小型多旋翼無(wú)人機(jī)獲取的視頻流信息。無(wú)人機(jī)采集的稻田原始圖像需要進(jìn)行壓縮、切割、歸一化、背景分離、閾值分割、去除噪聲等預(yù)處理操作后用于Haar-like特征提取和AdaBoost學(xué)習(xí)算法,最終進(jìn)行白穗識(shí)別。
稻田白穗的發(fā)生由多種病害或者蟲(chóng)害引起,病蟲(chóng)害發(fā)生之初一般危害單株或少量幾株水稻,并具有傳播性。隨著寄主植株危害程度加深,養(yǎng)分枯竭,轉(zhuǎn)移至臨近植株進(jìn)行危害并繁衍,若未進(jìn)行防治可造成大面積白穗現(xiàn)象發(fā)生。圖5為白穗發(fā)生初期至后期的水稻表征圖,本文主要針對(duì)圖中5a、5b階段進(jìn)行樣本采樣研究,此階段白穗危害并沒(méi)有大面積發(fā)生,使用無(wú)人機(jī)進(jìn)行早期識(shí)別并防治,可預(yù)防白穗大面積發(fā)生。對(duì)于c階段,白穗已經(jīng)大面積發(fā)生的狀況,防治已失去時(shí)效性,防治效果不明顯。
圖5 白穗危害表征圖
本研究主要采集水稻齊穗期至成熟期發(fā)生的白穂圖像,利用多旋翼無(wú)人機(jī)SPREADING WINGS S900平臺(tái),在稻田中獲取自然條件下的水稻白穂圖像,無(wú)人機(jī)飛行高度在1~1.5 m之間,采集時(shí)間集中在每天的10:00~14:00。如圖6所示,為了訓(xùn)練和測(cè)試研究提出的水稻白穂識(shí)別算法,所選取的水稻白穂圖像除了考慮光照影響外,還需要包括葉片遮擋、稻穗粘連或者復(fù)雜背景等干擾因素。為了方便計(jì)算將圖像分辨率壓縮為640×480 像素,存儲(chǔ)為jpg格式。
圖6 圖像采集各種干擾條件
試驗(yàn)期間共采集350幅水稻齊穗期至成熟期的稻田現(xiàn)場(chǎng)圖像,將圖像中的285幅做為訓(xùn)練集,在訓(xùn)練集圖像中選取包括白穂各種形態(tài)、光照、遮擋、粘連以及背景等干擾因素在內(nèi)的700個(gè)樣本點(diǎn),以保證訓(xùn)練集的包容性;剩余65幅圖像作為測(cè)試集,從中選取800個(gè)樣本點(diǎn)做為測(cè)試集樣本。表2為訓(xùn)練集和測(cè)試集中包含的各種形態(tài)的樣本數(shù)量,圖7列出了部分的正負(fù)樣本圖像。
Haar-like特征是是計(jì)算機(jī)視覺(jué)領(lǐng)域一種常用的特征描述算子,是用于物體識(shí)別的一種數(shù)字圖像特征。Haar-like特征值反映了圖像的灰度變化情況,通過(guò)改變特征模板的大小和位置,可在圖像子窗口中窮舉出大量的特征,達(dá)到辨別目標(biāo)物的目的?;贖aar-like特征識(shí)別一副待檢圖像中是否含有白穂,首先要對(duì)圖像進(jìn)行Haar-like特征提取,然后對(duì)白穂特征集和背景特征集進(jìn)行訓(xùn)練建立分類器。考慮到特征集的豐富性和識(shí)別速度,本文設(shè)計(jì)了4類Haar-like特征矩形:A類為邊緣特征、B類為線性特征、C類為中心特征、D類為一種擴(kuò)展Harr-like特征,如圖8所示。
表2 樣本集數(shù)量
圖7 訓(xùn)練樣本集舉例
圖8 Haar-like特征矩形示意圖
圖9為線性特征和擴(kuò)展特征在白穂圖像中的部分匹配說(shuō)明。由圖9a可知,線性特征對(duì)于圖像中垂直和傾斜的白穂局部灰度區(qū)域范圍變化的描述較為精確,白穂兩側(cè)邊緣鄰域部分的灰度比中間部分的灰度更深,線性特征對(duì)于單穗白穂的特征提取十分有效;圖9b所示,對(duì)于多穗白穗雜亂交差、粘連等情形,可利用擴(kuò)展特征進(jìn)行有效描述。
圖9 兩組特征在圖像上的部分匹配
在進(jìn)行Haar-like 特征提取時(shí),將樣本圖像歸一化到24像素×24像素的同一個(gè)尺度上。由Haar-like特征個(gè)數(shù)計(jì)算公式和其中×為圖像大小,×為矩形特征大小,表示矩形特征在水平和垂直方向的能放大的最大比例系數(shù),=+,可知A類特征矩形每組生成51 664個(gè)特征,B類特征矩形每組生成28 056個(gè)特征,C類特征矩形生成9 985個(gè)特征,D類特征矩形每組生成37 600個(gè)特征。
獲取矩形特征之后,對(duì)特征值進(jìn)行計(jì)算,為提高計(jì)算速度,強(qiáng)化算法實(shí)時(shí)性,引入積分圖算法實(shí)現(xiàn)Haar-like 特征的快速提取[22]。
