国产日韩欧美一区二区三区三州_亚洲少妇熟女av_久久久久亚洲av国产精品_波多野结衣网站一区二区_亚洲欧美色片在线91_国产亚洲精品精品国产优播av_日本一区二区三区波多野结衣 _久久国产av不卡

?

深度學(xué)習(xí)在植物葉部病害檢測(cè)與識(shí)別的研究進(jìn)展

2022-05-30 20:33邵明月張建華馮全柴秀娟張凝張文蓉
關(guān)鍵詞:識(shí)別卷積神經(jīng)網(wǎng)絡(luò)深度學(xué)習(xí)

邵明月 張建華 馮全 柴秀娟 張凝 張文蓉

摘要:植物病害準(zhǔn)確檢測(cè)與識(shí)別是其早期診斷與智能監(jiān)測(cè)的關(guān)鍵,是病蟲(chóng)害精準(zhǔn)化防治與信息化管理的核心。深度學(xué)習(xí)應(yīng)用于植物病害檢測(cè)與識(shí)別中,可以克服傳統(tǒng)診斷方法的弊端,大幅提升病害檢測(cè)與識(shí)別的準(zhǔn)確率,引起了廣泛關(guān)注。本文首先收集和介紹了部分公開(kāi)的植物病害圖像數(shù)據(jù)集,然后系統(tǒng)地綜述了近年來(lái)深度學(xué)習(xí)在植物病害檢測(cè)和識(shí)別中的研究應(yīng)用進(jìn)展,闡述了從早期檢測(cè)和識(shí)別算法到基于深度學(xué)習(xí)的檢測(cè)和識(shí)別算法的研究進(jìn)展,以及各算法的優(yōu)點(diǎn)和存在的問(wèn)題。調(diào)研了相關(guān)研究文獻(xiàn),提出了光照、遮擋、復(fù)雜背景、病害癥狀之間相似性、病害在不同時(shí)期癥狀會(huì)有不同的變化以及多種病害交疊共存是目前植物病害檢測(cè)和識(shí)別面臨的主要挑戰(zhàn)。并進(jìn)一步指出,將性能更好的神經(jīng)網(wǎng)絡(luò)、大規(guī)模數(shù)據(jù)集和農(nóng)業(yè)理論基礎(chǔ)相結(jié)合,是未來(lái)主要的發(fā)展趨勢(shì),同時(shí)還指出了多模態(tài)數(shù)據(jù)可以用于植物早期病害的識(shí)別,也是未來(lái)發(fā)展方向之一。本文可為植物病害識(shí)別的深入研究與發(fā)展提供參考。

關(guān)鍵詞:植物;葉部病害;深度學(xué)習(xí);病害檢測(cè);識(shí)別;卷積神經(jīng)網(wǎng)絡(luò);病害圖像數(shù)據(jù)集

中圖分類(lèi)號(hào):S432;TP391.4;TP183????? 文獻(xiàn)標(biāo)志碼:A?????????????? 文章編號(hào):SA202202005

引用格式:邵明月, 張建華, 馮全, 柴秀娟, 張凝, 張文蓉.深度學(xué)習(xí)在植物葉部病害檢測(cè)與識(shí)別的研究進(jìn)展[J].智慧農(nóng)業(yè)(中英文), 2022, 4(1):29-46.

SHAO Mingyue, ZHANG Jianhua, FENG Quan, CHAI Xiujuan, ZHANG Ning, ZHANG Wenrong. Research prog‐ress of deep learning in detection and recognition of plant leaf diseases[J]. Smart Agriculture, 2022, 4(1):29-46.(in Chinese with English abstract)

1 引言

植物病害是影響植物生長(zhǎng)的最復(fù)雜多變且難以克服的因素之一,是一種全球農(nóng)業(yè)生產(chǎn)和生態(tài)安全的生物災(zāi)害。發(fā)生植物病害不僅影響植物的正常生長(zhǎng),造成農(nóng)產(chǎn)品產(chǎn)量與品質(zhì)的降低,還會(huì)帶來(lái)糧食安全問(wèn)題[1,2]。近年來(lái),受全球氣候變暖、農(nóng)業(yè)水資源匱乏及農(nóng)業(yè)耕地面積減少等因素影響,植物病害變得更加普遍和頻繁[3]。據(jù)報(bào)道,每年僅由植物病害造成的全球經(jīng)濟(jì)損失高達(dá)2,200億美元[4]。加強(qiáng)對(duì)植物病害的防治和管理是保證農(nóng)作物高產(chǎn)、農(nóng)產(chǎn)品優(yōu)質(zhì)的關(guān)鍵。

植物病害防治的關(guān)鍵是能夠及時(shí)準(zhǔn)確地檢測(cè)病害危害區(qū)域,并對(duì)其病害類(lèi)型進(jìn)行辨識(shí)[5]。植物病害種類(lèi)繁多,全世界已有990多種植物病毒被確認(rèn)[6]。植物葉片感染病毒、真菌或生理病變后,感染部位的外部形態(tài)特征和內(nèi)部生理結(jié)構(gòu)均會(huì)發(fā)生變化,外部發(fā)生如形變、褪色、卷曲、腐爛、變色等,內(nèi)部主要為水分和色素含量變化等[7]。由于不同病害之間的受害癥狀呈現(xiàn)模糊性、復(fù)雜性和相似性,加之部分農(nóng)民的科技、文化素質(zhì)普遍偏低,不能精確診斷并掌握植物病害的發(fā)生與發(fā)展,往往在植物病害嚴(yán)重時(shí)才大劑量地噴灑農(nóng)藥,容易錯(cuò)過(guò)病害最佳防治時(shí)期,不但造成農(nóng)作物大量減產(chǎn),還嚴(yán)重污染環(huán)境[8,9]。因此,如何快速、簡(jiǎn)便、準(zhǔn)確地檢測(cè)植物病害發(fā)生區(qū)域并對(duì)其病害種類(lèi)進(jìn)行識(shí)別,為病害防治提供必要信息,已成為植物種植面臨的重要問(wèn)題。

深度學(xué)習(xí)概念在2006年由深度學(xué)習(xí)之父 Hinton 正式提出[10],但受當(dāng)時(shí)硬件條件限制,一直未得到學(xué)術(shù)界重視。直到2012年,Kri‐zhevsky等[11] 在圖形處理器(Graphics Process‐ing Unit , GPU )上實(shí)現(xiàn)了一種著名的卷積神經(jīng)網(wǎng)絡(luò) ( Convolutional? Neural? Networks , CNN )——AlexNet,該網(wǎng)絡(luò)在當(dāng)年斯坦福大學(xué)主辦的圖像識(shí)別挑戰(zhàn)賽(ImageNet Large-Scale Visual Recognition Challenge)上的分類(lèi)精度碾壓了其他早期計(jì)算機(jī)視覺(jué)方法。此后,各種 CNN 被陸續(xù)提出,深度學(xué)習(xí)得到快速發(fā)展。

深度學(xué)習(xí)是一種模擬人腦進(jìn)行分析學(xué)習(xí)的神經(jīng)網(wǎng)絡(luò),通過(guò)組合低層特征形成更加抽象的高層表示屬性類(lèi)別或特征,以發(fā)現(xiàn)數(shù)據(jù)的分布式特征表示,解決很多復(fù)雜的模式識(shí)別難題[12]。相對(duì)于淺層學(xué)習(xí),深度學(xué)習(xí)具有學(xué)習(xí)能力強(qiáng)、覆蓋范圍廣、適應(yīng)性好、性能上限高、可移植性好等優(yōu)點(diǎn)[13]。深度學(xué)習(xí)作為新人工智能崛起的代表,不斷取得重大進(jìn)展,很好地解決了人工智能界很多年來(lái)一直努力但仍沒(méi)有得到重要進(jìn)展的問(wèn)題[14]。近年來(lái),深度學(xué)習(xí)技術(shù)逐漸被引入到植物病害檢測(cè)與識(shí)別,在該領(lǐng)域獲得了較快的發(fā)展,并涌現(xiàn)出大量的研究成果,已成為本領(lǐng)域研究熱點(diǎn)。本文對(duì)近年來(lái)基于深度學(xué)習(xí)的植物葉部病害的檢測(cè)和識(shí)別研究進(jìn)行了綜述,分別從植物病害圖像數(shù)據(jù)集、植物病害目標(biāo)檢測(cè)研究進(jìn)展和植物病害識(shí)別研究進(jìn)展等三個(gè)方面進(jìn)行闡述,總結(jié)出植物病害檢測(cè)和識(shí)別目前面臨的挑戰(zhàn),并對(duì)未來(lái)的發(fā)展趨勢(shì)進(jìn)行了展望。

2 植物病害圖像數(shù)據(jù)集

數(shù)據(jù)集是構(gòu)建深度學(xué)習(xí)模型的基礎(chǔ),數(shù)據(jù)集的質(zhì)量和大小決定著深度學(xué)習(xí)模型能否構(gòu)建成功[15]。高質(zhì)量的數(shù)據(jù)集往往能夠提高模型訓(xùn)練的質(zhì)量和預(yù)測(cè)的準(zhǔn)確率,也可有效地提高神經(jīng)網(wǎng)絡(luò)模型的泛化能力。一個(gè)更高質(zhì)量的模型可以更精準(zhǔn)地區(qū)分植物病害的不同類(lèi)型及危害程度,從而提供更科學(xué)的防治措施。據(jù)調(diào)查,目前已形成了多個(gè)已公開(kāi)的植物病害圖像數(shù)據(jù)集,具體如表1所示。

2.1國(guó)外主要植物病害圖像數(shù)據(jù)集

(1) PlantVillage。該數(shù)據(jù)集是 David Hudhes和 Marcel Salathé聯(lián)合創(chuàng)建的公開(kāi)數(shù)據(jù)集,是目前被使用次數(shù)最多的數(shù)據(jù)庫(kù)之一。PlantVillage囊括了14種植物、26種病害,共38個(gè)類(lèi)別,包含54,036幅圖像,非常適用于植物病害檢測(cè)和識(shí)別模型的訓(xùn)練,但PlantVillage中的圖像大多是在實(shí)驗(yàn)室或單一背景下拍攝的,復(fù)雜自然條件下拍攝的圖像較少。

(2) Plant Pathology 2020- FGVC7。主要是高質(zhì)量帶注釋的蘋(píng)果圖像,包括蘋(píng)果黑星病、蘋(píng)果銹病、多種病害共存和健康葉片共3651張圖像。其中蘋(píng)果黑星病圖像1200張,蘋(píng)果銹病圖像1399張,多種病害共存圖像187張,健康葉片圖像865張。

(3) Cucumber Plant Diseases Dataset 。該數(shù)據(jù)集由 Karim Negm分享,共包含695張?zhí)镩g自然條件下拍攝的黃瓜有病的和健康的圖像。

(4) New? Plant? Diseasea? Dataset 。由SamirBhattarai使用數(shù)據(jù)增強(qiáng)技術(shù)重新創(chuàng)建的數(shù)據(jù)集。該數(shù)據(jù)集由38個(gè)不同的類(lèi)別組成,包括健康葉片和非健康葉片共87,000張,但該數(shù)據(jù)圖像背景為單一背景。

