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

?

基于改進(jìn)DenseNet和遷移學(xué)習(xí)的荷葉病蟲害識別模型

2023-07-28 02:28張國忠呂紫薇劉浩蓬劉婉茹龍長江黃成龍
關(guān)鍵詞:余弦荷葉卷積

張國忠,呂紫薇,劉浩蓬,劉婉茹,龍長江,黃成龍

基于改進(jìn)DenseNet和遷移學(xué)習(xí)的荷葉病蟲害識別模型

張國忠1,2,呂紫薇1,2,劉浩蓬1,2,劉婉茹1,2,龍長江1,2,黃成龍1,3※

(1. 華中農(nóng)業(yè)大學(xué)工學(xué)院,武漢 430070;2. 農(nóng)業(yè)農(nóng)村部長江中下游農(nóng)業(yè)裝備重點(diǎn)實(shí)驗(yàn)室,武漢 430070;3. 華中農(nóng)業(yè)大學(xué)作物遺傳改良國家重點(diǎn)實(shí)驗(yàn)室,武漢 430070)

病蟲害的發(fā)生將會嚴(yán)重影響蓮藕品質(zhì)與產(chǎn)量,開展病害診斷與識別對藕田病蟲害及時(shí)對癥對病診治、提升蓮藕生產(chǎn)質(zhì)量與經(jīng)濟(jì)效益具有重要意義。該研究以荷葉病蟲害高效、準(zhǔn)確識別為目標(biāo),提出了一種基于改進(jìn)DenseNet和遷移學(xué)習(xí)的荷葉病蟲害識別模型。采用分支結(jié)構(gòu)對模型的淺層特征提取模塊進(jìn)行改進(jìn),并在Dense Block與TransitionLayer中引入Squeeze and Excitation注意力機(jī)制模塊和銳化的余弦卷積,最后基于Plantvillage數(shù)據(jù)集進(jìn)行遷移學(xué)習(xí),實(shí)現(xiàn)了91.34%的識別準(zhǔn)確率。該研究實(shí)現(xiàn)了對荷葉腐敗病、病毒病、斜紋夜蛾、葉腐病、葉斑病的識別,并將改進(jìn)后的模型推廣應(yīng)用于基于無人機(jī)圖像的藕田病蟲害檢測,實(shí)現(xiàn)了病害分布可視化,可對蓮藕病蟲害的智能化防治提供有益指導(dǎo)。

模型;無人機(jī);病蟲害識別;荷葉;DenseNet;注意力機(jī)制;余弦相似度;遷移學(xué)習(xí)

0 引 言

蓮又名荷、蓮藕,是中國重要的水生蔬菜,其根莖、葉、花、種子均具有較高的食用與觀賞價(jià)值[1-3]。隨著生活水平的提升和飲食多樣化的發(fā)展,人們對綠色、高產(chǎn)、高品質(zhì)農(nóng)產(chǎn)品的需求日益提升。近年來,腐敗病、病毒病、斜紋夜蛾等病蟲害高發(fā),其中腐敗病可導(dǎo)致蓮藕減產(chǎn)40%以上,病蟲害的發(fā)生將嚴(yán)重影響蓮藕的品質(zhì)與產(chǎn)量[4-5]。蓮為大面積水田作物,其獨(dú)特的農(nóng)藝和生長特點(diǎn)使得田間穿行困難,不利于病蟲害的監(jiān)測與識別,加大了病蟲害的防控難度。目前,蓮的病蟲害鑒別主要依賴人工判別與病原檢測技術(shù)。人工判別主觀性強(qiáng),而病原檢測存在成本高、采樣難度大等不足之處,且多數(shù)種植戶缺乏專業(yè)的病蟲害防治知識,無法實(shí)現(xiàn)對蓮病蟲害的準(zhǔn)確判別。隨著人工智能技術(shù)與農(nóng)業(yè)的融合,基于機(jī)器學(xué)習(xí)與模式識別的圖像分類、目標(biāo)檢測等技術(shù)在植物病害檢測中得到廣泛應(yīng)用,對提升識別效率與降低病蟲害防控成本具有顯著成效。

葉片病斑的顏色和形態(tài)特征是病害識別的重要依據(jù)[6]。蓮雖為根莖類作物,但病蟲害特征主要集中于葉片,葉片感染不同病害時(shí)具有不同的顏色、形態(tài)特征[7-8],因此荷葉的生長狀態(tài)可以反映蓮病蟲害的發(fā)生情況[9]。傳統(tǒng)機(jī)器學(xué)習(xí)方法利用圖像處理技術(shù)提取葉片病斑的顏色、形狀、紋理等特征進(jìn)行植物病害識別,其分類性能依賴于病斑特征的提取效果,且算法具有專用性,泛化能力較差[10]。深度學(xué)習(xí)技術(shù)通過自動提取圖像特征可快速、準(zhǔn)確實(shí)現(xiàn)病蟲害識別,近年來被廣泛應(yīng)用于植物病蟲害識別領(lǐng)域,并取得了優(yōu)異成果。如黃建平等[11]提出一種基于神經(jīng)結(jié)構(gòu)搜索的植物葉片病害識別方法,根據(jù)特定數(shù)據(jù)集自動學(xué)習(xí)、搜索合適的深度神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu),實(shí)現(xiàn)了葉片病害的分類識別。XU等[12]通過引入自約束注意力增強(qiáng)分支改進(jìn)X-ception模型,增強(qiáng)了模型對局部特征的提取能力,實(shí)現(xiàn)了對花生病害的識別。以上研究的數(shù)據(jù)集均在背景單一、環(huán)境條件一致的實(shí)驗(yàn)室內(nèi)采集,而實(shí)際田間環(huán)境存在背景復(fù)雜、光照條件不一致、遮擋等干擾因素。田間環(huán)境下獲取的數(shù)據(jù)集比實(shí)驗(yàn)室條件下的數(shù)據(jù)集其模型識別準(zhǔn)確率低約30%~40%[13]。提高應(yīng)用于田間病蟲害識別的模型性能有利于病蟲害自動識別技術(shù)的推廣與應(yīng)用。復(fù)雜背景中存在與疾病特征相似的元素導(dǎo)致病害識別難度增加,為解決該問題,丁永軍等[14]以VGG-16模型為基礎(chǔ)構(gòu)建卷積膠囊網(wǎng)絡(luò)對百合病害進(jìn)行識別,提升了模型的抗噪能力,更適用于實(shí)際生產(chǎn)環(huán)境。孫俊等[15]通過多尺度特征融合、嵌入坐標(biāo)注意力機(jī)制等設(shè)計(jì)手段改善了MobileNet-V2中存在的感興趣區(qū)域分散以及特征提取尺度單一等問題。蘇仕芳等[16]采用TensorFlow_Lite將模型部署到智能手機(jī)移動終端,實(shí)現(xiàn)了移動端對葡萄葉片病害的快捷、智能化診斷。ZHOU等[17]提出一種基于區(qū)域提議和漸進(jìn)學(xué)習(xí)的PRPNet模型,以98.26%的平均識別準(zhǔn)確率實(shí)現(xiàn)了復(fù)雜背景下蔬菜病害的識別。彭紅星等[18]在MobileNet V2的反向殘差模塊中嵌入坐標(biāo)注意力機(jī)制,并采用深度可分離卷積設(shè)計(jì)了雙分支特征融合模塊,能快速實(shí)現(xiàn)葡萄病蟲害的識別。徐艷蕾等[19]采用雙分支結(jié)構(gòu)對圖像的局部特征與全局特征進(jìn)行融合,在自然環(huán)境下蘋果葉片病害的識別精度達(dá)到80%。目前,基于CNN對作物病害自動識別的研究主要集中于番茄、蘋果、玉米、水稻、黃瓜、小麥等[12,20-22],以水生蔬菜病蟲害智能識別為例,現(xiàn)有研究主要集中于芋頭病害的識別,如陳林琳等[23]和王佳[24]基于機(jī)器學(xué)習(xí)方法分別實(shí)現(xiàn)了對芋頭和香芋病害的識別,仍缺乏蓮藕病蟲害智能化識別的相關(guān)研究。