提取的Haar-like 特征輸入到Adaboost進(jìn)行訓(xùn)練學(xué)習(xí),Adaboost是一種迭代算法,其通過(guò)改變數(shù)據(jù)分布來(lái)實(shí)現(xiàn),根據(jù)每次訓(xùn)練集中每個(gè)樣本的分類是否正確,以及上次的總體分類的準(zhǔn)確率,來(lái)確定每個(gè)樣本的權(quán)值。將修改過(guò)權(quán)值的新數(shù)據(jù)集送給下層分類器進(jìn)行訓(xùn)練,最后將每次訓(xùn)練得到的分類器融合起來(lái),作為最終的決策分類器[23-24]。學(xué)習(xí)過(guò)程中將每一個(gè)基于Haar-like特征的判別閾值作為一個(gè)弱分類器,通過(guò)Adaboost算法學(xué)習(xí),得到樣本分布權(quán)重不同的測(cè)試樣本集,每次訓(xùn)練學(xué)習(xí)都會(huì)加大誤判樣本的權(quán)重,減小分類正確樣本的權(quán)重。將改變分布權(quán)重的樣本和其他新樣本構(gòu)成新的訓(xùn)練樣本集,進(jìn)行下一次的學(xué)習(xí)訓(xùn)練。經(jīng)過(guò)次迭代循環(huán)后,會(huì)得到個(gè)弱分類器,將個(gè)弱分類器的權(quán)重進(jìn)行級(jí)聯(lián)最終得到強(qiáng)分類器[25]。
經(jīng)過(guò)Adaboost算法訓(xùn)練學(xué)習(xí)后,得到的強(qiáng)分類器將各級(jí)循環(huán)迭代中各級(jí)弱分類器的誤判率降至最低。將待識(shí)別樣本所提取的Haar-like 特征值作為強(qiáng)分類器的輸入,強(qiáng)分類器根據(jù)特征值權(quán)值給出一個(gè)判斷是否為水稻白穂的評(píng)估值,當(dāng)為1時(shí)表示分類結(jié)果為水稻白穂,若等于?1,則該檢測(cè)樣本不是水稻白穂。
試驗(yàn)硬件環(huán)境為Intel(R) Pentium(R) CPU G3220 3.00 GHz,軟件環(huán)境為Windows 7,VS2010,Intel OpenCV 2.4.3。水稻白穗樣本庫(kù)中共收集350張圖像,其中正樣本190張,負(fù)樣本160張。試驗(yàn)中,選取155張正樣本和130張負(fù)樣本作為訓(xùn)練集,應(yīng)用訓(xùn)練集的420個(gè)正樣本點(diǎn)和280個(gè)負(fù)樣本點(diǎn)訓(xùn)練學(xué)習(xí)生成強(qiáng)分類器,另外剩余的35張正樣本圖像和30張負(fù)樣本圖像作為測(cè)試集,即285張訓(xùn)練集圖像、65張測(cè)試集圖像。訓(xùn)練過(guò)程中,強(qiáng)分類器的檢測(cè)率設(shè)定為0.91,誤判率設(shè)為0.000 1;單層弱分類器檢測(cè)率設(shè)定為0.95,目標(biāo)誤判率設(shè)為0.3;特征窗口大小設(shè)置為24×24。
本文設(shè)計(jì)了4類不同的Haar-like特征,為試驗(yàn)不同特征對(duì)白穗識(shí)別性能的貢獻(xiàn),設(shè)計(jì)了4種Haar-like 特征組合的對(duì)比試驗(yàn),包括獨(dú)立特征試驗(yàn)和多種特征組合試驗(yàn)。試驗(yàn)中將白穗正確識(shí)別率和誤識(shí)別率作為各特征識(shí)別貢獻(xiàn)性能的評(píng)價(jià)指標(biāo),選取測(cè)試集中550個(gè)樣本進(jìn)行測(cè)試,其中190個(gè)白穗樣本,360個(gè)非白穗樣本,各種形態(tài)的樣本數(shù)量如表3所示,試驗(yàn)結(jié)果如表4所示。
表3 樣本集數(shù)量
由表4可看出,本文提出的4種Haar-like 特征都具有80%以上的白穗識(shí)別率,不同特征的組合試驗(yàn)中,C類+D類組合特征正確識(shí)別率最高,誤識(shí)別率最低,分別為94.21%和3.33%;其余特征組合試驗(yàn)正確識(shí)別率和誤識(shí)別率與4種Haar-like 特征單獨(dú)試驗(yàn)性能相當(dāng),甚至低于單獨(dú)特征試驗(yàn)。由試驗(yàn)數(shù)據(jù)得知,第10組試驗(yàn)C類+D類特征是4種Haar-like 特征及其組合特征中分類性能最好的,因此本文識(shí)別算法采用C類+D類Haar-like 組合特征。
表4 Haar-like 特征種類對(duì)白穗識(shí)別的貢獻(xiàn)
注:A類為邊緣特征,B類為線性特征,C類為中心特征,D類為擴(kuò)展特征。
Note: Class A is marginal feature; Class B is linear feature; Class C is central feature; Class D is extended feature.