(5) PlantDoc。由 Singh等[16]創(chuàng)建,共包含2598張圖像,涵蓋13種植物,17種病害類(lèi)型。

(6) Rice Diseases Image Dataset 。涵蓋水稻褐斑病、葉斑病、鐵甲蟲(chóng)病及健康葉片共5447張圖像。

(7) PlantPathology Apple Dataset 。由 Plant‐ village衍生而來(lái),包括蘋(píng)果黑星病、黑腐病、雪松蘋(píng)果銹和健康葉片4種,共包含3171張圖像。

(8) New Plant Diseases Dataset (Augment‐ ed)。該數(shù)據(jù)集是一個(gè)相關(guān)番茄的數(shù)據(jù)集,基于Plantvillage數(shù)據(jù),經(jīng)數(shù)據(jù)增廣技術(shù)衍生而來(lái),包括9個(gè)番茄病害和1個(gè)健康葉片,共 22, 900張圖像。

(9) PlantifyDr Dataset 。該數(shù)據(jù)集包含10種不同植物類(lèi)型,主要為蘋(píng)果、甜椒、櫻桃、柑橘、玉米、葡萄、桃子、土豆、草莓和番茄。37種植物病害類(lèi)型共125,000張圖像。

(10) Plant Disease Recognition Dataset 。該數(shù)據(jù)集包含健康、粉狀、生銹3種共1530張圖像。

(11) Corn Leaf Diseases ( NLB )。包括患病和健康玉米葉片共4, 115張。

2.2國(guó)內(nèi)植物病害圖像數(shù)據(jù)集

(1) PDD271。Liu 等[17] 收集的植物病害數(shù)據(jù)集 PDD271,包括220,592幅植物葉片圖像,涵蓋271種植物病害類(lèi)別。其中每個(gè)植物病害類(lèi)別至少包含500張圖像,200多株植物。且每張圖像采集時(shí)都由專(zhuān)家及時(shí)標(biāo)注。采集后,由未參與標(biāo)注工作的專(zhuān)家進(jìn)行檢查,以保證標(biāo)簽的正確性。

(2) 農(nóng)業(yè)病蟲(chóng)害研究圖庫(kù)( IDADP )。該數(shù)據(jù)集是由中國(guó)科學(xué)院合肥智能機(jī)械研究所、亞熱帶農(nóng)業(yè)生態(tài)研究所和遙感與數(shù)字地球研究所共同建設(shè)的綜合數(shù)據(jù)庫(kù),涵蓋大田作物、水果和蔬菜等多種植物類(lèi)型及真菌、細(xì)菌和病毒等多種病害種類(lèi)。每種病蟲(chóng)害包括幾百到幾千張圖像。圖像絕大部分用單反相機(jī)拍攝,分辨率不小于兩千萬(wàn)像素(6000×4000,5472×3648),少部分用手機(jī)拍攝,分辨率像素為4128×2322。而且絕大多數(shù)圖像都是從自然條件下拍攝的,可應(yīng)用于復(fù)雜自然條件下植物病害的識(shí)別和檢測(cè)。

(3) 植物疾病癥狀圖像數(shù)據(jù)庫(kù) ( PDDB )。該數(shù)據(jù)集是一個(gè)免費(fèi)的數(shù)據(jù)庫(kù),涵蓋21種植物,171種病害類(lèi)型,共收集接近5萬(wàn)幅圖像。85%的圖像是在真實(shí)條件下拍攝,其他圖像在可控條件下拍攝。圖像全部由數(shù)碼相機(jī)和移動(dòng)設(shè)備拍攝,分辨率在1~24萬(wàn)像素之間[18],且都經(jīng)過(guò)了專(zhuān)家標(biāo)注。

(4) 水稻葉病數(shù)據(jù)集。該數(shù)據(jù)集圖像是在陽(yáng)光直射下以白色背景拍攝的,主要為水稻白葉枯病、褐斑病和黑穗病,每種病害均拍攝40張,共120張。

3 植物病害目標(biāo)檢測(cè)研究進(jìn)展

植物病害目標(biāo)檢測(cè)是利用計(jì)算機(jī)視覺(jué)技術(shù)在復(fù)雜自然條件下檢測(cè)出植物病害侵染區(qū)域及其準(zhǔn)確位置,是植物病害準(zhǔn)確分類(lèi)識(shí)別和病害危害程度評(píng)估的前提,也是植物病害區(qū)域準(zhǔn)確定位并引導(dǎo)植保裝備對(duì)靶噴藥的關(guān)鍵[19]。

早期植物病害目標(biāo)檢測(cè)算法采用滑動(dòng)窗口策略選出候選區(qū)域,然后提取候選區(qū)域特征,最后使用分類(lèi)器進(jìn)行分類(lèi),從而獲得目標(biāo)區(qū)域,如 V-J ( Viola-Jones )檢測(cè)、方向梯度直方圖( His‐togram of Oriented Gradient , HOG )檢測(cè)和有關(guān)可變形部件模型 ( Deformable? Part? Model, DPM )算法等?;瑒?dòng)窗口方法是設(shè)置不同的尺度和寬度對(duì)圖像進(jìn)行遍歷,雖然這種方法應(yīng)用于病害定位檢測(cè)可以不錯(cuò)過(guò)任何一個(gè)病害區(qū)域目標(biāo),但產(chǎn)生的多余候選窗口會(huì)帶來(lái)較大的計(jì)算量,且將病害圖像全部遍歷一遍要花費(fèi)較多時(shí)間,導(dǎo)致檢測(cè)的實(shí)效性差[20]。另外,候選區(qū)域的特征提取采用手工方式,提取的特征較多集中在病害顏色、形狀等底層特征,造成病害檢測(cè)的魯棒性差。分類(lèi)器采用Adaboost、支持向量機(jī)(Support Vector Machine , SVM )等進(jìn)行識(shí)別,識(shí)別速度慢、準(zhǔn)確率低。

3.1基于目標(biāo)檢測(cè)框架的植物病害檢測(cè)

基于深度學(xué)習(xí)的 R-CNN (Region-CNN )系列、YOLO (You Only Look Once)、 SSD ( Sin ‐gle Shot MultiBox Detector)以及CenterNet等新檢測(cè)算法顯著優(yōu)于早期的植物目標(biāo)檢測(cè)算法?;谏疃葘W(xué)習(xí)的目標(biāo)檢測(cè)框架可以分為二階檢測(cè)器( Two-stage) 和一階檢測(cè)器 ( One-stage) 兩大類(lèi)[10]。

3.1.1 基于二階檢測(cè)器的植物病害檢測(cè)

二階檢測(cè)器首先使用候選框生成器生成稀疏的候選框集,并從每個(gè)候選框中提取特征,然后使用區(qū)域分類(lèi)器預(yù)測(cè)候選框區(qū)域的類(lèi)別(如圖1所示)。如基于區(qū)域建議的 CNN ,包括 R-CNN、Fast-RCNN 、Faster-RCNN 及其變體。2014年,Girshick等[21]在一篇會(huì)議論文中提出了 R-CNN,首次使用卷積神經(jīng)網(wǎng)絡(luò)提取圖像特征,開(kāi)啟了利用深度學(xué)習(xí)進(jìn)行目標(biāo)檢測(cè)的大門(mén)。在 R-CNN 基礎(chǔ)上,Girshick [22]提出了 Fast-RCNN ,解決了 R-CNN 在候選區(qū)域選擇的過(guò)程中出現(xiàn)大量重疊框的問(wèn)題。經(jīng)過(guò) R-CNN和 Fast RCNN的積淀,Ren等[23]在 2016年提出了 Faster-RCNN ,將特征提取、邊界框回歸和分類(lèi)集成到一個(gè)網(wǎng)絡(luò)中,使綜合性能有較大提高,在檢測(cè)速度方面尤為明顯。Fuentes 等[24]應(yīng)用Faster-RCNN+VGGNet/ResNet的檢測(cè)框架對(duì)番茄病蟲(chóng)害區(qū)域進(jìn)行定位檢測(cè),其圖像庫(kù)中類(lèi)別有10種病害,平均精度均值(mean? Average? Precision ,mAP ) 值達(dá)到了85.98%,從此 Faster-RCNN逐漸被應(yīng)用到植物病害區(qū)域檢測(cè)上。劉闐宇等[25,26]采用 Faster-RCNN框架,分別采用 ZF Net 和VGGNet作為骨干網(wǎng)絡(luò),能準(zhǔn)確定位葡萄葉片和葉片上的病斑。Ozguven和Adem [27]通過(guò)改變 Faster-RCNN 模型的參數(shù)實(shí)現(xiàn)對(duì)甜菜葉斑病的自動(dòng)檢測(cè),對(duì)155幅甜菜圖像進(jìn)行了訓(xùn)練和測(cè)試,獲得了95.48%的總體分類(lèi)正確率。Bari等[28]使用 Faster-RCNN對(duì)水稻患病葉片圖像和健康葉片圖像進(jìn)行檢測(cè),識(shí)別葉片患病準(zhǔn)確率均在98%以上,表明 Faster-RCNN 可以相對(duì)準(zhǔn)確實(shí)時(shí)地檢測(cè)水稻常見(jiàn)病害。Zhou 等[29] 提出了一種基于 FCM-KM 和 Faster-RCNN 融合的水稻病害快速檢測(cè)方法,以 3010幅圖像為研究基礎(chǔ)數(shù)據(jù)集,得到稻瘟病、白葉枯病和紋枯病的檢測(cè)精度分別為96.71%、97.53%和98.26%,檢測(cè)時(shí)間分別為0.65、0.82和0.53 s。Xie等[30]提出了一種基于改進(jìn)的深度卷積神經(jīng)網(wǎng)絡(luò)——Faster DR-IACNN 模型,在自建的葡萄葉疾病數(shù)據(jù)集 (Grape? Leaf? Disease? Dataset,GLDD )上展開(kāi)研究,并引入了 Inception-v 1模塊、Inception-ResNetv2模塊和壓縮和激勵(lì)網(wǎng)絡(luò)(Squeeze-and-Excitation Networks ,SENet ),該模型具有較高的特征提取能力,mAP精度為81.1%,檢測(cè)速度為15.01 f/s 。上述研究表明,基于二階檢測(cè)器的植物病害目標(biāo)檢測(cè),在檢測(cè)準(zhǔn)確度方面獲得了較好的病害檢測(cè)效果,但由于檢測(cè)速度慢,只能用在實(shí)時(shí)性要求不高的場(chǎng)景中。

3.1.2 基于一階檢測(cè)器的植物病害檢測(cè)