荷葉病蟲害特征具有一定的相似性,輕癥病害特征不顯著,受復(fù)雜水田環(huán)境與光照條件等因素的影響增加了病蟲害識別難度,基于深度學(xué)習(xí)技術(shù)進(jìn)行荷葉病蟲害的自動識別可以提升藕田病蟲害的防控效率。本文基于DenseNet[25]模型結(jié)構(gòu)進(jìn)行改進(jìn),通過引入多層小卷積提升感受野,提升淺層網(wǎng)絡(luò)的特征表達(dá)能力;引入SE(squeeze and excitation)注意力機(jī)制模塊[26]提升對病害特征的學(xué)習(xí)能力;采用銳化的余弦卷積(sharpened cosine similarity convolution,SCS)[27]代替?zhèn)鹘y(tǒng)卷積以避免梯度爆炸,減小方差;通過遷移學(xué)習(xí)模型在Plantvillage數(shù)據(jù)集上的知識提升模型的收斂速度避免過擬合。經(jīng)過以上改進(jìn),期冀提升模型對細(xì)節(jié)特征的提取能力和抗干擾能力,從而實(shí)現(xiàn)對荷葉病蟲害的快速識別與檢測。

1 材料與方法

1.1 數(shù)據(jù)采集

荷葉主要病蟲害與特征描述如表1所示。為提升模型對不同來源數(shù)據(jù)的泛化能力,本研究于2022年6—9月在國家種質(zhì)武漢水生蔬菜資源圃、武漢市江夏區(qū)武當(dāng)村,采用多種設(shè)備進(jìn)行數(shù)據(jù)采集。采用大疆御MAVIC2無人機(jī)采集藕田內(nèi)部圖像,為保留葉片病害特征結(jié)合蓮藕種植農(nóng)藝特點(diǎn),將飛行高度設(shè)置為距離荷葉冠層3 m;采用2 000萬像素的佳能EOS6D數(shù)碼相機(jī),以及OnePlus 6T、小米8智能手機(jī)距離荷葉冠層1~2 m在田邊進(jìn)行數(shù)據(jù)采集,共獲取6類荷葉病蟲害圖片2 295張,其中腐敗病322張、病毒病327張、葉斑病382張、葉腐病490張、斜紋夜蛾302張、健康葉片472張,數(shù)據(jù)集示例如圖1所示。

表1 荷葉主要病蟲害及其特征

圖1 荷葉病蟲害圖像示例

1.2 數(shù)據(jù)預(yù)處理

由于采集的原始圖像來自不同設(shè)備,圖像分辨率大小不一致且包含過多冗余信息,對原始數(shù)據(jù)進(jìn)行裁剪并將大小統(tǒng)一調(diào)整為300×300像素。為提升模型泛化能力,避免測試集內(nèi)數(shù)據(jù)信息帶入訓(xùn)練集,將數(shù)據(jù)集按6∶4的比例隨機(jī)劃分為訓(xùn)練集和測試集。

數(shù)據(jù)增強(qiáng)按照數(shù)據(jù)存儲方式可分為離線數(shù)據(jù)增強(qiáng)和在線數(shù)據(jù)增強(qiáng)兩種。離線數(shù)據(jù)增強(qiáng)將數(shù)據(jù)集進(jìn)行擴(kuò)增后存儲在磁盤中,大量數(shù)據(jù)擴(kuò)增會增加存儲成本;在線數(shù)據(jù)增強(qiáng)又稱動態(tài)數(shù)據(jù)增強(qiáng),其使用生成器在訓(xùn)練期間對即將喂入網(wǎng)絡(luò)的數(shù)據(jù)進(jìn)行動態(tài)變換,擴(kuò)充后的數(shù)據(jù)直接喂入網(wǎng)絡(luò)而不進(jìn)行存儲,降低了存儲成本[28]。本文采用在線數(shù)據(jù)增強(qiáng)的方法對讀取的數(shù)據(jù)進(jìn)行數(shù)據(jù)變換。由于數(shù)據(jù)增強(qiáng)中包含隨機(jī)因子,因此每輪次讀取的數(shù)據(jù)均不同從而達(dá)到數(shù)據(jù)增強(qiáng)效果。數(shù)據(jù)集大小影響模型的泛化能力,相同數(shù)據(jù)集采用不同的數(shù)據(jù)增強(qiáng)策略對模型性能提升效果不同[29]。采用縮放(resize)、隨機(jī)裁剪(random resized crop,RRC)、隨機(jī)旋轉(zhuǎn)(random rotation,RR)、隨機(jī)水平翻轉(zhuǎn)(random horizontal flip,RHF)、隨機(jī)調(diào)整銳度(random adjust sharpness,RAS)共5種數(shù)據(jù)增強(qiáng)方法進(jìn)行組合,以探究不同數(shù)據(jù)增強(qiáng)方式對荷葉病害數(shù)據(jù)集的模型性能影響。