Haar-like特征描述能力與AdaBoost學(xué)習(xí)過(guò)程中訓(xùn)練次數(shù)共同決定著稻田白穂識(shí)別算法的優(yōu)劣,Haar-like某一類特征或者多類特征的組合在AdaBoost學(xué)習(xí)過(guò)程中并不是訓(xùn)練次數(shù)越多,識(shí)別性能越高,相反在達(dá)到一定的訓(xùn)練次數(shù)后,識(shí)別率不但不會(huì)增加,反而會(huì)因訓(xùn)練次數(shù)的增加影響識(shí)別時(shí)間,進(jìn)而降低目標(biāo)識(shí)別效率。因此本文針對(duì)Haar-like特征分類性能最好的C類+D類特征組合進(jìn)行了AdaBoost訓(xùn)練次數(shù)試驗(yàn),以期在保證識(shí)別率的前提下確定最少的訓(xùn)練時(shí)間,達(dá)到最好的識(shí)別效果。試驗(yàn)中選取300個(gè)樣本進(jìn)行試驗(yàn),其中100個(gè)白穗樣本,200個(gè)非白穗樣本。在樣本選取上,參閱多篇AdaBoost文獻(xiàn)[26-29],正負(fù)樣本比例設(shè)置在1∶1.5~1∶3之間,數(shù)量不固定。本次試驗(yàn)樣本基本參照了本文2.2節(jié)表3的測(cè)試集樣本,并進(jìn)行了一定比例的縮減,另外為了保證試驗(yàn)結(jié)果的可靠性,驗(yàn)證算法的魯棒性,樣本中人為選取了多組具有典型特征的樣本,最終形成了100個(gè)白穗樣本,200個(gè)非白穗樣本,共300容量的試驗(yàn)樣本集。各種形態(tài)的樣本數(shù)量如表5所示,不同訓(xùn)練次數(shù)的對(duì)比試驗(yàn)結(jié)果如表6所示。
表5 樣本集數(shù)量
表6 AdaBoost訓(xùn)練次數(shù)對(duì)比試驗(yàn)(C類+D類)
由C類+D類特征組合的訓(xùn)練次數(shù)對(duì)比試驗(yàn)數(shù)據(jù)可知,訓(xùn)練次數(shù)在25 000次以下時(shí),隨著訓(xùn)練次數(shù)的增加,分類器的分類性能會(huì)逐漸提高,但當(dāng)訓(xùn)練次數(shù)在25 000~50 000次之間時(shí),正確識(shí)別率基本不增加,誤識(shí)別率減小也不明顯,而識(shí)別時(shí)間增加顯著,降低了識(shí)別效率。因此,對(duì)于C類+D類特征組合的AdaBoost學(xué)習(xí),在當(dāng)前試驗(yàn)訓(xùn)練樣本容量情況下,訓(xùn)練次數(shù)選擇25 000次為最佳。
利用經(jīng)過(guò)AdaBoost訓(xùn)練學(xué)習(xí)得到的強(qiáng)分類器對(duì)測(cè)試集中的65張圖像進(jìn)行在線識(shí)別試驗(yàn)。經(jīng)統(tǒng)計(jì),65張無(wú)人機(jī)拍攝的稻田圖像中,包含了423個(gè)白穂樣本。首先對(duì)待識(shí)別圖像進(jìn)行背景分離、閾值分割、去除噪聲等預(yù)處理操作,然后進(jìn)行窗口檢測(cè)、積分圖計(jì)算、Haar-like特征提取等操作,最后將Haar-like特征輸入到強(qiáng)分類器中進(jìn)行分類判別,輸出識(shí)別結(jié)果。圖10為試驗(yàn)過(guò)程中一副含有白穂的樣本圖像識(shí)別過(guò)程。
圖10 圖像預(yù)處理及識(shí)別結(jié)果
試驗(yàn)結(jié)果表明,65張包含了多種白穂形態(tài)、光照影響、葉片遮擋、稻穗粘連以及背景干擾等的圖像中的423個(gè)白穂樣本,有396個(gè)白穂樣本能夠被識(shí)別,正確識(shí)別率達(dá)到93.62%;背景圖像中有23處被誤識(shí)別為白穂,誤識(shí)別率為5.44%。白穂未識(shí)別和誤識(shí)別的情況如圖11所示。圖11a所示為因光照強(qiáng)烈造成的白穗誤識(shí)別,紅色框圖為識(shí)別到的白穗。對(duì)照?qǐng)D11a可看出,圖11b圖中右下角2個(gè)紅色框圖所標(biāo)識(shí)的白穗為誤識(shí)別,原因?yàn)闊o(wú)人機(jī)拍攝時(shí),其飛行姿態(tài)和拍攝角度使得正常稻穗反射強(qiáng)光所致。