一階檢測(cè)器直接對(duì)特征圖上每個(gè)位置的對(duì)象進(jìn)行類(lèi)別預(yù)測(cè),不經(jīng)過(guò)二階檢測(cè)器中的區(qū)域建議步驟(具體步驟見(jiàn)圖2),如YOLO、SSD及其變體。YOLO 是 Redmon 等[31]于 2016 年提出的一種一階段檢測(cè)算法。YOLO 的設(shè)計(jì)不同于Faster-RCNN,它將檢測(cè)過(guò)程整合為單個(gè)網(wǎng)絡(luò)同時(shí)實(shí)現(xiàn)目標(biāo)區(qū)域預(yù)測(cè)和分類(lèi)的回歸過(guò)程。YOLO 并不生成候選框,而是將圖像劃分成網(wǎng)格,以網(wǎng)格為中心確定目標(biāo)邊界和類(lèi)別,與 Faster-RCNN 相比,YOLO 在滿足更高精度的同時(shí)大大提高了檢測(cè)速度。 Bhatt等[32]在復(fù)雜自然條件下的茶園采集圖像,并提出了一種基于YOLOv3的病蟲(chóng)害檢測(cè)方法,在確保系統(tǒng)實(shí)時(shí)可用性的同時(shí),實(shí)現(xiàn)了mAP為 86%,交并比(Intersection-over-Union,IOU)為 50%。Maski和Thondiyath[33]提出了幾個(gè)輕量級(jí)的 YO‐LO 模型,用于移動(dòng)農(nóng)業(yè)機(jī)器人對(duì)植物病害的檢測(cè),主要針對(duì)木瓜環(huán)斑病建立了一個(gè)大規(guī)模的數(shù)據(jù)集,在此基礎(chǔ)上 tiny-YOLOv4 算法的mAP最高可達(dá) 99.9%。MobileNetV2-YOLOv3 算法在疾病嚴(yán)重程度檢測(cè)方面的最高mAP約為98.39%。李昊等[34]基于改進(jìn)的 YOLOv4實(shí)現(xiàn)柑橘病蟲(chóng)害葉片檢測(cè),并根據(jù)檢測(cè)目標(biāo)框?qū)崿F(xiàn)柑橘病害葉片的局部分割,結(jié)合DenseNet算法對(duì)分割出來(lái)的葉片進(jìn)行病害檢測(cè),檢測(cè)準(zhǔn)確率達(dá)到95.46%。

針對(duì) YOLO 的缺陷, Liu 等[35] 于2016年提出了 SSD 。相較于 YOLO , SSD的改進(jìn)主要包括3個(gè)方面:一是提取不同尺度的特征圖,解決了 YOLO不能準(zhǔn)確檢測(cè)小目標(biāo)的問(wèn)題;二是設(shè)計(jì)了多個(gè)不同尺度的先驗(yàn)框;三是在 VGG16網(wǎng)絡(luò)中增加6個(gè)卷積層來(lái)預(yù)測(cè)邊界框偏移量,解決了 YOLO定位不準(zhǔn)的問(wèn)題。Sun等[36]提出了一種可部署在移動(dòng)設(shè)備上的輕量級(jí)的 MEAN-SSD 病害檢測(cè)模型。 MEAN-SSD 是通過(guò)引入 MEAN 塊(Mobile End AppleNet block)和所有3×3卷積核都替換為 MEAN 塊的 Inception 模塊構(gòu)建而成,mAP能夠達(dá)到83.12%,速度達(dá)到12.53 f/s 。Sun 等[37]提出了一種多尺度特征融合的改進(jìn)的 SSD 算法,該方法結(jié)合了數(shù)據(jù)預(yù)處理、特征融合、特征共享、疾病檢測(cè)等步驟,用于檢測(cè)復(fù)雜背景下玉米葉枯病,mAP比原 SSD 算法的mAP高了20%左右(從71.80%提高到91.83%)。同時(shí)傳輸速度也從24 f/s提高到28.4 f/s ,達(dá)到了實(shí)時(shí)檢測(cè)25 f/s 的標(biāo)準(zhǔn)。Selvaraj 等[38]對(duì)比了幾種著名目標(biāo)檢測(cè)框架和不同骨干網(wǎng)絡(luò)組合對(duì)香蕉病蟲(chóng)害檢測(cè)的效果,數(shù)據(jù)庫(kù)包括了10種香蕉病蟲(chóng)害,共3萬(wàn)余張圖像,發(fā)現(xiàn) SSD框架和MobileNet v1的組合檢測(cè)總體效果最好。雖然經(jīng)過(guò)不斷地改進(jìn)和優(yōu)化,一階段檢測(cè)算法在植物病害精度和速度上都有所提高,但錨框的存在仍然令這種檢測(cè)方法不夠精簡(jiǎn)。

3.2基于無(wú)錨框的植物病害檢測(cè)

2019年,Zhou等[39]提出了一種無(wú)錨框的檢測(cè)算法——CenterNet,該算法是在CornerNet的基礎(chǔ)上改進(jìn)而來(lái),由原來(lái)對(duì)兩個(gè)關(guān)鍵點(diǎn)(即圖像的左上角和右下角)的檢測(cè)改為對(duì)圖像中心點(diǎn)的估計(jì)。由于該算法去掉了生成錨框這一操作,并且由熱力圖估計(jì)損失,省去了一些耗時(shí)的操作,所以很大程度上提升了檢測(cè)性能。目前,基于CenterNet的病害檢測(cè)研究還較少,但CenterNet已被證明可以應(yīng)用于自然條件下的目標(biāo)檢測(cè)。夏雪等[40]通過(guò)CenterNet檢測(cè)網(wǎng)絡(luò)與MobileNet v3相結(jié)合,構(gòu)建一個(gè)新的網(wǎng)絡(luò)——M-CenterNet,對(duì)自然條件下果樹(shù)上的蘋(píng)果進(jìn)行檢測(cè)。并與 Cen‐terNet和 SSD做對(duì)比,發(fā)現(xiàn)所提網(wǎng)絡(luò)不論是檢測(cè)精度還是檢測(cè)速度上都比CenterNet和 SSD 要好很多,尤其是檢測(cè)速度上,比這兩種網(wǎng)絡(luò)提高了1倍左右。Albattah等[41]提出了一種改進(jìn)的 Cen ‐terNet算法,以PlantVillageKaggle數(shù)據(jù)庫(kù)為主要數(shù)據(jù)來(lái)源,以 DenseNet-77為基礎(chǔ)網(wǎng)絡(luò)對(duì)深層次關(guān)鍵點(diǎn)進(jìn)行提取,然后分別對(duì)番茄、蘋(píng)果、葡萄等在內(nèi)的14種植物26類(lèi)病害及12類(lèi)健康葉片進(jìn)行識(shí)別,從多方面分析得出,改進(jìn)的CenterNet方法比目前最新的EfficientNet方法能夠更準(zhǔn)確地識(shí)別植物病害。無(wú)錨框的檢測(cè)算法在性能上優(yōu)于基于錨框的檢測(cè)算法,是今后病害區(qū)域檢測(cè)方面的主要研究方向。

3.3植物病害目標(biāo)檢測(cè)分析與展望

近年來(lái)植物病害檢測(cè)研究進(jìn)展見(jiàn)表2~表4??梢钥闯觯槍?duì)大豆、玉米、馬鈴薯、蘋(píng)果、葡萄等植物病害目標(biāo)檢測(cè),一階段檢測(cè)算法和二階檢測(cè)算法都獲得了較好的檢測(cè)效果。但目前文獻(xiàn)表明,在病斑邊界框(bounding box)標(biāo)識(shí)時(shí)較為混亂,一些文獻(xiàn)對(duì)一張葉片上大的病斑單獨(dú)框出,一些文獻(xiàn)則對(duì)小且多的病斑往往采用一個(gè)框,對(duì)沒(méi)有明顯邊界的病害則往往不考慮在病害檢測(cè)任務(wù)范疇內(nèi)。同時(shí),對(duì)植物生長(zhǎng)的復(fù)雜自然場(chǎng)景中進(jìn)行病害目標(biāo)檢測(cè)研究較少,這種場(chǎng)景下密集、小目標(biāo)檢測(cè)算法有待進(jìn)一步研究,同時(shí)還需應(yīng)對(duì)復(fù)雜自然條件下可能出現(xiàn)的光照、陰影、復(fù)雜背景、遮擋、疊加、小病斑檢測(cè)等難點(diǎn)。

如今已提出的病害檢測(cè)算法均對(duì)特定的數(shù)據(jù)集有較好的檢測(cè)效果,但若數(shù)據(jù)集發(fā)生了改變,則可能會(huì)導(dǎo)致檢測(cè)效果不佳,所以在未來(lái),提高模型的魯棒性是值得研究的一個(gè)方向。另外,早期病害的檢測(cè)研究仍處于空白階段,主要因?yàn)閿?shù)據(jù)采集的困難。早期病害部位信息較少,研究者無(wú)法保證準(zhǔn)確識(shí)別病害種類(lèi)與病斑位置,但早期病害檢測(cè)更有利于防止病菌的傳播與發(fā)展,有效防治植物病害,所以今后應(yīng)重視開(kāi)展對(duì)早期病害檢測(cè)的開(kāi)發(fā)研究,以期達(dá)到及時(shí)防治、減少損失的目的。目前對(duì)植物的病害檢測(cè)還處于有人工干預(yù)的半自動(dòng)化過(guò)程,探索全自動(dòng)化的病害檢測(cè)方法也將是未來(lái)主要研究方向之一。

4 植物病害識(shí)別研究進(jìn)展

植物病害識(shí)別是指對(duì)病害圖像進(jìn)行處理、分析和理解,以辨識(shí)不同種類(lèi)病害對(duì)象的技術(shù),是植物病害及時(shí)有效防治的主要依據(jù)。

早期植物病害識(shí)別方法中,病害特征的提取和選擇是依據(jù)先驗(yàn)經(jīng)驗(yàn)人工完成的,識(shí)別性能好壞主要取決于所提取與選擇的特征是否能充分表達(dá)待識(shí)別特定對(duì)象的信息和具有高可分性,以及與后續(xù)分類(lèi)器的匹配性。傳統(tǒng)的圖像特征包括形狀、顏色、紋理特征等,曾被廣泛地用于植物病害的分類(lèi)。形狀特征包括基于物體邊界形狀的方法,如傅里葉形狀描述符以及基于區(qū)域的形狀表示方法,如形狀不變矩、小波變換方法、小波輪廓描述符等。顏色特征包括圖像各個(gè)顏色分量的一階/二階灰度、直方圖、均值等參數(shù);紋理特征如灰度共生矩陣、分形特征、小波、Gabor等,以及描述分形特征的分形維數(shù),與圖像的粗糙程度直接相關(guān),用于對(duì)病斑紋理變化的描述。為提高識(shí)別性能,常見(jiàn)的策略是采用各種圖像特征的組合以及分類(lèi)器的集成。人工設(shè)計(jì)提取優(yōu)秀的識(shí)別特征的要求較高,需要從事此項(xiàng)工作的人具有豐富的工程技能和領(lǐng)域?qū)I(yè)知識(shí),但受植物類(lèi)型、生長(zhǎng)階段、病害種類(lèi)、環(huán)境光照等因素的影響,植物病害癥狀復(fù)雜,人工無(wú)法對(duì)特征設(shè)計(jì)進(jìn)行細(xì)致的優(yōu)化,導(dǎo)致復(fù)雜場(chǎng)景下的植物病害識(shí)別效果不佳。

深度學(xué)習(xí)可以使用通用學(xué)習(xí)程序從原始數(shù)據(jù)(如圖像像素)中自動(dòng)學(xué)習(xí)特征,能避免人工特征工程的局限。近年來(lái),作為深度學(xué)習(xí)核心的各種卷積神經(jīng)網(wǎng)絡(luò)被陸續(xù)提出,如 ZF Net (2013)、VGG? (2014)、GoogleNet? (2014)、ResNet(2015)、DenseNet (2017)、MobileNet (2017)以及EfficientNet (2019)等,同時(shí),各個(gè)卷積神經(jīng)網(wǎng)絡(luò)都在不斷更新與優(yōu)化中,以適應(yīng)不同的任務(wù)和性能要求?;谏疃葘W(xué)習(xí)的植物病害識(shí)別基本步驟見(jiàn)圖3。