1.3 荷葉病害分類模型比較

深度學(xué)習(xí)模型通過構(gòu)建深層網(wǎng)絡(luò)結(jié)構(gòu)實(shí)現(xiàn)端到端的自動學(xué)習(xí)[30]。AlexNet[31]、VGG[32]、ResNet[33]、ResNeXt[34]、DenseNet等深度學(xué)習(xí)模型廣泛應(yīng)用于圖像識別領(lǐng)域。為選擇性能最優(yōu)模型進(jìn)行改進(jìn),本文采用相同的訓(xùn)練策略進(jìn)行試驗(yàn)。為適應(yīng)模型大小將圖片統(tǒng)一縮放至224×224像素并進(jìn)行歸一化,受硬件條件約束批處理大?。˙atch-size)設(shè)置為16,模型迭代次數(shù)共100 epoch,初始學(xué)習(xí)率設(shè)為0.01,每經(jīng)過30 epoch,學(xué)習(xí)率衰減為原來的0.1,所采用隨機(jī)梯度下降(stochastic gradient descent,SGD)優(yōu)化器。在自建的荷葉病蟲害數(shù)據(jù)集上對VGG-16、AlexNet、ResNet50、ResnNeXt50、DenseNet121模型進(jìn)行訓(xùn)練,模型準(zhǔn)確率分別為66.23%、75.55%、76.32%、78.18%、81.47%。模型準(zhǔn)確率與損失曲線如圖2所示,VGG-16的損失值最快收斂到穩(wěn)定狀態(tài),但模型準(zhǔn)確率較低,DenseNet121準(zhǔn)確率最高且損失值在第65 epoch后逐漸到達(dá)穩(wěn)定狀態(tài),故本文基于DenseNet121(下文簡述為DenseNet)模型進(jìn)行改進(jìn)。

圖2 模型性能比較

1.4 DenseNet模型結(jié)構(gòu)改進(jìn)

DenseNet通過建立層與層之間的密集連接將前面所有層的特征圖作為該層的輸入,在很大程度上解決了深度卷積網(wǎng)絡(luò)梯度消失的問題,加強(qiáng)了特征的傳遞和重用。密集連接的正則化效應(yīng)在一定程度上可以降低小數(shù)據(jù)集的過擬合問題。此外,由于不需要重新學(xué)習(xí)冗余的特征映射,減少了網(wǎng)絡(luò)的參數(shù)量。

1.4.1 Stem模塊

淺層網(wǎng)絡(luò)可以實(shí)現(xiàn)對圖像邊緣、細(xì)節(jié)紋理等信息的提取。如圖3a所示,Stem-A為DenseNet的Stem模塊,由步長為2的7×7卷積層和步長為2的3×3最大池化組成。如圖3b所示,Stem-B在1×1卷積前添加步長為2的最大池化作為新添加路徑,可減少特征丟失,提升模型性能[35]。如圖3c所示,參考VGG網(wǎng)絡(luò)采用3個(gè)3×3的卷積堆疊代替7×7卷積可以獲得更大感受野,提升特征的表達(dá)能力。為提升淺層網(wǎng)絡(luò)對圖像的邊緣、紋理等特征的提取能力,采用步長為2的3×3卷積進(jìn)行快速降維,然后通過分支結(jié)構(gòu)將3×3卷積與最大池化組合,再利用1×1的卷積進(jìn)行降維,以較少的計(jì)算開銷提升模型的準(zhǔn)確率[36],如圖3d所示。

1.4.2 銳化的余弦卷積

傳統(tǒng)的多層神經(jīng)網(wǎng)絡(luò)使用上一層的輸出向量與輸入權(quán)重向量的點(diǎn)積作為激活函數(shù)的輸入,神經(jīng)元的方差較大使模型對輸入分布的變化敏感,泛化效果差,并加劇了內(nèi)部協(xié)變量的移位,從而降低了訓(xùn)練速度。將余弦相似度與卷積網(wǎng)絡(luò)相結(jié)合,通過余弦歸一化[37](cosine normalization)(如式(1)所示)將激活函數(shù)限制在-1~1,可避免梯度爆炸,減小方差。兩個(gè)完全不同的輸入向量可能具有相似的余弦相似度,且當(dāng)輸入接近0時(shí),余弦相似度在數(shù)值上會變得不穩(wěn)定。因此,本文采用銳化的余弦卷積作為傳統(tǒng)卷積的一種替代方案。銳化的余弦相似度(sharpened cosine similarity,SharpCosSim)如式(2)所示,通過添加兩個(gè)超參數(shù)和進(jìn)一步改進(jìn)余弦相似度。余弦相似度的峰值為1,利用參數(shù)以把余弦相似度提高到指數(shù)次方,在輸入向量中添加另一個(gè)參數(shù)即預(yù)期本底噪聲的大小,將防止匹配項(xiàng)對噪聲進(jìn)行注冊,將銳化的余弦相似度與卷積相結(jié)合,即引入銳化的余弦卷積(SCS)具有更好的特征提取能力。

注:conv代表卷積;1×1、3×3、7×7為卷積核尺寸;2×2為池化核尺寸;s為步長;p為邊界擴(kuò)充;maxpool為最大池化。下同。

1.4.3 Squeeze and Excitation注意力機(jī)制模塊

SE注意力機(jī)制在注意力模式下可以確定不同通道之間的權(quán)重關(guān)系,提升對感興趣通道的注意力,其結(jié)構(gòu)如圖4所示,SE注意力機(jī)制模塊通過全局平均池化對特征圖進(jìn)行壓縮,實(shí)現(xiàn)全局上下文信息的融合,然后通過全連接層→線性整流函數(shù)(Rectified linear units,Relu)→全連接層→Sigmoid激活函數(shù)(Sigmoid)生成每個(gè)特征通道的權(quán)重,最后將原始輸入的特征圖與獲得的通道權(quán)重進(jìn)行點(diǎn)乘,在通道維度上實(shí)現(xiàn)對特征的標(biāo)定。

注:H、W、C分別表示圖像的高、寬、通道數(shù),F(xiàn)C表示全連接層。下同。

改進(jìn)前后模型結(jié)構(gòu)對比如表2所示,DenseNet由Stem模塊、4個(gè)Dense Block和3個(gè)Transition Layer組成。本文將DenseNet的Stem模塊(Stem-A如圖3a所示)改為具有更大感受野的Stem-D(如圖3d所示),通過分支結(jié)構(gòu)以較小的計(jì)算開銷提升模型對淺層特征的提取能力;將Dense Block與Transition Layer內(nèi)的卷積更換為銳化的余弦卷積可以提升模型的泛化能力和對邊緣等特征的提取能力;在Dense Block每個(gè)Denselayer的1×1卷積前以及Transition Layer內(nèi)引入SE注意力機(jī)制模塊以提升模型對通道特征的敏感性。

表2 模型結(jié)構(gòu)

注:SE block為SE注意力機(jī)制模塊;SCS為銳化的余弦卷積。

Note: SE block represents SE attention mechanism block; SCS represents sharpened cosine similarity convolution.