而對(duì)于圖片上部的4株白穗因光照均勻,無(wú)明顯反射現(xiàn)象,均能正確識(shí)別。因此在強(qiáng)光照射下,個(gè)別稻穗因反射強(qiáng)光而呈現(xiàn)高亮區(qū)域,造成誤識(shí)別;圖11b中白色框圖內(nèi)標(biāo)識(shí)為未識(shí)別到的目標(biāo)白穂,因稻穗粘連遮擋嚴(yán)重所致,對(duì)于稻穗粘連不嚴(yán)重的圖像可以達(dá)到正確識(shí)別的效果;圖11c中白色框圖標(biāo)識(shí)為未識(shí)別到的白穂,因目標(biāo)物被稻葉遮擋嚴(yán)重所致,而對(duì)于最左邊紅色框圖標(biāo)識(shí)的白穗,因其被稻葉遮擋不嚴(yán)重可以被識(shí)別。分析以上錯(cuò)誤識(shí)別原因,后續(xù)算法改進(jìn)中應(yīng)著重考慮強(qiáng)光干擾和拍攝角度的配合,以及稻葉、稻穗遮擋等干擾情況,近一步提高識(shí)別算法的性能。
圖11 白穗誤識(shí)別和未識(shí)別示意圖
為進(jìn)一步驗(yàn)證本算法的有效性,采用2.2節(jié)Haar-Like特征性能分析所用樣本(樣本數(shù)量如表3所示),應(yīng)用輪廓波紋理識(shí)別算法進(jìn)行對(duì)比試驗(yàn)。輪廓波特征提取識(shí)別算法是現(xiàn)有研究中應(yīng)用較為廣泛,識(shí)別精度較高的算法之一,可以很好的捕捉各種角度的圖像邊緣特征,具有旋轉(zhuǎn)不變性的紋理特征提取特點(diǎn),尤其適用于多角度、多方向樣本圖像的特征提取[30-32]。試驗(yàn)中,輪廓波特征提取時(shí),分解層次為4,取分解的各高頻子帶的能量和低頻子帶的均值及標(biāo)準(zhǔn)差作為識(shí)別特征,采用AdaBoost機(jī)器學(xué)習(xí)進(jìn)行分類識(shí)別。圖12為某一樣本稻穗的輪廓波分解圖,試驗(yàn)結(jié)果如表7所示。
圖12 稻穗原圖與輪廓波分解圖
表7 輪廓波對(duì)比試驗(yàn)
由表7知,相同試驗(yàn)樣本情況下,輪廓波識(shí)別算法的白穗正確識(shí)別率為91.05%,誤識(shí)別率為6.67%;C類+D類Haar-like 組合特征識(shí)別算法白穗正確識(shí)別率為94.21%,誤識(shí)別率為3.33%。試驗(yàn)結(jié)果表明,C類+D類Haar-like 組合特征識(shí)別算法在白穗正確識(shí)別率和誤識(shí)別率上均優(yōu)于輪廓波識(shí)別算法。在光照影響突出的樣本中,輪廓波誤識(shí)別率高達(dá)20%,顯著高于Haar-like 組合特征識(shí)別算法。針對(duì)此試驗(yàn)結(jié)果,對(duì)比試驗(yàn)中將輪廓波特征造成誤識(shí)別的樣本再次利用Haar-like特征識(shí)別算法進(jìn)行識(shí)別,試驗(yàn)結(jié)果顯示,輪廓波識(shí)別算法對(duì)強(qiáng)光照?qǐng)D像樣本的誤識(shí)別相對(duì)較多,圖13列出了輪廓波識(shí)別算法中誤識(shí)別的部分非白穗樣本圖像,而這些樣本有一部分在Haar-like特征識(shí)別算法中并未造成誤識(shí)別(框圖中為輪廓波算法因強(qiáng)光照誤識(shí)別的正常稻穗),表明在強(qiáng)光干擾下輪廓波識(shí)別算法較Haar-like特征識(shí)別算法的識(shí)別能力明顯不足。
注:框圖中為輪廓波算法因強(qiáng)光照誤識(shí)別的正常稻穗。
本文以小型多旋翼無(wú)人機(jī)為采集平臺(tái),對(duì)采集到的稻田圖像進(jìn)行預(yù)處理后作為研究對(duì)象,提出了一種基于Haar-like特征和AdaBoost學(xué)習(xí)的稻田白穂識(shí)別算法,并進(jìn)行驗(yàn)證試驗(yàn),試驗(yàn)結(jié)果和結(jié)論如下:
1)針對(duì)小型多旋翼無(wú)人機(jī)現(xiàn)場(chǎng)采集的稻田圖像,提出了基于AdaBoost算法的水稻白穗識(shí)別方法,該方法引入4類Haar-like特征,構(gòu)建一個(gè)強(qiáng)分類器。試驗(yàn)表明4類Haar-like特征及其組合特征中,C類+D類Haar-like 組合特征對(duì)分類器性能提升強(qiáng)于其他特征。