4.1基于深度網(wǎng)絡(luò)的植物病害識(shí)別

相較于早期植物病害識(shí)別方法,基于深度網(wǎng)絡(luò)的植物病害識(shí)別可以自動(dòng)對(duì)圖像進(jìn)行預(yù)處理,不再需要通過(guò)人工處理圖像,因此大大提高了病害識(shí)別的效率。同時(shí),隨著網(wǎng)絡(luò)深度的增加,模型的學(xué)習(xí)能力更強(qiáng),提取的特征更豐富。但由于網(wǎng)絡(luò)深度的增加,訓(xùn)練過(guò)程會(huì)消耗大量的時(shí)間,并且可能會(huì)出現(xiàn)在訓(xùn)練時(shí)過(guò)擬合的問(wèn)題。不少研究人員嘗試在植物病害識(shí)別中采用以 CNN 為主的深度神經(jīng)網(wǎng)絡(luò),以提高識(shí)別的準(zhǔn)確性。Kawa‐saki等[68]最早采用 CNN 進(jìn)行病害識(shí)別,基于健康葉片與2種病害葉片類(lèi)型的800幅黃瓜病害圖像訓(xùn)練 CNN ,訓(xùn)練時(shí)用4-fold交叉驗(yàn)證,最終獲得了94.9%的準(zhǔn)確率。Sladojevic等[69]從互聯(lián)網(wǎng)上搜集了2589張圖像,包括桃、梨、蘋(píng)果等6種植物葉片的13種病害、正常葉片和背景共計(jì)15種類(lèi)型,通過(guò)仿射、投影和旋轉(zhuǎn)等數(shù)據(jù)增廣技術(shù)增加了圖像數(shù)量;采用CaffeNet和遷移學(xué)習(xí),特征提取層參數(shù)在 ImageNet 上預(yù)訓(xùn)練,只在最后分類(lèi)的全連接層進(jìn)行了精調(diào)(fine-tuning),該網(wǎng)絡(luò)的平均分類(lèi)準(zhǔn)確度達(dá)到了96.3%。 Mohanty 等[70]? 比較了AlexNet和GoogleNet兩種網(wǎng)絡(luò)性能,訓(xùn)練數(shù)據(jù)庫(kù)為PlantVillage,發(fā)現(xiàn)當(dāng)訓(xùn)練數(shù)據(jù)與測(cè)試數(shù)據(jù)之比是8:2并采用遷移學(xué)習(xí)時(shí),在GoogleNet上平均識(shí)別準(zhǔn)確率高達(dá)99.34%。Ram ‐charan等[71]在田間拍攝了2756張木薯葉片圖像,將剪切的15,000張小葉片圖像構(gòu)造數(shù)據(jù)庫(kù),采用了 Inception v3和遷移學(xué)習(xí)網(wǎng)絡(luò)模型,并與傳統(tǒng)的 SVM和 K最鄰近( KNN ,K-NearestNeighbor)分類(lèi)器進(jìn)行了對(duì)比,識(shí)別3種病害和2種蟲(chóng)害,發(fā)現(xiàn) Inception v3的平均準(zhǔn)確率為93%,高于傳統(tǒng)分類(lèi)器。為解決網(wǎng)絡(luò)收斂時(shí)間長(zhǎng)和參數(shù)內(nèi)存需求大的問(wèn)題,孫俊等[72]在AlexNet的基礎(chǔ)上采用批歸一化方法以及加入全局池化層和縮減特征圖數(shù)目的方法,得到8種改進(jìn)模型,對(duì)PlantVil‐lage中的14種植物共26種病害進(jìn)行識(shí)別,其最優(yōu)模型的平均測(cè)試識(shí)別準(zhǔn)確率達(dá)到99.56%。Lu 等[73]采集了9230張大田小麥圖像用于訓(xùn)練和測(cè)試,包含6種常見(jiàn)小麥病害,采用深度學(xué)習(xí)框架基于多實(shí)例學(xué)習(xí)的弱監(jiān)督方法,在完成診斷功能的同時(shí),可以通過(guò)全卷積網(wǎng)絡(luò)(Fully Convolu‐tional Network ,F(xiàn)CN )算法分割出病害區(qū)域,平均識(shí)別精度為97.95%。Ferentinos [74] 比較了5種典型 CNN 模型的病害識(shí)別精度,該研究使用了一種包括25種植物和57種病害的數(shù)據(jù)庫(kù),共計(jì)87,000張葉片病害圖像,其中37.3%是在大田條件下拍攝,62.7%是在實(shí)驗(yàn)室中拍攝;訓(xùn)練方法是從頭訓(xùn)練而非遷移訓(xùn)練,結(jié)果表明 VGG 模型效果最好,平均識(shí)別準(zhǔn)確率為99.53%。趙建敏等[75]基于 TensorFlow框架,搭建8層CNN+soft‐ max分層卷積神經(jīng)網(wǎng)絡(luò)模型,自動(dòng)學(xué)習(xí)到256個(gè)病害圖像特征,采用softmax分類(lèi)器識(shí)別病害,簡(jiǎn)單背景單一病斑識(shí)別準(zhǔn)確率達(dá)到87%。Xing等[76]在自然環(huán)境下采集了17種柑橘類(lèi)害蟲(chóng)圖像和7種柑橘類(lèi)病害圖像共12,561張,對(duì)這些圖像進(jìn)行了增廣處理,對(duì)比6種網(wǎng)絡(luò)模型的識(shí)別效果,其中在DenseNet基礎(chǔ)上做了一些簡(jiǎn)化的Weakly DenseNet的效果最好,平均準(zhǔn)確率為93.42%。曾偉輝等[77]提出了一種高階殘差和參數(shù)共享反饋的卷積神經(jīng)網(wǎng)絡(luò)模型,高階殘差子網(wǎng)絡(luò)為病害表觀提供豐富細(xì)致的特征表達(dá),以提高模型識(shí)別精度,參數(shù)共享反饋?zhàn)泳W(wǎng)絡(luò)用來(lái)抑制原深層特征中的背景噪聲,可提高模型的魯棒性,該網(wǎng)絡(luò)在PlantVillage上分類(lèi)效果優(yōu)于幾種傳統(tǒng)的分類(lèi)器。Ji 等[78] 使用多個(gè) CNN 組成 United‐Model 提取互補(bǔ)的病害特征用于葡萄病害的識(shí)別,在PlantVillage上取得了很好識(shí)別效果。

4.2基于輕量型網(wǎng)絡(luò)的植物病害識(shí)別

近些年,為了適應(yīng)復(fù)雜自然場(chǎng)景中數(shù)據(jù)資源有限的局限性,研究人員展開(kāi)了對(duì)輕量型網(wǎng)絡(luò)的研究。劉洋等[79]在安卓手機(jī)上實(shí)現(xiàn)了PlantVil‐lage數(shù)據(jù)集病害圖像識(shí)別,比較了MobileNet和Inception v3兩種輕量級(jí)網(wǎng)絡(luò)的性能,結(jié)果表明兩種網(wǎng)絡(luò)識(shí)別精度相差很小,但基于MobileNet模型的程序更小,運(yùn)行速度更快。王春山等[80]提出了改進(jìn)型多尺度殘差網(wǎng)絡(luò),通過(guò)改變殘差層連接方式,將大卷積核分解,減少了模型參數(shù),設(shè)計(jì)了輕量級(jí)病害識(shí)別模型。Saleem 等[81]利用Xception對(duì)PlantVillage中的26種疾病進(jìn)行分類(lèi),通過(guò)對(duì)先進(jìn)的 CNN 體系結(jié)構(gòu)和改進(jìn)的深度學(xué)習(xí)模型進(jìn)行對(duì)比分析,選出了識(shí)別效果最好的Xception模型,然后通過(guò)各種優(yōu)化器訓(xùn)練,選取提高性能最多的 Adam優(yōu)化器,對(duì)數(shù)據(jù)集中植物病害的識(shí)別準(zhǔn)確率最終達(dá)到99.81%。De Ocamop和Dadios [82]用輕量型網(wǎng)絡(luò)MobileNet對(duì)植物病害進(jìn)行分類(lèi),證明MobileNet在數(shù)據(jù)集相對(duì)較小的情況下也能有效地對(duì)植物病害進(jìn)行分類(lèi)。

4.3植物病害同步檢測(cè)與識(shí)別

植物病害檢測(cè)和識(shí)別不僅可以采取先檢測(cè)病害區(qū)域(即感興趣區(qū)域或 IOT區(qū)域)再進(jìn)行病害識(shí)別的分步操作,還可以同步進(jìn)行,即對(duì)病害區(qū)域檢測(cè)的同時(shí)判別病害種類(lèi)。De Luna等[83] 以自然條件下采集的4923張患病及健康的番茄葉片圖像為研究對(duì)象,設(shè)計(jì)了一個(gè)自動(dòng)圖像采集系統(tǒng)和一個(gè)番茄病害檢測(cè)和識(shí)別模型。利用以AlexNet為基礎(chǔ)特征提取網(wǎng)絡(luò)的 F-RCNN 實(shí)現(xiàn)病害檢測(cè)的同時(shí),對(duì)番茄葉片病害進(jìn)行了分類(lèi),該模型最終準(zhǔn)確率達(dá)到95.75%。Rashid 等[84]針對(duì)病害檢測(cè)和識(shí)別研究無(wú)法檢測(cè)作物種類(lèi)和作物病害的問(wèn)題,建立了一個(gè)多層次深度學(xué)習(xí)模型來(lái)識(shí)別馬鈴薯葉片病害,首先使用 YOLO v5對(duì)馬鈴薯葉片進(jìn)行檢測(cè),同時(shí)使用深度學(xué)習(xí)技術(shù) PD ‐ DCNN對(duì)馬鈴薯葉片上的病害進(jìn)行分類(lèi),得到最終準(zhǔn)確率為99.75%;同時(shí),他們還與 VGG16、 InceptionResNetV2、DenseNet_121、DenseNet169、Xception等模型進(jìn)行了比較,證實(shí)該模型在精度和計(jì)算代價(jià)方面具有優(yōu)勢(shì)。Kiratiratanapruk等[85] 使用了4種卷積神經(jīng)網(wǎng)絡(luò)預(yù)訓(xùn)練模型( Faster- RCNN、RetinaNet、YOLOv3和 Mask RCNN)來(lái)檢測(cè)6種常見(jiàn)的水稻病害,以在自然條件下拍攝的6330張照片為數(shù)據(jù)集,并在該數(shù)據(jù)集上使用了上述4種模型,同時(shí)實(shí)現(xiàn)了病葉檢測(cè)和病害分類(lèi)功能,YOLOv3模型最終表現(xiàn)良好,準(zhǔn)確率達(dá)到79.19%。

4.4植物病害識(shí)別分析與展望

根據(jù)卷積神經(jīng)網(wǎng)絡(luò)模型規(guī)模,神經(jīng)網(wǎng)絡(luò)可分為深度網(wǎng)絡(luò)和輕量型網(wǎng)絡(luò),近年植物病害識(shí)別研究進(jìn)展總結(jié)見(jiàn)表5和表6(第40頁(yè))。