1.5 AdaMax優(yōu)化器

采用隨機(jī)梯度下降算法(stochastic gradient descent,SGD)時(shí)模型的收斂速度快,但選擇合適的學(xué)習(xí)率較困難。學(xué)習(xí)率過小使模型收斂緩慢。學(xué)習(xí)率過高時(shí)模型收斂波動較大,容易收斂到局部最優(yōu)解,且容易被困在鞍點(diǎn)。AdaMax是Adam(adaptive moment estimation)基于無窮范數(shù)的一種變體[38]。在Adam中,單個(gè)權(quán)重的更新規(guī)則是將其梯度與當(dāng)前和過去梯度的2范數(shù)成反比例縮放。AdaMax是在Adam基礎(chǔ)上將基于2范數(shù)的更新規(guī)則泛化到基于L范數(shù)的更新規(guī)則。AdaMax會因范數(shù)較大在數(shù)值上變得不穩(wěn)定,但當(dāng)范數(shù)→∞時(shí),如式(3)所示,時(shí)刻的指數(shù)加權(quán)無窮范數(shù)u基于∞范數(shù)更新規(guī)則會得到一個(gè)簡單、穩(wěn)定的算法,學(xué)習(xí)率的邊界范圍更簡單,且當(dāng)初始值00時(shí),不需要糾正初始化偏差。

1.6 試驗(yàn)設(shè)置與模型性能評價(jià)指標(biāo)

所有試驗(yàn)均在Ubuntu 20.04 LTS 64位系統(tǒng)環(huán)境下運(yùn)行,模型采用深度學(xué)習(xí)開源框架Pytorch1.10.1和Python3.8.0搭建。計(jì)算機(jī)搭載的處理器為Intel Core i7-10700K@3.80GHz八核,運(yùn)行內(nèi)存32 GB,GPU為GTX 3070Ti 8G。輸入模型的圖像尺寸大小為224×224像素,受硬件條件的約束,批處理大?。˙atch-size)設(shè)置為16,模型迭代次數(shù)共100 epoch,初始學(xué)習(xí)率設(shè)置為0.01,采用余弦退火學(xué)習(xí)率更新策略。

本文采用準(zhǔn)確率(Accuracy)、精確率(Precision)、召回率(Recall)、準(zhǔn)確率與召回率加權(quán)調(diào)和平均值—1值(1-score)來衡量模型的分類性能,以上評價(jià)指標(biāo)越高模型性能越好。

2 試驗(yàn)結(jié)果與分析

2.1 不同動態(tài)數(shù)據(jù)增強(qiáng)方式性能評估

采用SGD優(yōu)化器對不同數(shù)據(jù)增強(qiáng)方式進(jìn)行試驗(yàn),結(jié)果如表3所示。方案2將圖片先縮放至256×256像素再隨機(jī)裁剪至224×224像素比直接縮放至224×224像素準(zhǔn)確率提升1.43個(gè)百分點(diǎn),圖片縮放尺寸變化較大時(shí)可能丟失較多細(xì)節(jié)信息而降低模型識別效果;在方案2的基礎(chǔ)上引入隨機(jī)水平翻轉(zhuǎn)、隨機(jī)旋轉(zhuǎn)、隨機(jī)調(diào)整銳度等不同數(shù)據(jù)增強(qiáng)方法后模型準(zhǔn)確率均得到提升,其中隨機(jī)調(diào)整銳度效果最佳。結(jié)果表明數(shù)據(jù)增強(qiáng)方案8效果最佳,準(zhǔn)確率為85.01%。故選用方案8作為數(shù)據(jù)增強(qiáng)方案可減少圖片過度縮放導(dǎo)致的細(xì)節(jié)信息丟失,同時(shí)對圖片進(jìn)行水平翻轉(zhuǎn)和隨機(jī)調(diào)整銳度以模擬不同的拍攝角度和質(zhì)量可提升模型泛化能力。為驗(yàn)證AdaMax優(yōu)化器的優(yōu)越性,本研究基于數(shù)據(jù)增強(qiáng)方案8將SGD優(yōu)化器替換為AdaMax優(yōu)化器后模型的準(zhǔn)確率為88.05%,較方案8提升了3.04個(gè)百分點(diǎn),因此后續(xù)試驗(yàn)均采用該優(yōu)化策略。

表3 不同數(shù)據(jù)增強(qiáng)方式下DenseNet準(zhǔn)確率

注:Resize(224,224):將原圖像縮放至224×224像素;Resize256:對圖像的短邊按原圖長寬比統(tǒng)一縮放至256像素;RRC:隨機(jī)裁剪圖像并將其調(diào)整至224×224像素;RHF:以給定概率對圖片進(jìn)行隨機(jī)水平翻轉(zhuǎn);RR30:將圖片隨機(jī)旋轉(zhuǎn)30°;RAS:以給定概率隨機(jī)調(diào)整圖像銳度。

Note: Resize(224,224): Resize the input image to 224×224; Resize256: Uniformly resize the short edges of the image to 256 according to the aspect ratio of the original image; RRC: Random resized crop a random portion of image and resize it to 224×224; RHF: Random horizontal flip the given image randomly with a given probability; RR30: Random rotate the image by 30°; RAS: Random adjust the sharpness of the image randomly with a given probability.