2)本文提出的,C類+D類Haar-like 組合特征可以有效的抑制絕大多數(shù)的稻田背景、稻葉遮擋、稻穗黏連等復(fù)雜情況的影響,在測(cè)試集樣本容量為550的情況下(正樣本190個(gè),負(fù)樣本360個(gè)),稻田白穗的正確識(shí)別率和誤識(shí)別率分別為94.21%和3.33%,不過(guò)對(duì)于高強(qiáng)度光照和嚴(yán)重遮擋情況下的識(shí)別,算法還有待進(jìn)一步優(yōu)化和提高。
3)同輪廓波特征算法進(jìn)行對(duì)比試驗(yàn)表明,C類+D類Haar-like 組合特征識(shí)別效果優(yōu)于輪廓波特征識(shí)別。利用C類+D類Haar-like 組合特征,經(jīng)過(guò)AdaBoost訓(xùn)練學(xué)習(xí)得到的強(qiáng)分類器對(duì)測(cè)試集中的65張圖像,423個(gè)白穂樣本進(jìn)行在線識(shí)別試驗(yàn)。結(jié)果表明:白穗識(shí)別率可達(dá)93.62%,誤識(shí)別率為5.44%。
[1] 高春庭. 水稻白穗的發(fā)生原因及其防治措施[J]. 科學(xué)種養(yǎng),2011(10):29.
Gao Chunting. The cause of the occurrence of rice white panicle and its control measures[J]. Scientific Farming Magazine, 2011(10): 29. (in Chinese with English abstract)
[2] 劉占宇,孫華生,黃敬峰. 基于學(xué)習(xí)矢量量化神經(jīng)網(wǎng)絡(luò)的水稻白穗和正常穗的高光譜識(shí)別[J]. 中國(guó)水稻科學(xué),2007,21(6):664-668.
Liu Zhanyu, Sun Huasheng, Huang Jingfeng. Classification of empty and healthy panicles in rice plants by hyperspectral reflectance based on learning vector quantization (LVQ) neural network[J]. Chinese J Rice Sci, 2007, 21(6): 664-668. (in Chinese with English abstract)
[3] 宋懷波,張衛(wèi)園,張欣欣,等. 基于模糊集理論的蘋(píng)果表面陰影去除方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2014,30(3):135-141.
Song Huaibo, Zhang Weiyuan, Zhang Xinxin, et al. Shadow removal method of apples based on fuzzy set theory[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2014, 30(3): 135-141. (in Chinese with English abstract)
[4] 趙川源,何東健,喬永亮. 基于多光譜圖像和數(shù)據(jù)挖掘的多特征雜草識(shí)別方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2013,29(2):192-198.
Zhao Chuanyuan, He Dongjian, Qiao Yongliang. Identification method of multi-feature weed based on multi-spectral images and data mining[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2013, 29(2): 192-198. (in Chinese with English abstract)
[5] 周天娟,張鐵中,楊麗,等. 基于數(shù)學(xué)形態(tài)學(xué)的相接觸草莓果實(shí)的分割方法及比較研究[J]. 農(nóng)業(yè)工程學(xué)報(bào),2007,23(9):164-168.