植物病害圖像數(shù)據(jù)庫(kù)無(wú)論是開(kāi)放還是私有的,都有一個(gè)共同點(diǎn),即病害部位基本在圖像中間,且占據(jù)圖像的大部分空間,這表明目前的研究似乎集中在人機(jī)交互場(chǎng)景,是在部分可控拍攝條件下進(jìn)行病害識(shí)別,與植保機(jī)器人面臨的實(shí)際情況相差較大。大量實(shí)驗(yàn)表明,基于數(shù)據(jù)庫(kù)識(shí)別效果最好的 CNN 在不同類(lèi)型病害圖像庫(kù)表現(xiàn)并非始終最好,這表明需要通過(guò)大量試驗(yàn)才能為特定場(chǎng)景的病害識(shí)別篩選出最適合的網(wǎng)絡(luò)模型。雖然有學(xué)者研究了基于移動(dòng)設(shè)備的病害識(shí)別,但處理速度通常在數(shù)百毫秒以上,并非實(shí)時(shí)識(shí)別。

從深度學(xué)習(xí)在植物病害識(shí)別的研究進(jìn)展看,在研究?jī)?nèi)容和網(wǎng)絡(luò)模型等方面逐漸發(fā)生轉(zhuǎn)變。在研究?jī)?nèi)容方面,逐漸從提升 CNN 的病害識(shí)別準(zhǔn)確率轉(zhuǎn)變成運(yùn)算速度提升,從實(shí)驗(yàn)室單一背景植物病害識(shí)別轉(zhuǎn)變成自然條件下復(fù)雜背景植物病害識(shí)別,從靜態(tài)的病害圖像檢測(cè)轉(zhuǎn)變成動(dòng)態(tài)的視頻檢測(cè),從單一植物葉部病害特征提取轉(zhuǎn)變成對(duì)根、莖、葉、花、果不同植物器官的病害癥狀特征提取。在 CNN 模型方面,從 VGG、GoogleNet、ResNet、DenseNet等網(wǎng)絡(luò)向 Mo ‐bileNet、EfficientNet輕量型網(wǎng)絡(luò)轉(zhuǎn)變,保持性能不降低情況下盡量壓縮模型參數(shù)、提升網(wǎng)絡(luò)運(yùn)行速度,確保網(wǎng)絡(luò)模型可在算力有限的人工智能邊緣計(jì)算平臺(tái)上運(yùn)行。

5 面臨的挑戰(zhàn)與展望

5.1面臨的挑戰(zhàn)

雖然近年來(lái)國(guó)內(nèi)外研究者基于深度學(xué)習(xí)技術(shù)開(kāi)展了大量的植物病害目標(biāo)檢測(cè)與分類(lèi)識(shí)別研究,推動(dòng)了檢測(cè)與識(shí)別準(zhǔn)確度的提高,但在實(shí)際應(yīng)用場(chǎng)景中,植物病害檢測(cè)與識(shí)別仍然面臨著諸多挑戰(zhàn)。

(1) 光照變化導(dǎo)致目標(biāo)區(qū)域準(zhǔn)確定位難。在實(shí)際植物種植環(huán)境中,一日之內(nèi)光照變化劇烈、不同背景下的反光、不同氣象條件等影響因素,造成植物病斑目標(biāo)區(qū)域難以準(zhǔn)確定位。自然光照條件下,拍攝的角度、高度或者地點(diǎn)可能會(huì)導(dǎo)致部分圖片中病斑位置的顏色深淺不一,使得病斑特征不明顯,從而影響分類(lèi)識(shí)別準(zhǔn)確度。

(2) 背景復(fù)雜導(dǎo)致目標(biāo)的準(zhǔn)確檢測(cè)難度大。在實(shí)際植物種植環(huán)境中獲取的病害圖像背景有可能會(huì)包括葉片、樹(shù)干、莖稈、根部、土壤、雜草、秸稈、地膜、落葉、石頭、積水、陰影等,在復(fù)雜背景條件下獲取的植物病害圖像對(duì)于病斑的目標(biāo)檢測(cè)難度較大。同時(shí),植物病斑顏色形狀等有可能與背景中的其他對(duì)象相似,造成目標(biāo)檢測(cè)的準(zhǔn)確率降低。

(3) 遮擋導(dǎo)致目標(biāo)特征缺失、噪聲重疊。目前,大多數(shù)研究者都避免對(duì)復(fù)雜環(huán)境下植物病害的識(shí)別,采用直接截取所采集圖像感興趣的區(qū)域的方法,很少考慮遮擋問(wèn)題。遮擋問(wèn)題在復(fù)雜自然環(huán)境中普遍存在,包括由葉片姿態(tài)變化引起的葉片遮擋、分支遮擋、外部光照引起的光遮擋以及不同遮擋類(lèi)型引起的混合遮擋。遮擋條件下植物病害識(shí)別的難點(diǎn)在于特征缺失和遮擋引起的噪聲重疊。不同的遮擋條件對(duì)識(shí)別算法有不同程度的影響,導(dǎo)致誤檢甚至漏檢。

(4) 病害相似性導(dǎo)致錯(cuò)判或者誤判。不同的病害引起的癥狀具有相似性。癥狀是判斷病害種類(lèi)的主要依據(jù)之一,若不同種類(lèi)病害的發(fā)病癥狀極為相似,通過(guò)二維圖像無(wú)法準(zhǔn)確地辨識(shí),需要獲取更多維度信息如深度信息、光譜信息、紅外信息、熒光信息等,才能準(zhǔn)確判斷出植物病害類(lèi)別。

(5) 病害癥狀變化導(dǎo)致病害識(shí)別難度大。病原菌可以在植物不同時(shí)期進(jìn)行侵染,發(fā)病時(shí)又會(huì)因植物的品種、生育期和器官表現(xiàn)出不同的癥狀,同一種病害在不同的危害時(shí)期或不同侵染程度下表現(xiàn)出不同癥狀。同一種病害危害植物的不同組織或植物器官癥狀會(huì)有差異,如嫩芽、子葉、真葉、果實(shí)、莖稈、根部等呈現(xiàn)出來(lái)的癥狀各有不同。同一種病害在同種植物器官上也會(huì)呈現(xiàn)不同的癥狀類(lèi)型,比如棉花黃萎病常見(jiàn)的癥狀有4種類(lèi)型,分別為黃斑型、葉枯型、萎蔫型和落葉型,這對(duì)病害識(shí)別提出了很大的挑戰(zhàn)。

(6) 多重病害交疊導(dǎo)致植物病害的檢測(cè)和識(shí)別準(zhǔn)確率低。目前提到的病害檢測(cè)和識(shí)別都是基于每片葉子上均是一種病害或一種病害特征最為明顯的情況而研究的。但在自然條件下,常見(jiàn)多種病害同時(shí)存在于單片葉子的情況,還存在病害與蟲(chóng)害相互重疊現(xiàn)象,使植物病害檢測(cè)和識(shí)別成為一項(xiàng)復(fù)雜的工作。

5.2展望

深度學(xué)習(xí)作為新一代人工智能技術(shù)有著兩方面的優(yōu)勢(shì):一是可隨著數(shù)據(jù)規(guī)模的增加不斷提升其性能;二是可以從數(shù)據(jù)中直接提取特征,削減對(duì)每一個(gè)問(wèn)題設(shè)計(jì)特征提取器的工作量。因此,作為大數(shù)據(jù)時(shí)代的算法利器,深度學(xué)習(xí)技術(shù)受到各個(gè)國(guó)家的高度重視,關(guān)于卷積神經(jīng)網(wǎng)絡(luò)的研究有很多。

基于深度學(xué)習(xí)的植物病害的檢測(cè)和識(shí)別技術(shù)目前已經(jīng)發(fā)展的較為成熟,但在自然條件下依舊面臨著很多的挑戰(zhàn)??傮w而言,難點(diǎn)主要集中于背景和病害本身特性的復(fù)雜上,為了克服這些難點(diǎn),需要性能更好的神經(jīng)網(wǎng)絡(luò)和更豐富的數(shù)據(jù)集。在未來(lái),深度神經(jīng)網(wǎng)絡(luò)的性能將不斷提升,深度神經(jīng)網(wǎng)絡(luò)節(jié)點(diǎn)功能不斷豐富,深度神經(jīng)網(wǎng)絡(luò)工程化應(yīng)用技術(shù)不斷深化。數(shù)據(jù)集也將從單一可控背景下拍攝收集轉(zhuǎn)換為自然條件復(fù)雜背景下采集。另外,在數(shù)據(jù)模態(tài)方面,也將從單一模態(tài)的視覺(jué)通道向擁有近紅外光譜、高光譜、紅外熱圖像、深度、熒光等多模態(tài)數(shù)據(jù)轉(zhuǎn)變,病害識(shí)別的準(zhǔn)確率進(jìn)一步提升,同時(shí)可以利用多模信息開(kāi)展早期植物病害的判別。隨著深度學(xué)習(xí)技術(shù)的快速發(fā)展,如何將最新深度學(xué)習(xí)技術(shù)與植物病害檢測(cè)和識(shí)別相結(jié)合,解決檢測(cè)器定位能力差、識(shí)別模型精度低、算法泛化性能弱以及構(gòu)建大規(guī)模數(shù)據(jù)集合難等方面的問(wèn)題,形成可面向復(fù)雜自然條件、多種植物病害、應(yīng)用于實(shí)際大田作業(yè)的模型算法,提升田間病害的智能監(jiān)測(cè)水平,創(chuàng)制出適用于田間作業(yè)的智能植保機(jī)械裝備,是該領(lǐng)域未來(lái)主要的研究方向。

同時(shí)對(duì)植物病害的檢測(cè)和識(shí)別并不是一個(gè)簡(jiǎn)單的工作,不能僅靠植物表面圖像來(lái)獲得可靠的結(jié)果。并且,在實(shí)際應(yīng)用場(chǎng)景中,植物葉片上時(shí)常有多種病害共存,這將對(duì)檢測(cè)和識(shí)別產(chǎn)生很大的干擾,未來(lái)植物病害檢測(cè)和識(shí)別應(yīng)該更好地與農(nóng)業(yè)理論基礎(chǔ)相結(jié)合,在對(duì)圖像分析的基礎(chǔ)上,還要考慮環(huán)境因素、作物生長(zhǎng)規(guī)律以及病菌的生物學(xué)特性等多種因素,進(jìn)一步提高病害識(shí)別和檢測(cè)結(jié)果的實(shí)用性,同時(shí)讓多種病害同時(shí)檢測(cè)和識(shí)別成為可能。

參考文獻(xiàn):

[1] NAVROZIDIS? I, ALEXANDRIDIS? T K,? DIMITRA ‐KOS A, et al. Identification of purple spot disease on asparagus? crops? across? spatial? and? spectral? scales[J]. Computers and Electronics in Agriculture, 2018, 148:322-329.

[2] MA J, DU K, ZHENG F, et al. A recognition methodfor? cucumber? diseases? using? leaf? symptom? images based on deep convolutional neural network[J]. Com ‐puters and Electronics in Agriculture, 2018, 154:18-24.

[3] CHEN P, XIAO Q, ZHANG J, et al. Occurrence pre‐diction? of cotton? pests? and? diseases? by? bidirectional long short-term memory networks with climate and at‐mosphere circulation[J]. Computers and Electronics in Agriculture, 2020, 176: ID 105612.

[4] 張建華, 韓書(shū)慶, 翟治芬, 等.改進(jìn)自適應(yīng)分水嶺方法分割棉花葉部粘連病斑[J].農(nóng)業(yè)工程學(xué)報(bào) , 2018, 34(24):165-174.