2.2 荷葉病蟲害識別最優(yōu)Stem模塊

為提升模型淺層網(wǎng)絡(luò)的特征提取能力,在DenseNet中插入不同的Stem模塊進(jìn)行消融試驗(yàn),結(jié)果如表4所示。引入Stem-B模塊較原模型(DenseNet-Stem-A)的計(jì)算量和參數(shù)量變化較小,但模型準(zhǔn)確率僅提升0.36個(gè)百分點(diǎn)。Stem-C擴(kuò)大了模型的感受野,準(zhǔn)確率提升1.10個(gè)百分點(diǎn),但參數(shù)量提升了0.07 M,計(jì)算量提升0.84 G。引入Stem-D比原模型的準(zhǔn)確率提高1.10個(gè)百分點(diǎn),參數(shù)量和計(jì)算量僅分別增加0.02 M、0.02 G,該模塊在不過多增加計(jì)算量和參數(shù)量的基礎(chǔ)上提升了模型的特征提取能力。

2.3 DenseNet優(yōu)化策略消融試驗(yàn)

1)DenseNet模型引入不同模塊的消融試驗(yàn)結(jié)果如表 5所示。在DenseNet中引入銳化的余弦卷積(SCS)比表4原模型的準(zhǔn)確率和1值分別提升0.99、1.18個(gè)百分點(diǎn),驗(yàn)證了銳化的余弦卷積能夠提升模型對特征的提取能力,引入SE注意力機(jī)制模塊后模型性能稍有提升。DenseNet-SE-SCS-Stem-D將SCS、SE注意力機(jī)制模塊、Stem-D模塊進(jìn)行組合,模型的準(zhǔn)確率、精確率、召回率、F1值分別提升至90.13%、91.33%、89.82%、90.05%。

遷移學(xué)習(xí)可以減小模型訓(xùn)練的代價(jià),讓卷積神經(jīng)網(wǎng)絡(luò)更適應(yīng)小樣本數(shù)據(jù),提升模型的泛化能力?;赑lantvillage數(shù)據(jù)集對DenseNet-SE-SCS-Stem-D進(jìn)行遷學(xué)習(xí),學(xué)習(xí)率設(shè)置為0.001,進(jìn)行遷移學(xué)習(xí)后模型的準(zhǔn)確率、精確率、召回率、1值分別為91.34%、91.43%、91.23%、91.29%,進(jìn)一步提升了模型的性能,其中準(zhǔn)確率較優(yōu)化前表3方案1提升9.87個(gè)百分點(diǎn)。

表4 引入不同Stem模塊的模型性能比較

表5 不同改進(jìn)方法模型性能對比

2)為驗(yàn)證遷移學(xué)習(xí)后模型的性能,測試集的混淆矩陣如圖5所示,通過比較混淆矩陣中主對角線上預(yù)測正確的樣本量可知,改進(jìn)后的模型提升了對各類病害的識別能力,模型對健康荷葉、葉斑病、病毒病識別準(zhǔn)確率分別提升了11.70、17.11、20.77個(gè)百分點(diǎn)。但模型將葉斑病錯(cuò)誤識別為病毒病的數(shù)量提升,推測是由于模型對細(xì)節(jié)紋理信息提取能力的增強(qiáng),使得改進(jìn)后模型將部分輕度卷曲且光照不均的葉斑病錯(cuò)誤識別為病毒病。

注:0為腐敗病;1為健康葉片;2為斜紋夜蛾;3為葉斑?。?為葉腐病,5為病毒病。下同。

3)圖6為模型改進(jìn)前后病害預(yù)測結(jié)果示例和類激活特征圖可視化結(jié)果,圖中9個(gè)特征圖依次為表2模型結(jié)構(gòu)圖中8個(gè)Layers和模型最后一個(gè)Batch Normalization(BN)層的類激活特征圖。改進(jìn)后模型的淺層卷積對葉片細(xì)節(jié)紋理信息的提取能力提升,能更好地實(shí)現(xiàn)對細(xì)小病斑的識別。改進(jìn)后模型能更加準(zhǔn)確地判斷病斑在葉片上的發(fā)生區(qū)域,如對腐敗病更加關(guān)注葉片邊緣特征,對葉斑病、斜紋夜蛾、葉腐病更加關(guān)注葉片內(nèi)部病斑區(qū)域,對幼嫩健康葉片與病毒病則更關(guān)注全局信息。

注:pred_label代表圖片的預(yù)測類別標(biāo)簽,True label表示圖片的真實(shí)類別標(biāo)簽,pred_score表示預(yù)測為該類別的概率,pred_class表示預(yù)測標(biāo)簽對應(yīng)的病害類別。

2.4 基于無人機(jī)圖像的藕田病蟲害識別

采用大疆御MAVIC2無人機(jī)將飛行高度設(shè)置為6.5 m拍攝藕田圖片,并對荷葉圖片的葉片區(qū)域進(jìn)行標(biāo)定獲取葉片的坐標(biāo)信息。批量讀取葉片的坐標(biāo)信息并對圖片進(jìn)行裁剪,將裁剪后的圖片送入遷移學(xué)習(xí)后的DenseNet-SE-SCS-Stem-D模型進(jìn)行推理預(yù)測,并根據(jù)預(yù)測結(jié)果在無人機(jī)圖片的相應(yīng)坐標(biāo)區(qū)域添加不同顏色掩膜以生成藕田荷葉病蟲害分布圖,如圖7所示。

通過荷葉病蟲害分布圖可以直觀了解到藕田內(nèi)各類病害的分布情況與嚴(yán)重程度,實(shí)現(xiàn)藕田內(nèi)荷葉病蟲害的評估。根據(jù)荷葉標(biāo)定數(shù)據(jù)與推理結(jié)果可獲取區(qū)域內(nèi)各類病害的發(fā)生比例,其中病毒病為1.86%,腐敗病為19.25%,健康為32.30%,葉斑病為27.33%,斜紋夜蛾為9.94%,葉腐病9.32%。在后期研究中可將目標(biāo)檢測網(wǎng)絡(luò)與荷葉病蟲害分類識別網(wǎng)絡(luò)相結(jié)合,先對荷葉進(jìn)行自動提取,再利用分類網(wǎng)絡(luò)進(jìn)行識別以降低復(fù)雜田間背景對不同病害識別的影響,以實(shí)現(xiàn)大面積藕田的病蟲害識別與動態(tài)監(jiān)測。

圖7 藕田荷葉病蟲害分布圖

3 結(jié) 論

本文提出了一種基于改進(jìn)DenseNet和遷移學(xué)習(xí)的荷葉病蟲害識別模型,并將其應(yīng)用于基于無人機(jī)圖像的藕田荷葉病蟲害檢測,得出以下結(jié)論:

1)采用動態(tài)數(shù)據(jù)增強(qiáng),對不同數(shù)據(jù)增強(qiáng)組合方法進(jìn)行比較,結(jié)果表明對圖片先縮放再隨機(jī)裁剪比直接縮放至相同大小模型的準(zhǔn)確率更高,大幅度進(jìn)行圖片縮放可能會導(dǎo)致圖片部分細(xì)節(jié)信息丟失,影響模型識別效果。隨機(jī)縮放、隨機(jī)裁剪、隨機(jī)旋轉(zhuǎn)、隨機(jī)調(diào)整銳度的數(shù)據(jù)增強(qiáng)方式相組合更具優(yōu)勢。

2)在DenseNet模型的Dense Block模塊和Transition模塊中引入銳化的余弦卷積,驗(yàn)證了銳化的余弦卷積提升模型性能的有效性;通過優(yōu)化模型淺層特征提取模塊,引入Squeeze and Excitation注意力機(jī)制模塊提升了模型對細(xì)節(jié)特征的提取能力和對病害的識別能力。荷葉病蟲害的識別準(zhǔn)確率達(dá)到91.34%。

3)將改進(jìn)后的分類模型應(yīng)用于無人機(jī)圖片進(jìn)行藕田內(nèi)荷葉病蟲害識別。對荷葉所在區(qū)域標(biāo)定裁剪后進(jìn)行分類識別,并根據(jù)模型預(yù)測結(jié)果生成不同掩膜添加在無人機(jī)圖片上生成藕田荷葉病蟲害分布圖,實(shí)現(xiàn)了藕田荷葉病蟲害的識別,解決了藕田內(nèi)部病蟲害監(jiān)測困難的問題,為藕田的病蟲害識別與動態(tài)監(jiān)測奠定基礎(chǔ)。

本文實(shí)現(xiàn)了對荷葉病蟲害的自動分類識別,為高效、準(zhǔn)確、動態(tài)監(jiān)測藕田病蟲害提供了新方法,為藕田植保無人機(jī)的變量施藥和飛行路徑規(guī)劃提供了信息支撐。

[1] 李峰,周雄祥,柯衛(wèi)東,等. 湖北省蓮產(chǎn)業(yè)發(fā)展調(diào)研報(bào)告[J]. 湖北農(nóng)業(yè)科學(xué),2020,59(23):101-106. LI Feng, ZHOU Xiongxiang, KE Weidong, et al. Research report on lotus industry development in Hubei province[J]. Hubei Agricultural Sciences, 2020, 59(23): 101-106. (in Chinese with English abstract)

[2] 程婷婷,惠小涵,尚欣欣,等. 10個(gè)產(chǎn)地蓮藕營養(yǎng)成分分析與品質(zhì)綜合評價(jià)[J]. 食品工業(yè)科技,2021,42(8):320-325. CHENG Tingting, XI Xiaohan, SHANG Xinxin, et a1. Nutrient composition analysis and quality comprehensive evaluation of lotus root in 10 producing areas[J]. Science and Technology of Food Industry, 2021, 42(8): 320-325. (in Chinese with English abstract)

[3] 賈彩鳳,時(shí)江蕾,周敬娜,等. 白洋淀地區(qū)水生藥用植物資源開發(fā)利用研究進(jìn)展[J]. 醫(yī)學(xué)研究與教育,2019,36(1):28-33. JIA Caifeng, SHI Jianglei, ZHOU Jingna, et al. Advances on exploitation of aquatic medicinal plant resources in Baiyangdian Lake[J]. Medical Research and Education, 2019, 36(1): 28-33. (in Chinese with English abstract)

[4] 魏林,梁懷志. 蓮藕病蟲草害識別與綜合防治[M]. 北京:中國農(nóng)業(yè)科學(xué)技術(shù)出版社,2013.

[5] 唐光美,楊馨語,張小樂,等. 蓮藕腐敗病致病菌的鑒定與防治[J]. 云南大學(xué)學(xué)報(bào)(自然科學(xué)版),2022,44(6):1280-1287. TANG Guangmei, YANG Xinyu, ZHANG Xiaole, et al. Identification and control of pathogenic fungi of lotus rhizome rot[J]. Journal of Yunnan University (Natural Sciences Edition), 2022, 44(6): 1280-1287. (in Chinese with English abstract)

[6] 于明,李若曦,閻剛,等. 基于顏色掩膜網(wǎng)絡(luò)和自注意力機(jī)制的葉片病害識別方法[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2022,53(8):337-344. YU Ming, LI Ruoxi, YAN Gang, et al. Crop diseases recognition method via fusion color mask and self-attention mechanism[J]. Transactions of the Chinese Society for Agricultural Machinery, 2022, 53(8): 337-344. (in Chinese with English abstract)

[7] 呂佩珂,蘇慧蘭,高振江. 多年生蔬菜水生蔬菜病蟲害診治原色圖鑒[M]. 北京:化學(xué)工業(yè)出版社,2013.

[8] 高澤普. 不同設(shè)施栽培模式對蓮藕生長發(fā)育的影響及腐敗病防治效果研究[D]. 邯鄲:河北工程大學(xué),2021. GAO Zepu. The Effect of Different Cultivation Modes on the Growth and Development of Lotus Root and the Study on the Control Effectiveness of Lotus Rot Disease[D]. Handan: Hebei University of Engineering, 2021. (in Chinese with English abstract)

[9] 丁自立,吳金平,刁英,等. 蓮藕主要病害及其綜合防治[J]. 湖北農(nóng)業(yè)科學(xué),2011,50(15):3078-3079. DING Zili, WU Jinping, DIAO Ying, et al. The main disease and comprehensive prevention of[J]. Hubei Agricultural Sciences, 2011, 50(15): 3078-3079. (in Chinese with English abstract)

[10] 孫亮,柯宇航,劉輝,等. 計(jì)算機(jī)視覺技術(shù)在植物病害識別上的研究進(jìn)展[J]. 熱帶生物學(xué)報(bào),2022,13(6):651-658. SUN Liang, KE Yuhang, LIU Hui, et al. Advances in recognition of plant diseases based on computer vision[J]. Journal of Tropical Biology, 2022, 13(6): 651-658. (in Chinese with English abstract)

[11] 黃建平,陳鏡旭,李克新,等. 基于神經(jīng)結(jié)構(gòu)搜索的多種植物葉片病害識別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2020,36(16):166-173. HUANG Jianping, CHEN Jingxu, LI Kexin, et al. Identification of multiple plant leaf diseases using neural architecture search[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(16): 166-173. (in Chinese with English abstract)

[12] XU L, CAO B, NING S, et al. Peanut leaf disease identification with deep learning algorithms[J]. Molecular Breeding, 2023, 43(4): 25.