Zhou Tianjuan, Zhang Tiezhong, Yang Li, et al. Comparison of two algorithms based on mathematical morphology for segmentation of touching strawberry fruits[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2007, 23(9): 164-168. (in Chinese with English abstract)
[6] Hayashi S, Ganno K, Ishii Y, et al. Robotic harvesting system for eggplants[J]. Japan Agricultural Research Quarterly, 2002, 36(3): 163-168.
[7] Van Henten E J, Van Tuijl B A J, Hemming J, et al. Field test of an autonomous cucumber picking robot[J]. Biosystems Engineering, 2003, 86(3): 305-313.
[8] 謝忠紅,姬長(zhǎng)英,郭小清,等. 基于改進(jìn)Hough變換的類圓果實(shí)目標(biāo)檢測(cè)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2010,26(7):157-162. Xie Zhonghong, Ji Changying, Guo Xiaoqing, et al. An object detection method for quasi-circular fruits based on improved Hough transform[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2010, 26(7): 157-162. (in Chinese with English abstract)
[9] 蔡健榮,趙杰文. 自然場(chǎng)景下成熟水果的計(jì)算機(jī)視覺(jué)識(shí)別[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2005,36(2):61-64.
Cai Jianrong, Zhao Jiewen. Recognition of mature fruit in natural scene using computer vision[J]. Transactions of the Chinese Society of Agricultural Machinery, 2005, 36(2): 61-64. (in Chinese with English abstract)
[10] 詹文田,何東健,史世蓮. 基于Adaboost算法的田間獼猴桃識(shí)別方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2013,29(23):140-146.
Zhan Wentian, He Dongjian, Shi Shilian. Recognition of kiwifruit in field based on Adaboost algorithm[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2013, 29(23): 140-146. (in Chinese with English abstract)
[11] 宋懷波,何東健,潘景朋. 基于凸殼理論的遮擋蘋(píng)果目標(biāo)識(shí)別與定位方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2012,28(22):174-180.
Song Huaibo, He Dongjian, Pan Jingpeng. Recognition and localization methods of occluded apples based on convex hull theory[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(22): 174-180. (in Chinese with English abstract)
[12] 李昕,李立君,高自成,等. 基于偏好人工免疫網(wǎng)絡(luò)多特征融合的油茶果圖像識(shí)別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2012,28(14):133-137.
Li Xin, Li Lijun, Gao Zicheng, et al. Image recognition of camellia fruit based on preference for ai NET multi-features integration[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(14): 133-137. (in Chinese with English abstract)
[13] 趙源深,貢亮,周斌,等. 番茄采摘機(jī)器人非顏色編碼化目標(biāo)識(shí)別算法研究[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2016,47(7):1-7.
ZhaoYuanshen, Gong Liang, Zhou Bin, et al. Object recognition algorithm of tomato harvesting robot using non-color coding approach[J]. Transactions of the Chinese Society of Agricultural Machinrey, 2016, 47(7): 1-7. (in Chinese with English abstract)
[14] 劉占宇,孫華生,黃敬峰. 基于學(xué)習(xí)矢量量化神經(jīng)網(wǎng)絡(luò)的水稻白穗和正常穗的高光譜識(shí)別[J]. 中國(guó)水稻科學(xué),2007,21(6):664-668.
Liu Zhanyu, Sun Huasheng, Huang Jingfeng. Classification of empty and healthy panicles in rice plants by hyperspectral reflectance based on learning vector quantization(LVQ) neural network[J]. Chinese J Rice Sci, 2007, 21(6): 664-668. (in Chinese with English abstract)
[15] 劉占宇,祝增榮,趙敏,等. 基于主成分分析和人工神經(jīng)網(wǎng)絡(luò)的稻穗健康狀態(tài)的高光譜識(shí)別[J]. 浙江農(nóng)業(yè)學(xué)報(bào),2011,23(3):607-616. Liu Zhanyu, Zhu Zengrong, Zhao Min, et al. Hyperspectral discrimination of different health conditions in rice panicles based on principal component analysis and artificial neural network[J]. Acta Agriculture Zhejiangensis, 2011, 23(3): 607-616. (in Chinese with English abstract)
[16] 白曉東. 基于圖像的水稻關(guān)鍵發(fā)育期自動(dòng)觀測(cè)技術(shù)研究[D]. 武漢:華中科技大學(xué),2014. Bai Xiaodong. Image-based Automatic Observation Technology for Critical Growth Stages of Rice[D]. Wuhan: Huazhong University of Science & Technology, 2014. (in Chinese with English abstract)
[17] Kyosuke Yamamoto. On plant detection of intact tomato fruits using image analysis and machine learning methods[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2014, 14(7): 12191-12206.