ZHANG? J, HAN? S, ZHAI Z,? et? al. Segmentation? of cotton leaf adhesiosis spot by improved adaptive water‐ shed method [J]. Transactions of the CSAE, 2018, 34(24):165-174.

[5] VISHNOI V K, KUMAR K, KUMAR B. Plant diseasedetection using? computational? intelligence? and? imageprocessing[J]. Journal? of Plant? Diseases? and? Protec‐tion, 2021, 128(1):19-53.

[6] 高秀美, 曹長(zhǎng)余, 邵增順, 等.園林植物病蟲(chóng)害發(fā)生特點(diǎn)與防治對(duì)策[J].中國(guó)農(nóng)學(xué)通報(bào), 2001, 17(1):70-71.

GAO X, CAO C, SHAO Z, et al. Occurrence character‐istics and control countermeasures of garden plant dis‐eases and insect pests[J]. Chinese Agricultural ScienceBulletin, 2001, 17(1):70-71.

[7] 劉媛.基于深度學(xué)習(xí)的葡萄葉片病害識(shí)別方法研究[D].蘭州:甘肅農(nóng)業(yè)大學(xué), 2018.

LIU Y. Research? on? grape? leaf disease? identificationmethod? based? on? deep? learning[D]. Lanzhou: GansuAgricultural University, 2018.

[8] ZHANG? J, KONG? F, WU? J,? et? al. Automatic? imagesegmentation method for cotton leaves with disease un‐der natural environment[J]. Journal of Integrative Agri‐culture 2018, 17(8):1800-1814.

[9] ZHANG J, KONG F, ZHAI Z, et al. Robust image seg‐mentation method for cotton leaf under natural condi‐tions? based? on? immune? algorithm? and? PCNN? algo‐rithm[J]. International? Journal? of Pattern? Recognition& Artificial Intelligence, 2018, 32(5):3671-3678.

[10] HINTON G E, OSINDERO S, TEH Y W. A fast learn‐ing algorithm for deep belief nets[J]. Neural Computa‐tion, 2006, 18(7):1527-1554.

[11] KRIZHEVSKY A, SUTSKEVER I, HINTON G E. Im‐agenet? classification? with? deep? convolutional? neuralnetworks[J]. Advances in Neural Information process‐ing systems, 2012, 25:1097-1105.

[12] DARGAN S, KUMAR M, AYYAGARI M R, et al. Asurvey of deep learning and its applications: A new par‐adigm? to? machine? learning[J]. Archives? of Computa‐tional? Methods? in? Engineering, 2020, 27(4): 1071-1092.

[13] ZHANG? Q,? LIU Y,? GONG? C,? et? al. Applications? ofdeep learning for dense scenes analysis in agriculture:A review[J]. Sensors, 2020, 20(5): ID 1520.

[14] SANTOS? L,? SANTOS? F N,? OLIVEIRA P M,? et? al.Deep? learning? applications? in? agriculture: A short re‐view[C]// Iberian Robotics Conference. Berlin,German:Springer, Cham, 2019:139-151.

[15] BARBEDO J G A. Impact of dataset size and varietyon the effectiveness of deep learning and transfer learn‐ing? for plant? disease? classification[J]. Computers? andElectronics in Agriculture, 2018, 153:46-53.

[16] SINGH D, JAIN N, JAIN P, et al. PlantDoc: A datasetfor? visual? plant? disease? detection[C]// The 7th ACMIKDD? CoDS? and 25th? COMAD. New? York,? USA:ACM, 2020:249-253.

[17] LIU X, MIN W, MEI? S, et al. Plant disease recogni‐tion: A large-scale benchmark dataset and a visual re‐gion and loss reweighting approach[J]. IEEE Transac‐tions on Image Processing, 2021, 30:2003-2015.

[18] BARBEDO? J? G A,? KOENIGKAN? L V,? HALFELD-VIEIRA B A, et al. Annotated plant pathology databas‐ es for image-based detection and recognition of diseas‐ es[J]. IEEE Latin America Transactions, 2018, 16(6):1749-1757.

[19]張文靜.基于卷積神經(jīng)網(wǎng)絡(luò)的煙草病害識(shí)別與檢測(cè)[D].泰安:山東農(nóng)業(yè)大學(xué), 2021.

ZHANG W. Tobacco disease recognition and detection based on convolutional neural network[J]. Tai'an: Shan‐ dong Agricultural University, 2021.

[20] LECUN Y, BENGIO Y, HINTON G. Deep learning[J].Nature, 2015, 521:436-444.

[21] GIRSHICK R, DONAHUE J, DARRELL T, et al. Richfeature hierarchies for accurate object detection and se‐ mantic? segmentation[C]// The? IEEE? Conference? on Computer Vision and Pattern Recognition. Piscataway, New York, USA: IEEE, 2014:580-587.

[22] GIRSHICK R. Fast r-cnn[C]// The IEEE InternationalConference? on? Computer? Vision. Piscataway,? New York, USA: IEEE, 2015:1440-1448.

[23] REN S, HE K, GIRSHICK R, et al. Faster R-CNN : To‐wards real-time object detection with region proposal networks[J]. Advances in Neural Information Process‐ing Systems, 2015, 28:91-99.

[24] FUENTES A, YOON S, KIM S C, et al. A robust deep-learning-based detector for real-time tomato plant dis‐ eases? and? pests? recognition[J]. Sensors, 2017, 17(9):ID 2022.

[25]劉闐宇, 馮全.基于卷積神經(jīng)網(wǎng)絡(luò)的葡萄葉片檢測(cè)[J].西北大學(xué)學(xué)報(bào)(自然科學(xué)版), 2017, 47(4):505-512.

LIU T, FENG Q. Grape leaf detection based on convo‐lutional neural network[J]. Journal of Northwest Uni‐versity? (Natural? Science? Edition),? 2017,? 47(4):505-512.

[26]劉闐宇, 馮全, 楊森.基于卷積神經(jīng)網(wǎng)絡(luò)的葡萄葉片病害檢測(cè)方法[J].東北農(nóng)業(yè)大學(xué)學(xué)報(bào) , 2018, 49(3):73-83.

LIU T, FENG Q, YANG S. Detection of grape leaf dis‐ eases based on convolutional neural network[J]. Jour‐nal of Northeast Agricultural University, 2018, 49(3):73-83.

[27] OZGUVEN M M, ADEM K. Automatic detection andclassification? of leaf spot disease in? sugar beet using deep learning algorithms[J]. Physica A: Statistical Me‐chanics and its Applications, 2019, 535: ID 122537.

[28] BARI B? S, ISLAM M N, RASHID M, et al. A real-time? approach? of diagnosing? rice? leaf disease? using deep? learning-based? faster? R-CNN? framework[J]. PeerJ Computer Science, 2021, 7: ID e432.

[29] ZHOU G, ZHANG W, CHEN A, et al. Rapid detectionof rice disease based on FCM-KM and faster R-CNN fusion[J]. IEEE Access, 2019, 7:143190-143206.

[30] XIE X, MA Y, LIU B, et al. A deep-learning-based real-time? detector? for? grape? leaf diseases using? improvedconvolutional? neural? networks[J]. Frontiers? in? PlantScience, 2020, 11: ID 751.

[31] REDMON J, DIVVALA S, GIRSHICK R, et al. Youonly look once: Unified, real-time object detection[C]//The IEEE Conference on Computer Vision and PatternRecognition. Piscataway,? New? York,? USA: IEEE,2016:779-788.

[32] BHATT P V, SARANGI S, PAPPULA S. Detection ofdiseases and pests on images captured in uncontrolledconditions? from? tea? plantations[C]// Autonomous Airand Ground Sensing Systems for Agricultural Optimi‐zation? and? Phenotyping? IV. International? Society? forOptics? and? Photonics. Baltimore,? Maryland,? UnitedStates: SPIE, 2019: ID 1100808.

[33] MASKI P, THONDIYATH A. Plant disease detectionusing advanced deep learning algorithms: A case studyof papaya ring spot disease[C]//20216th InternationalConference on Image, Vision and Computing (ICIVC).Piscataway, New York, USA: IEEE, 2021:49-54.

[34]李昊, 劉海隆, 劉生龍.基于深度學(xué)習(xí)的柑橘病蟲(chóng)害動(dòng)態(tài)識(shí)別系統(tǒng)研發(fā)[J].中國(guó)農(nóng)機(jī)化學(xué)報(bào), 2021, 42(9):195-201, 208.

LI H, LIU H, LIU S. Development of dynamic recogni‐tion system of citrus pests and diseases based on deeplearning[J]. Journal of Chinese Agricultural Mechaniza‐tion, 2021, 42(9):195-201, 208.

[35] LIU W, ANGUELOV D, ERHAN D, et al. SSD : Sin‐gle? shot? multibox? detector[C]// European? Conferenceon Computer Vision. Berlin, German: Springer, Cham,2016:21-37.

[36] SUN H, XU H, LIU B, et al. MEAN-SSD : A novel re‐al-time detector for apple leaf diseases using improvedlight-weight? convolutional? neural? networks[J]. Com ‐puters? and? Electronics? in Agriculture, 2021, 189: ID106379.

[37] SUN? J, YANG Y,? HE? X,? et? al. Northern? maize? leafblight? detection? under? complex? field? environmentbased? on? deep? learning[J]. IEEE? Access, 2020, 8:33679-33688.

[38] SELVARAJ M G, VERGARA A, RUIZ H, et al. AI-powered banana? diseases? and pest? detection[J]. PlantMethods, 2019, 15(1):1-11.

[39] ZHOU? X, WANG? D,? KR?HENB?HL P. Objects? aspoints[J/OL]. arXiv preprint arXiv:1904.07850, 2019.

[40]夏雪, 孫琦鑫, 侍嘯, 等.基于輕量級(jí)無(wú)錨點(diǎn)深度卷積神經(jīng)網(wǎng)絡(luò)的樹(shù)上蘋(píng)果檢測(cè)模型[J].智慧農(nóng)業(yè)(中英文), 2020, 2(1):99-110.

XIA X, SUN Q, SHI X, et al. Apple detection modelbased? on? lightweight? anchor-free? deep? convolutionalneural? network[J]. Smart? Agriculture, 2020, 2(1):99-110.

[41] ALBATTAH W, NAWAZ M, JAVED A, et al. A? noveldeep? learning method? for? detection? and? classification of plant? diseases[J]. Complex & Intelligent? Systems, 2021:1-18.

[42] ZHANG K, WU Q, CHEN Y. Detecting soybean leafdisease? from? synthetic? image? using? multi-feature? fu‐sion? faster? R-CNN [J]. Computers? and? Electronics? in Agriculture, 2021, 183: ID 106064.

[43] HU G, WANG H, ZHANG Y, et al. Detection and se‐verity analysis of tea leaf blight based on deep learn‐ing[J]. Computers & Electrical Engineering, 2021, 90:ID 107023.

[44] ESER S. A deep learning based approach for the detec‐tion of diseases in pepper and potato leaves[J]. Anado‐lu? Journal? of? Agricultural? Sciences, 2021, 36(2):167-178.

[45] REHMAN Z, KHAN M A, AHMED F, et al. Recogniz‐ing apple leaf diseases using a novel parallel real ‐ time processing? framework? based? on? MASK? RCNN? and transfer? learning: An? application? for? smart? agricul‐ture[J]. IET? Image? Processing, 2021, 15(10):2157-2168.