[13] ABADE A, FERREIRA PA, DE BARROS V F. Plant diseases recognition on images using convolutional neural networks: A systematic review[J]. Computers and Electronics in Agriculture, 2021, 185: 106125.

[14] 丁永軍,張晶晶,李民贊. 基于卷積膠囊網(wǎng)絡(luò)的百合病害識別研究[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2020,51(12):246-251. DING Yongjun, ZHANG Jingjing, LI Minzan. Disease detection of lily based on convolutional capsule network[J]. Transactions of the Chinese Society for Agricultural Machinery, 2020, 51(12): 246-251. (in Chinese with English abstract)

[15] 孫俊,朱偉棟,羅元秋,等. 基于改進(jìn) MobileNet-V2 的田間農(nóng)作物葉片病害識別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2021,37(22):161-169. SUN Jun, ZHU Weidong, LUO Yuanqiu, et al. Recognizing the diseases of crop leaves in fields using improved Mobilenet-V2[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(22): 161-169. (in Chinese with English abstract)

[16] 蘇仕芳,喬焰,饒?jiān)? 基于遷移學(xué)習(xí)的葡萄葉片病害識別及移動端應(yīng)用[J]. 農(nóng)業(yè)工程學(xué)報(bào),2021,37(10):127-134. SU Shifang, QIAO Yan, RAO Yuan. Recognition of grape leaf diseases and mobile application based on transfer learning[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2021, 37(10): 127-134. (in Chinese with English abstract)

[17] ZHOU J, LI J, WANG C, et al. A vegetable disease recognition model for complex background based on region proposal and progressive learning[J]. Computers and Electronics in Agriculture, 2021, 184: 106101.

[18] 彭紅星,徐慧明,劉華鼐,等. 融合雙分支特征和注意力機(jī)制的葡萄病蟲害識別模型[J]. 農(nóng)業(yè)工程學(xué)報(bào),2022,38(10):156-165. PENG Hongxing, XU Huiming, LIU Huanai, et al. Model for identifying grape pests and diseases based on two-branch feature fusion and attention mechanism[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2022, 38(10): 156-165. (in Chinese with English abstract)

[19] 徐艷蕾,孔朔琳,陳清源,等. 基于Transformer的強(qiáng)泛化蘋果葉片病害識別模型[J]. 農(nóng)業(yè)工程學(xué)報(bào),2022,38(16):198-206. XU Yanlei, KONG Shuolin, CHEN Qingyuan, et al. Model for identifying strong generalization apple leaf disease using Transformer[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2022, 38(16): 198-206. (in Chinese with English abstract)

[20] 鮑文霞,孫慶,胡根生,等. 基于多路卷積神經(jīng)網(wǎng)絡(luò)的大田小麥赤霉病圖像識別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2020,36(11):174-181. BAO Wenxia, SUN Qing, HU Gensheng, et al. Image recognition of field wheat scab based on multi-way convolutional neural network[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2020, 36(11): 174-181. (in Chinese with English abstract)

[21] 謝州益,馮亞枝,胡彥蓉,等. 基于 ResNet18 特征編碼器的水稻病蟲害圖像描述生成[J]. 農(nóng)業(yè)工程學(xué)報(bào),2022,38(12):197-206. XIE Zhouyi, FENG Yazhi, HU Yanrong, et al. Generating image description of rice pests and diseases using a ResNet18 feature encoder[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2022, 38(12): 197-206. (in Chinese with English abstract)

[22] 賈兆紅,張?jiān)矗鹾?,? 基于Res2Net和雙線性注意力的番茄病害時(shí)期識別方法[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2022,53(7):259-266. JIA Zhaohong, ZHANG Yuanyan, WANG Haitao, et al. Identification method of tomato disease period based on Res2Net and Bilinear attention mechanism[J]. Transactions of the Chinese Society for Agricultural Machinery, 2022, 53(7): 259-266. (in Chinese with English abstract)

[23] 陳林琳,姜大慶,蔡銀杰,等. 基于計(jì)算機(jī)視覺的芋頭病害監(jiān)測與識別[J]. 農(nóng)機(jī)化研究,2020,42(6):224-229. CHEN Linlin, JIANG Daqing, CAI Yinjie, et al. Monitoring and identification of the taro disease based on the computer vision[J]. Journal of Agricultural Mechanization Research, 2020, 42(6): 224-229. (in Chinese with English abstract)

[24] 王佳. 計(jì)算機(jī)視覺在香芋病害檢測中的應(yīng)用研究[J]. 農(nóng)機(jī)化研究,2020,42(8):241-244. WANG Jia. The applied research of computer vision in the taro disease detection[J]. Journal of Agricultural Mechanization Research, 2020, 42(8): 241-244. (in Chinese with English abstract)

[25] HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), San Juan: IEEE, 2017: 4700-4708.

[26] HU J, SHEN L, SUN G. Squeeze-and-excitation networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Salt Lake City: IEEE, 2018: 7132-7141.

[27] RAPHAEL P. Sharpened cosine distance as an alternative for convolutions[EB/OL]. 2022[2022-01-06]. https: //www. rpisoni. dev/posts/cossim-convolution/

[28] JIA S, WANG P, JIA P, et al. Research on data augmentation for image classification based on convolution neural networks[C]//Chinese Automation Congress (CAC), Jinan: IEEE, 2017: 4165-4170.

[29] 甘雨,郭慶文,王春桃,等. 基于改進(jìn)EfficientNet模型的作物害蟲識別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2022,38(1):203-211. GAN Yu, GUO Qingwen, WANG Chuntao, et al. Recognizing crop pests using an improved EfficientNet model[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2022, 38(1): 203-211. (in Chinese with English abstract)

[30] VISNOI V K, KUMAR K, KUMAR B. Plant disease detection using computational intelligence and image processing[J]. Journal of Plant Diseases and Protection, 2021, 128(1): 19-53.

[31] KRIZHEVSKY A, SUTSKEVER I, HINTON G E. Imagenet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90.

[32] SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[C]//International Conference on Learning Representations (ICLR), San Diego: ICLR, 2015: 1-14.

[33] HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), Las Vegas: IEEE, 2016: 770-778.

[34] XIE S, GIRSHICK R, DOLLáR P, et al. Aggregated residual transformations for deep neural networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), San Juan: IEEE, 2017: 1492-1500.