[18] Sengupta S, Lee W S. Identification and determination of the number of immature green citrus fruit in a canopy under different ambient light conditions[J]. Biosystems Engineering, 2014, 117(1): 51-61.
[19] 劉世晶,王帥,陳軍,等. 基于改進(jìn)主成分分析和AdaBoost算法的運(yùn)動(dòng)蝦苗識(shí)別方法[J]. 農(nóng)業(yè)工程學(xué)報(bào),2017,33(1):212-218.
Liu Shijing, Wang Shuai, Chen Jun, et al. Moving larval shrimps recognition based on improved principal component analysis and AdaBoost[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(1): 212-218. (in Chinese with English abstract)
[20] 文學(xué)志,方巍,鄭鈺輝. 一種基于類Haar特征和改進(jìn)AdaBoost分類器的車(chē)輛識(shí)別算法[J]. 電子學(xué)報(bào),2011,39(5):1121-1126.
Wen Xuezhi, Fang Wei, Zheng Yuhui. An algorithm based on Haar-like features and improved Adaboost classifier for vehiclerecognition[J]. Acta Electronica Sinica, 2011, 39(5): 1121-1126. (in Chinese with English abstract)
[21] Anandhi D, Valli S. An algorithm for multi-sensor image fusion using maximum a posteriori and nonsubsampled contourlet transform[J]. Computers & Electrical Engineering, 2018, 65: 139-152.
[22] Gatica G, Best S, Ceroni J, et al. Olive fruits recognition using neural networks[J]. Procedia Computer Science, 2013, 17: 412-419.
[23] Zhang C X, Zhang J S, Zhang G Y. An efficient modified boosting method for solving classification problems[J]. Journal of Computational & Applied Mathematics, 2008, 214(2): 381-392.
[24] Techo J, Nattee C, Theeramunkong T. Boosting-based ensemble learning with penalty profiles for automatic Thai unknown word recognition[J]. Computers & Mathematics with Applications, 2012, 63(6): 1117-1134.
[25] Duda R O, Hart P E, Stork D G. 模式分類[M]. 李宏?yáng)|,姚天翔,等譯. 北京:機(jī)械工業(yè)出版社,2003:383-387.
[26] 師黎,吳敏,張娟. 基于小訓(xùn)練樣本的AdaBoost人臉檢測(cè)[J].計(jì)算機(jī)工程,2011,37(8):199-201. Shi Li, Wu Min, Zhang Juan. AdaBoost face detection based on few training samples[J]. Computer Engineering, 2011, 37(8): 199-201. (in Chinese with English abstract)
[27] Farias G, Dormido-Canto S, Vega J, et al. Adaboost classification of TJ-II Thomson Scattering images[J]. Fusion Engineering & Design, 2017, 123: 759-763.
[28] Li L, Wang C, Li W, et al. Hyperspectral image classification by adaboost weighted composite kernel extreme learning machines[J]. Neurocomputing, 2018, 275: 1725-1733.
[29] Al-Shemarry M S, Li Y, Abdulla S, et al. Ensemble of adaboost cascades of 3l-Lbps classifiers for license plates detection with low quality images[J]. Expert Systems with Applications, 2018, 92: 216-235.
[30] Jiang Y, Liu C, Ma J. BYY harmony learning of t-mixtures with the application to image segmentation based on contourlet texture features[J]. Neurocomputing, 2016, 188: 262-274.
[31] Biswas S, Sil J. An efficient face recognition method using contourlet and curvelet transform[J]. Journal of King Saud University-Computer and Information Sciences, 2017(10) , https://doi.org/10.1016/j.jksuci.2017.10.010.
[32] Amorim P, Moraes T, Fazanaro D, et al. Electroencephalogram signal classification based on shearlet and contourlet transforms[J]. Expert Systems with Applications, 2017, 67: 140-147.