[46] ANANDHAN? K,? SINGH? A? S. Detection? of? paddycrops diseases and early diagnosis using faster regional convolutional neural networks[C]//2021 International Conference? on? Advance? Computing? and? Innovative Technologies? in? Engineering (ICACITE). Piscataway, New York, USA: IEEE, 2021:898-902.

[47] KUMAR P. Research paper on sugarcane disease detec‐tion model[J]. Turkish Journal of Computer and Mathe‐matics? Education (TURCOMAT), 2021, 12(6):5167-5174.

[48]李鑫然, 李書(shū)琴, 劉斌.基于改進(jìn) Faster R_CNN 的蘋(píng)果葉片病害檢測(cè)方法[J].計(jì)算機(jī)工程 , 2021, 47(11):298-304.

LI X,? LI? S,? LIU? B. Detection? of apple? leaf diseases based on improved Faster R_CNN [J]. Computer Engi‐neering, 2021, 47(11):298-304.

[49] WANG Q, QI F, SUN M, et al. Identification of tomatodisease types and detection of infected areas based on deep convolutional neural networks and object detec‐tion? techniques[J]. Computational? Intelligence? and Neuroscience, 2019, 2019(2): ID 9142753.

[50]喬虹, 馮全, 張芮, 等.基于時(shí)序圖像跟蹤的葡萄葉片病害動(dòng)態(tài)監(jiān)測(cè)[J].農(nóng)業(yè)工程學(xué)報(bào) , 2018, 34(17):167-175.

QIAO H, FENG Q, ZHANG R, et al. Dynamic moni‐toring of grape leaf diseases based on sequential image tracking[J]. Transactions? of the? CSAE, 2018, 34(17):167-175.

[51] WANG J, YANG J, YU L, et al. DBA_SSD : A novelend-to-end object detection using deep attention mod‐ ule for helping? smart device with vegetable and fruitleaf plant? disease? detection[J]. Information, 2021, 12(11):474.

[52] SHILL? A,? RAHMAN? M? A. Plant? disease? detectionbased on YOLOv3 and YOLOv4[C]//2021 Internation‐al Conference on Automation, Control and Mechatron‐ics? for? Industry 4.0(ACMI). Piscataway, New York,USA: IEEE, 2021.

[53] WANG X, LIU J. Multiscale parallel algorithm for ear‐ly detection of tomato gray mold in a complex naturalenvironment[J]. Frontiers? in? Plant? Science, 2021, 12:ID 620273.

[54] HE X, FANG K, QIAO B, et al. Watermelon diseasedetection based on deep learning[J]. International Jour‐nal of Pattern Recognition? and Artificial Intelligence,2021, 35(5): ID 2152004.

[55] ATILA ? , U?AR M, AKYOL K, et al. Plant leaf dis‐ease? classification? using? EfficientNet? deep? learningmodel[J]. Ecological Informatics, 2021, 61: ID 101182.

[56] LIU J, WANG X. Tomato diseases and pests detectionbased on improved YOLOv3 convolutional neural net‐work[J]. Frontiers in Plant Science, 2020, 11: ID 898.

[57] MORBEKAR A, PARIHAR A, JADHAV R. Crop dis‐ease? detection? using? YOLO[C]//2020 InternationalConference for Emerging Technology (INCET). Piscat‐away, New York, USA: IEEE, 2020.

[58] PONNUSAMY V,? COUMARAN A,? SHUNMUGAMA S, et al. Smart glass: Real-time leaf disease detectionusing YOLO transfer learning[C]//2020 InternationalConference on Communication and Signal Processing(ICCSP). Piscataway,? New? York,? USA: IEEE, 2020:1150-1154.

[59] LIU? J,? WANG? X. Early? recognition? of tomato? grayleaf? spot? disease? based? on? MobileNetv2-YOLOv3model[J]. Plant Methods, 2020, 16: ID 83.

[60] JIANG P, CHEN Y, LIU B, et al. Real-time detectionof apple? leaf diseases? using? deep? learning? approachbased? on? improved? convolutional? neural? networks[J].IEEE Access, 2019, 7:59069-59080.

[61] TIAN Y, YANG G, WANG Z, et al. Detection of applelesions in orchards based on deep learning methods ofcyclegan? and? YOLOv3-dense[J]. Journal? of? Sensors,2019: ID 7630926.

[62] RAMCHARAN A, MCCLOSKEY P, BARANOWSKIK, et al. A mobile-based deep learning model for cassa‐va? disease? diagnosis[J]. Frontiers? in? Plant? Science,2019, 10: ID 272.

[63] KAVITHA LAKSHMI R,? SAVARIMUTHU N. DPD-DS for plant disease detection based on instance seg‐mentation[J]. Journal of Ambient Intelligence and Hu‐manized Computing, 2021:1-11.

[64] LIU C, ZHU H, GUO W, et al. EFDet: An efficient de‐tection? method? for? cucumber? disease? under? naturalcomplex? environments[J]. Computers? and? Electronicsin Agriculture, 2021, 189: ID 106378.

[65] DWIVEDI? R,? DEY? S,? CHAKRABORTY? C,? et? al.Grape? disease? detection? network based? on? multi-task learning and attention features[J]. IEEE Sensors Jour‐nal, 2021: ID 99.

[66] NIHAR F, KHANOM N N, HASSAN S S, et al. Plantdisease detection through the implementation of diver‐sified and modified neural network algorithms[J]. Jour‐nal of Engineering Advancements, 2021, 2(1):48-57.

[67] NAGASUBRAMANIAN K, JONES? S,? SINGH A K,et al. Plant disease identification using explainable 3D deep learning on hyperspectral images[J]. Plant Meth‐ods, 2019, 15(1): ID 98.

[68] KAWASAKI Y, UGA H, KAGIWADA S, et al. Basicstudy of automated diagnosis of viral plant diseases us‐ing? convolutional? neural? networks[C]// International Symposium? on? Visual? Computing. Berlin, German: Springer, Cham, 2015:638-645.

[69] SLADOJEVIC? S, ARSENOVIC M, ANDERLA A, etal. Deep? neural? networks? based? recognition? of plant diseases by leaf image classification[J]. Computational Intelligence? and? Neuroscience,? 2016,? 2016:? ID3289801.

[70] MOHANTY? S? P, HUGHES? D? P,? SALATH?? M. Us‐ing? deep? learning? for? image-based? plant? disease? de‐tection[J]. Frontiers in Plant Science, 2016, 7: ID 1419.

[71] RAMCHARAN? A,? BARANOWSKI,? MCCLOSKEYP, et al. Deep learning for image-based cassava disease detection[J]. Frontiers in Plant Science, 2017, 23(10):1-7.

[72]孫俊, 譚文軍, 毛罕平, 等.基于改進(jìn)卷積神經(jīng)網(wǎng)絡(luò)的多種植物葉片病害識(shí)別[J].農(nóng)業(yè)工程學(xué)報(bào) , 2017, 33(19):209-215.

SUN J, TAN W, MAO H, et al. Recognition of plant leaf diseases based on improved convolutional neural network [J]. Transactions of the CSAE, 2017, 33(19):209-215.

[73] LU? J,? HU? J,? ZHAO? G,? et? al. An? in-field? automaticwheat? disease? diagnosis? system[J]. Computers? and Electronics in Agriculture, 2017, 142:369-379.

[74] FERENTINOS? K? P. Deep? learning? models? for? plantdisease? detection? and? diagnosis[J]. Computers? and Electronics in Agriculture, 2018, 145:311-318.

[75]趙建敏, 李艷, 李琦, 等.基于卷積神經(jīng)網(wǎng)絡(luò)的馬鈴薯葉片病害識(shí)別系統(tǒng)[J].江蘇農(nóng)業(yè)科學(xué) , 2018, 46(24):251-255.

ZHAO? J,? LI? Y,? LI? Q,? et? al. Recognition? system? of potato leaf disease based on convolutional neural net‐ work[J]. Jiangsu Agricultural? Sciences, 2018, 46(24):251-255.

[76] XING S, LEE M, LEE K K. Citrus pests and diseasesrecognition model using weakly dense connected con‐ volution network[J]. Sensors, 2019, 19(14): ID 3195.

[77] 曾偉輝, 李淼, 李增, 等.基于高階殘差和參數(shù)共享反饋卷積神經(jīng)網(wǎng)絡(luò)的農(nóng)作物病害識(shí)別[J].電子學(xué)報(bào) ,2019, 47(9):1979-1986.

ZENG W, LI M, LI Z, et al. Recognition of crop diseas‐es based on high-order residual and parameter sharingfeedback convolutional Neural network[J]. Acta Elec‐tronicaSinica, 2019, 47(9):1979-1986.

[78] JI M, ZHANG L, WU Q. Automatic grape leaf diseas‐es? identification? via? UnitedModel? based? on? multipleconvolutional neural networks[J]. Information Process‐ing in Agriculture, 2020, 7(3):418-426.

[79]劉洋, 馮全, 王書(shū)志.基于輕量級(jí) CNN的植物病害識(shí)別方法及移動(dòng)端應(yīng)用[J].農(nóng)業(yè)工程學(xué)報(bào) , 2019, 35(17):194-204.

LIU Y, FENG Q, WANG S. Plant disease recognitionmethod based on lightweight CNN and its mobile ap‐plication[J]. Transactions of the CSAE, 2019, 35(17):194-204.

[80]王春山, 周冀, 吳華瑞, 等.改進(jìn) Multi-scale ResNet的蔬菜葉部病害識(shí)別[J].農(nóng)業(yè)工程學(xué)報(bào) , 2020, 36(20):209-217.

WANG? C,? ZHOU? J,? WU? H,? et? al. Improved? multi-scale ResNet for identification of vegetable leaf diseas‐es[J]. Transactions? of? the? CSAE, 2020, 36(20):209-217.

[81] SALEEM M H, POTGIETER J, ARIF K M. Plant dis‐ease? classification: A comparative? evaluation? of con‐volutional neural networks and deep learning optimiz‐ers[J]. Plants, 2020, 9(10): ID 1319.

[82] DE OCAMPO A L P, DADIOS E P. Mobile platformimplementation? of lightweight? neural? network? modelfor plant? disease? detection? and? recognition[C]//2018IEEE 10th? International? Conference? on? Humanoid,Nanotechnology,? Information? Technology,? Communi‐cation? and? Control,? Environment? and? Management(HNICEM). Piscataway, New York, USA: IEEE, 2018:1-4.

[83] DE LUNA R G, DADIOS E P, BANDALA A A. Au‐tomated? image? capturing? system? for? deep? learning-based tomato plant leaf disease detection and recogni‐tion[C]// TENCON 2018-2018 IEEE Region 10 Con‐ference. Piscataway,? New? York,? USA: IEEE, 2018:1414-1419.

[84] RASHID J, KHAN I, ALI? G,? et? al. Multi-level deeplearning model? for potato? leaf disease recognition[J].Electronics, 2021, 10(17): ID 2064.

[85] KIRATIRATANAPRUK K, TEMNIRANRAT P, KIT ‐VIMONRAT A, et al., Using deep learning techniquesto detect rice diseases from images of rice fields[C]//International? Conference? on? Industrial,? Engineeringand Other Applications of Applied Intelligent Systems.Berlin, German: Springer, 2020:225-237.

[86] JIANG Z, DONG Z, JIANG W, et al. Recognition ofrice? leaf diseases? and? wheat? leaf diseases? based? on multi-task? deep? transfer? learning[J]. Computers? and Electronics in Agriculture, 2021, 186: ID 106184.