[35] HE T, ZHANG Z, ZHANG H, et al. Bag of tricks for image classification with convolutional neural networks[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach: IEEE, 2019: 558-567.

[36] WANG R J, LI X, LING C X. Pelee: A real-time object detection system on mobile devices[C]//Advances in neural information processing systems (NeurIPS), Montreal: Neural Information Processing Systems Foundation, 2018, 31: 1963-1972.

[37] LUO C, ZHAN J, XUE X, et al. Cosine normalization: Using cosine similarity instead of dot product in neural networks[C]//International Conference on Artificial Neural Networks (ICANN), Cham: Springer, 2018: 382-391.

[38] KINGMA D P, BA J. Adam: A Method for Stochastic Optimization[C]//International Conference for Learning Representations (ICLR), San Diego: ICLR, 2015: 1-13.

Model for identifying lotus leaf pests and diseases using improved DenseNet and transfer learning

ZHANG Guozhong1,2, LYU Ziwei1,2, LIU Haopeng1,2, LIU Wanru1,2, LONG Changjiang1,2, HUANG Chenglong1,3※

(1.,,430070,; 2.-,,430070,; 3.,.43000,)

Influenced by the ecological environment and other factors, the quality and yield of lotus root have been seriously affected by the occurrence of diseases and insect pests in recent years. With the improvement of living standards and the development of the lotus industry chain, people are looking for green food, high-yield and high-quality products. Nowadays, many farmers and planters are unable to accurately identify the diseases and pests of lotus due to lack of professional knowledge of diseases and insect pests control. There is a shortage of efficient, low-cost and automatic identification technology for the prevention and control of lotus diseases and insect pests. The diagnosis and identification of diseases and insect pests are of great significance for the prevention and control of diseases and insect pests in lotus fields. Over the past few years, deep learning technology has been widely used in the field of plant diseases and insect pests recognition to automatically extract the features of plant diseases and insect pests. In order to achieve an efficient and accurate diagnosis of lotus leaf diseases and insect pests, lotus leaf diseases and insect pests dataset was constructed and preliminary experiments were constructed on AlexNet, VGG-16, ResNet50, ResNeXt50, and DenseNet121 models. The experimental results indicated that DenseNet121 has the best performance on the dataset, lotus leaf diseases and insect pests identification model based on improved DenseNet was improved. Firstly, different methods for dynamic data enhancement were compared in this paper. The results show that resizing and randomly resizing the image is more accurate than directly resizing to the same size. The loss of detail information in part of the image is caused by over-compressing the image size, which affects the model’s recognition effect. The accuracy of the model was increased from 81.47% to 85.01% by using the data enhancement method of resize, random resized crop, random horizontal flip and random adjust sharpness. AdaMax optimizer was used to replace Stochastic Gradient Optimization optimizer and the accuracy of DenseNet model has been improved by 3.04 percentage points. The Stem block uses multi-layer small convolution for fast dimensionality reduction and a branch structure to combine convolution and maximum pooling. It improves the ability of the model to extract shallow features at a lower operating cost. The Squeeze and Excitation attention mechanism block and sharpen cosine similarity convolution were introduced in the Denselayer of the Dense Block and the Transition Layer. This method improved the recognition ability of the model to lotus leaf diseases, and verified the effectiveness of sharpen cosine convolution to improve the performance of the model. Transfer learning was performed on the Plantvillage dataset. The accuracy of the improved model is 91.34%, which 9.87 percentage points higher than before improvement and optimization. In order to solve the problem of monitoring diseases and insect pests in lotus fields, the improved model was applied to the identification of lotus field diseases and insect pests in UAV images. The calibration area of lotus leaf was cut and predicted by reasoning, then different masks were generated according to the model prediction results and added to the UAV image to generate a distribution map of lotus field diseases and insect pests. The recognition of lotus field diseases and insect pests in the UAV image was investigated, automatic classification and recognition of leaf spot, viral disease,, lotusleaf rot and lotus rhizome rot were realized. It provides a new method for efficient and accurate identification and dynamic monitoring of lotus diseases. It also supplies information supports for variable pesticide application and flight path planning in plant disease prevention and control based on UAV.

models; UAV; identification of pests and diseases; lotus leaf; DenseNet; attention mechanism; cosine similarity; transfer learning

2023-01-30

2023-04-07

國家特色蔬菜產(chǎn)業(yè)技術(shù)體系專項(xiàng)資助項(xiàng)目(CARS-24-D-02);湖北省高等學(xué)校優(yōu)秀中青年科技創(chuàng)新團(tuán)隊(duì)計(jì)劃項(xiàng)目(T201934);中央高校基本科研業(yè)務(wù)費(fèi)專項(xiàng)基金資助(2662020GXPY012)

張國忠,教授,博士生導(dǎo)師,研究方向?yàn)楝F(xiàn)代農(nóng)業(yè)裝備設(shè)計(jì)與測控。Email:zhanggz@mail.hzau.edu.cn

黃成龍,副教授,碩士生導(dǎo)師,研究方向?yàn)橹腔坜r(nóng)業(yè)技術(shù)與裝備、植物表型研究。Email:hcl@mail.hzau.edu.cn

10.11975/j.issn.1002-6819.202301122

TP391.4;S24

A

1002-6819(2023)-08-0188-09

張國忠,呂紫薇,劉浩蓬,等. 基于改進(jìn)DenseNet和遷移學(xué)習(xí)的荷葉病蟲害識別模型[J]. 農(nóng)業(yè)工程學(xué)報(bào),2023,39(8):188-196. doi:10.11975/j.issn.1002-6819.202301122 http://www.tcsae.org

ZHANG Guozhong, LYU Ziwei, LIU Haopeng, et al. Model for identifying lotus leaf pests and diseases using improved DenseNet and transfer learning[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2023, 39(8): 188-196. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.202301122 http://www.tcsae.org

猜你喜歡
余弦荷葉卷積
基于3D-Winograd的快速卷積算法設(shè)計(jì)及FPGA實(shí)現(xiàn)
荷葉像什么
從濾波器理解卷積
基于傅里葉域卷積表示的目標(biāo)跟蹤算法
荷葉圓圓
兩個(gè)含余弦函數(shù)的三角母不等式及其推論
分?jǐn)?shù)階余弦變換的卷積定理
荷葉傘
圖像壓縮感知在分?jǐn)?shù)階Fourier域、分?jǐn)?shù)階余弦域的性能比較
小螞蟻喜歡躺在荷葉上