Identification of diseased empty rice panicles based on Haar-like feature of UAV optical image
Wang Zhen1,2, Chu Guikun1, Zhang Hongjian1, Liu Shuangxi1,2, Huang Xincheng3, Gao Farui3, Zhang Chunqing4, Wang Jinxing1,2※
(1.,,271018,; 2.,271018,; 3.,273013,; 4.,,271018,)
Empty rice panicles are a common pest and disease characteristic in rice fields that affects the rice yield and quality. In order to achieve accurate prevention and control of pests and diseases in rice fields, in this study, a multi-rotor UAV-loaded industrial CCD digital camera was used as the image acquisition platform to rapidly and accurately identify and locate the empty rice panicles in large area rice fields based on the Haar-like feature extraction and Adaboost training algorithm. We used the method of UAV aerial photography technology to perform video capture of large area rice fields on a scheduled route. The interval frame number of the sample image was calculated by parameters such as the flight speed of the UAV, aerial video speed, aerial altitude, and the angle of camera, then the video of the rice field was processed by image disassembly, frame extraction, image mosaic, etc. to achieve efficient and rapid acquisition of image information of large area rice fields. The training sample database and the test sample database for the test were finally formed according to the position information of the rice field coordinate in the frame image extracted by the image extraction interval frame number. After many preprocessing operations, such as compression, cutting, normalization, background separation, threshold segmentation, noise removal, etc., the images in the training sample database and the test sample database were applied in the Haar-like feature extraction and AdaBoost training. In this study, we designed four kinds of Haar-like features, such as edge feature of class A, linear feature of class B, center feature of class C and extension feature of class D, these four kinds of Haar-like features and their combination features were rapidly extracted by the integral diagram calculation, then input the extracted Haar-like features into Adaboost training. During the calculation process, we took each discrimination threshold based on the Haar-like features as a weak classifier to give iterative cycle training, after T times iterative cycles. Then T weak classifiers were obtained, and the strong classifier was obtained after cascading the weights of the T weak classifiers. After the Adaboost training, the obtained strong classifier minimized the misjudgment rate of weak classifiers at all levels in each cycle of iteration. We then took the Haar-like eigenvalue extracted by the unrecognized samples as the input of the strong classifier, based on eigenvalue weight, the strong classifier gave a assessed value H to judge whether it was the empty rice panicles or not. When the H was 1, it meant that the classification result was empty rice panicles. When the H was -1, the tested sample was not the empty rice panicles. In this way, identification of empty rice panicles was realized. In order to ensure the diversity and adequacy of the test samples, the influence of the interference factors such as various forms of the empty rice panicles, lighting, shielding, adhesion and background etc. were fully considered. Two hundred and eight five images and a total of 700 positive and negative samples in the training sample database were used for Haar-like feature extraction and AdaBoost learning training. Sixty five images and a total of 800 positive and negative samples in the test sample database were used to verify the performance of strong classifier. The experimental results showed that among the four Haar-like features and their combined features, the class C and class D Haar-like combined features had better performance in improving classifiers than other features. The strong classifiers generated by this combined features were then used to identify the 423 empty rice panicles samples in the test, among which, three hundred and ninety six were identified, and the recognition rate was 93.62%. Our results demonstrated that this method could effectively inhibit the influence of complex backgrounds such as the rice leaves shielding, rice panicles adhesion and lighting etc., and it was also suitable for field identification of empty rice panicles in natural environment. In the study, this method was compared with algorithms that used texture recognition, such as shear waves, contour waves, curve waves, etc. The experiment showed that this method has significant advantages both in the accuracy and the speed of recognition.
unmanned aerial vehicle; algorithms; diseases; empty rice panicles; Haar-like feature
10.11975/j.issn.1002-6819.2018.20.010
TP391.41
A
1002-6819(2018)-20-0073-10
2018-04-03
2018-07-05
國(guó)家公益性行業(yè)農(nóng)業(yè)科研專項(xiàng)(201303005);山東省現(xiàn)代農(nóng)業(yè)產(chǎn)業(yè)技術(shù)體系創(chuàng)新項(xiàng)目;山東省“雙一流”獎(jiǎng)補(bǔ)資金資助(SYL2017XTTD14)
王 震,講師,博士生,主要從事精準(zhǔn)農(nóng)業(yè)信息化研究。Email:wangzhenxky@sdau.edu.cn
王金星,博士生導(dǎo)師,主要從事農(nóng)業(yè)機(jī)械化工程研究。Email:jinxingw@163.com
王 震,褚桂坤,張宏建,劉雙喜,黃信誠(chéng),高發(fā)瑞,張春慶,王金星. 基于無(wú)人機(jī)可見(jiàn)光圖像Haar-like特征的水稻病害白穂識(shí)別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(20):73-82. doi:10.11975/j.issn.1002-6819.2018.20.010 http://www.tcsae.org
Wang Zhen, Chu Guikun, Zhang Hongjian, Liu Shuangxi, Huang Xincheng, Gao Farui, Zhang Chunqing, Wang Jinxing. Identification of diseased empty rice panicles based on Haar-like feature of UAV optical image[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(20): 73-82. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2018.20.010 http://www.tcsae.org