[87] ABBAS A, JAIN S, GOUR M, et al. Tomato plant dis‐ease detection using transfer learning with C-GAN syn‐ thetic images[J]. Computers and Electronics in Agricul‐ture, 2021, 187: ID 106279.

[88] CHELLAPANDI B, VIJAYALAKSHMI M, CHOPRAS. Comparison? of pre-trained? models? using? transfer learning for detecting plant disease[C]//2021 Interna‐tional Conference on Computing, Communication, and Intelligent? Systems (ICCCIS). Piscataway, New York, USA: IEEE, 2021:383-387.

[89]樊湘鵬 , 許燕 , 周建平 , 等.基于遷移學(xué)習(xí)和改進(jìn)CNN 的葡萄葉部病害檢測(cè)系統(tǒng)[J].農(nóng)業(yè)工程學(xué)報(bào) , 2020, 37(6):151-159.

FAN X, XU Y, ZHOU J, et al. Grape leaf disease detec‐tion? system based? on? transfer? learning? and? improved CNN [J]. Transactions? of? the? CSAE, 2020, 37(6):151-159.

[90] JIANG F, LU Y, CHEN Y, et al. Image recognition offour rice leaf diseases based on deep learning and sup‐ port vector machine[J]. Computers and Electronics in Agriculture, 2020, 179: ID 105824.

[91] BARMAN? U,? CHOUDHURY R? D,? SAHU? D,? et? al.Comparison of convolution neural networks for smart‐ phone? image? based? real? time? classification? of citrus leaf disease[J]. Computers and Electronics in Agricul‐ture, 2020, 177: ID 105661.

[92] DANG L M, HASSAN S I, SUHYEON I, et al. UAVbased? wilt? detection? system? via? convolutional? neural networks[J]. Sustainable? Computing: Informatics? and Systems, 2020, 28: ID 100250.

[93] HOWLADER M R, HABIBA U, FAISAL R H, et al.Automatic? recognition? of? guava? leaf? diseases? using deep convolution neural network[C]//2019 Internation‐ al Conference on Electrical, Computer and Communi‐ cation? Engineering (ECCE). Piscataway,? New? York, USA: IEEE :2019.

[94] COULIBALY? S,? KAMSU-FOGUEM? B,? KAMISSO ‐KO D, et al. Deep neural networks with transfer learn‐ing? in? millet? crop? images[J]. Computers? in? Industry, 2019, 108:115-120.

[95] HU G, YANG X, ZHANG Y, et al. Identification of tealeaf diseases by using an improved deep convolutional neural network[J]. Sustainable Computing: Informatics and Systems, 2019, 24: ID 100353.

[96] SIBIYA M, SUMBWANYAMBE M. A computationalprocedure? for? the? recognition? and? classification? of maize leaf diseases out of healthy leaves using convo‐lutional neural networks[J]. AgriEngineering, 2019, 1(1):119-131.

[97]王艷玲, 張宏立, 劉慶飛, 等.基于遷移學(xué)習(xí)的番茄葉片病害圖像分類(lèi)[J].中國(guó)農(nóng)業(yè)大學(xué)學(xué)報(bào), 2019, 24(6):124-130.

WANG Y, ZHANG H, LIU Q, et al. Image classifica‐tion? of tomato? leaf diseases? based? on? transfer? learn‐ing[J]. Journal of China Agricultural University, 2019,24(6):124-130.

[98] SINGH U P, CHOUHAN S S, JAIN S, et al. Multilay‐er convolution neural network for the classification ofmango leaves infected by anthracnose disease[J]. IEEEAccess, 2019, 7:43721-43729.

[99] PICON A, ALVAREZ-GILA A, SEITZ M, et al. Deepconvolutional neural networks? for mobile capture de‐vice-based? crop? disease? classification? in? the? wild[J].Computers and Electronics in Agriculture, 2019, 161:280-290.

[100] ABDALLA A, CEN H, WAN L, et al. Fine-tuning con‐volutional neural network with transfer learning for se‐mantic segmentation of ground-level oilseed rape imag‐es? in? a? field? with? high? weed? pressure[J]. Computersand Electronics in Agriculture, 2019, 167: ID 105091.

[101] ATOLE R R, PARK D. A multiclass deep convolution‐al neural network? classifier? for? detection? of commonrice? plant? anomalies[J]. International? Journal? of Ad‐vancedComputer? Science? and Applications, 2018, 9(1):67-70.

[102] ZHANG X, QIAO Y, MENG F, et al. Identification ofmaize leaf diseases using improved deep convolutionalneural? networks[J]. IEEE? Access, 2018, 6:30370-30377.

[103] LIU B, ZHANG Y, HE D J, et al. Identification of ap‐ple? leaf diseases based? on? deep? convolutional? neuralnetworks[J]. Symmetry, 2018, 10(1): ID 11.

[104]張建華 , 孔繁濤 , 吳建寨 , 等.基于改進(jìn) VGG 卷積神經(jīng)網(wǎng)絡(luò)的棉花病害識(shí)別模型[J].中國(guó)農(nóng)業(yè)大學(xué)學(xué)報(bào),2018, 23(11):161-171.

ZHANG J, KONG F, WU J, et al. Cotton disease rec‐ognition model based on improved VGG convolutionalneural network[J]. Journal of China Agricultural Uni‐versity, 2018, 23(11):161-171.

[105] RANGARAJAN A K,? PURUSHOTHAMAN? R,? RA ‐MESH A. Tomato crop disease classification using pre-trained deep learning algorithm[J]. Procedia ComputerScience, 2018, 133:1040-1047.

[106] DE CHANT C, WIESNER-HANKS T, CHEN S, et al.Automated? identification? of? northern? leaf? blight-in‐fected? maize? plants? from? field? imagery? using? deeplearning[J]. Phytopathology, 2017, 107(11):1426-1432.

[107] LU Y, YI S, ZENG N, et al. Identification of rice dis‐eases? using? deep? convolutional? neural? networks[J].Neurocomputing, 2017, 267:378-384.

[108] OPPENHEIM D, SHANI G. Potato disease classifica‐tion using convolution neural networks[J]. Advances inAnimal Biosciences, 2017, 8(2):244-249.

[109]FUJITA E,? KAWASAKI Y,? UGA H,? et? al. Basic? in‐vestigation? on? a robust? and practical plant? diagnostic system[C]//201615th IEEE International Conference on Machine Learning and Applications (ICMLA). Pis‐cataway, New York, USA: IEEE, 2016:989-992.

[110] SRINIDHI V V, SAHAY A, DEEBA K. Plant patholo‐gy disease detection in apple leaves using deep convo‐lutional neural networks: Apple? leaves? disease? detec‐tion using EfficientNet and DenseNet[C]//20215th In‐ternational? Conference? on? Computing? Methodologies and Communication (ICCMC). Piscataway, New York,USA: IEEE, 2021:1119-1127.

[111] ZHOU J, LI J, WANG C, et al. A vegetable disease rec‐ognition model for complex background based on re‐gion proposal and progressive learning[J]. Computersand Electronics in Agriculture, 2021, 184: ID 106101.

[112] ZHANG P, YANG L, LI D. EfficientNet-B4-Ranger: Anovel method for greenhouse cucumber disease recog‐nition under natural complex environment[J]. Comput‐ers? and? Electronics? in? Agriculture, 2020, 176: ID105652.

Research Progress of Deep Learning in Detection and Recognition of Plant Leaf Diseases

SHAO Mingyue1 , ZHANG Jianhua1* , FENG Quan2 , CHAI Xiujuan1 ,ZHANG Ning1 , ZHANG Wenrong1

(1. Agricultural Information Institute of Chinese Academy of Agricultural Sciences/Key Laboratory of AgriculturalBig Data, Ministry of Agriculture and Rural Affairs, Beijing 10081, China;2. School of Mechanical and ElectricalEngineering, Gansu Agricultural University, Lanzhou 730070, China )

Abstract: Accurate detection and recognition of plant diseases is the key technology to early diagnosis and intelligent monitoring of plant diseases, and is the core of accurate control and information management of plant diseases and insect pests. Deep learn‐ing can overcome the disadvantages of traditional diagnosis methods and greatly improve the accuracy of diseases detection and recognition, and has attracted a lot of attention of researchers. This paper collected the main public plant diseases image data sets all over the world, and briefly introduced the basic information of each data set and their websites, which is convenient to download and use. And then, the application of deep learning in plant disease detection and recognition in recent years was sys‐tematically reviewed. Plant disease target detection is the premise of accurate classification and recognition of plant disease and evaluation of disease hazard level. It is also the key to accurately locate plant disease area and guide spray device of plant pro‐tection equipment to spray drug on target. Plant disease recognition refers to the processing, analysis and understanding of dis‐ ease images to identify different kinds of disease objects, which is the main basis for the timely and effective prevention and control of plant diseases. The research progress in early disease detection and recognition algorithm was expounded based on depth of learning research, as well as the advantages and existing problems of various algorithms were described. It can be seen from this review that the detection and recognition algorithm based on deep learning is superior to the traditional detection and recognition algorithm in all aspects. Based on the investigation of research results, it was pointed out that the illumination, shel‐tering, complex background, different disorders with similar symptoms, different changes of disease symptoms in different peri‐ods, and overlapping coexistence of multiple diseases were the main challenges for the detection and recognition of plant diseas‐ es. At the same time, the establishment of a large-scale and more complex data set that meets the specific research needs is also a difficulty that need to face together. And at further, we point out that the combination of the better performance of the neural network, large-scale data set and agriculture theoretical basis is a major trend of the development of the future. It is also pointed out that multimodal data can be used to identify early plant diseases, which is also one of the future development direction.???? Key words: plants; leaf disease; deep learning; disease detection; recognition; convolutional neural network; plant diseases im‐ age data set

猜你喜歡
識(shí)別卷積神經(jīng)網(wǎng)絡(luò)深度學(xué)習(xí)
基于深度卷積神經(jīng)網(wǎng)絡(luò)的物體識(shí)別算法
MOOC與翻轉(zhuǎn)課堂融合的深度學(xué)習(xí)場(chǎng)域建構(gòu)
大數(shù)據(jù)技術(shù)在反恐怖主義中的應(yīng)用展望
深度學(xué)習(xí)算法應(yīng)用于巖石圖像處理的可行性研究
基于深度卷積網(wǎng)絡(luò)的人臉年齡分析算法與實(shí)現(xiàn)
青島市中山公園園林樹(shù)木易混淆品種識(shí)別
基于卷積神經(jīng)網(wǎng)絡(luò)的樹(shù)葉識(shí)別的算法的研究
論犯罪危險(xiǎn)人格的識(shí)別
靖宇县| 东明县| 九龙城区| 平阴县| 定远县| 沁源县| 六枝特区| 瑞金市| 新竹市| 多伦县| 冷水江市| 嘉黎县| 鄂伦春自治旗| 姚安县| 丰台区| 湘潭市| 怀宁县| 乌鲁木齐县| 金寨县| 闽清县| 抚宁县| 西乡县| 南雄市| 安徽省| 河西区| 九龙县| 南阳市| 海淀区| 江安县| 凌云县| 泰顺县| 太白县| 香河县| 岑溪市| 岳普湖县| 永定县| 英德市| 揭阳市| 泸州市| 赫章县| 赤壁市|