王生生,王 順,張 航,溫長吉
?
基于輕量和積網(wǎng)絡(luò)及無人機(jī)遙感圖像的大豆田雜草識(shí)別
王生生1,王 順2,張 航2,溫長吉3
(1. 吉林大學(xué)計(jì)算機(jī)科學(xué)與技術(shù)學(xué)院,長春 130012;2. 吉林大學(xué)軟件學(xué)院,長春 130012; 3. 吉林農(nóng)業(yè)大學(xué)信息技術(shù)學(xué)院,長春 130118)
為提高機(jī)器視覺在無人機(jī)等小型嵌入式設(shè)備中雜草識(shí)別的準(zhǔn)確率,該文以大豆苗中常見禾本科雜草和闊葉型雜草為研究對(duì)象,針對(duì)傳統(tǒng)和積網(wǎng)絡(luò)在圖像分類任務(wù)中模型參數(shù)多、訓(xùn)練時(shí)間長、含有較多冗余節(jié)點(diǎn)和子樹的問題,該文改進(jìn)傳統(tǒng)和積網(wǎng)絡(luò)的學(xué)習(xí)過程,提出一種以小批量數(shù)據(jù)作為輸入的輕量和積網(wǎng)絡(luò)。在結(jié)構(gòu)學(xué)習(xí)中,當(dāng)積節(jié)點(diǎn)作用域內(nèi)的變量個(gè)數(shù)小于一定閾值時(shí),合并積節(jié)點(diǎn)為多元葉節(jié)點(diǎn),否則將積節(jié)點(diǎn)重組為和積混合結(jié)構(gòu),并對(duì)邊緣節(jié)點(diǎn)進(jìn)行裁剪,有效降低了模型的參數(shù)量和復(fù)雜度。在參數(shù)學(xué)習(xí)中,提出貝葉斯矩匹配更新網(wǎng)絡(luò)參數(shù),使得模型對(duì)小樣本的學(xué)習(xí)效率更高。最后結(jié)合均值聚類算法應(yīng)用于無人機(jī)圖像中的雜草識(shí)別。試驗(yàn)結(jié)果表明,利用該方法對(duì)無人機(jī)圖像中大豆苗、禾本科雜草、闊葉型雜草以及土壤的平均識(shí)別準(zhǔn)確率達(dá)99.5%,高于傳統(tǒng)和積網(wǎng)絡(luò)和傳統(tǒng)AlexNet。并且模型平均參數(shù)量僅為傳統(tǒng)和積網(wǎng)絡(luò)的33%,內(nèi)存需求最大時(shí)減少了549 M,訓(xùn)練時(shí)間最多減少了688.79 s。該研究可為輕量和積網(wǎng)絡(luò)模型在無人機(jī)噴灑農(nóng)藥中的雜草識(shí)別提供參考。
無人機(jī);遙感;識(shí)別;和積網(wǎng)絡(luò);結(jié)構(gòu)學(xué)習(xí);參數(shù)學(xué)習(xí);雜草
雜草是糧食生產(chǎn)的主要制約因素之一[1]。雜草管理的一個(gè)重要目標(biāo)是區(qū)分禾本科雜草和闊葉型雜草,因?yàn)檫@2個(gè)雜草群落可通過不同除草劑適當(dāng)控制[2]。作物田間雜草的識(shí)別方法主要有人工識(shí)別法、遙感識(shí)別法和機(jī)器視覺識(shí)別法3種[3]。
在雜草控制中,采用人工廣泛噴灑除草劑的方法不僅會(huì)造成除草劑的浪費(fèi),還會(huì)造成環(huán)境污染。與此同時(shí),也導(dǎo)致了農(nóng)產(chǎn)品的安全和生態(tài)問題,如化學(xué)農(nóng)藥殘留和雜草群落進(jìn)化產(chǎn)生抗藥性等[4]。在精細(xì)農(nóng)業(yè)中,精準(zhǔn)識(shí)別作物幼苗和雜草,合理使用農(nóng)藥尤為重要[5-6]。
無人機(jī)遙感技術(shù)以其低成本、高分辨率、高靈活性的特點(diǎn)使其成為精細(xì)農(nóng)業(yè)中空中采集圖像的新型工具[7]。如Castro AID等利用低空航拍高光譜圖像通過植被指數(shù)、光譜角制圖等方法繪制谷類和豆類中的十字花科雜草,由識(shí)別的結(jié)果定點(diǎn)噴施農(nóng)藥,節(jié)約了71.7%~95.4%的除草劑[8]。Ishida T等利用無人機(jī)遙感搭載液晶可調(diào)諧濾波器獲得的高光譜圖像,結(jié)合光照和陰影2種光譜反射率對(duì)植被、土壤、雜草等進(jìn)行分類,對(duì)植被的分類準(zhǔn)確率達(dá)94.5%[9]。Barrero O等將無人機(jī)獲取的低分辨率多光譜圖像和高分辨率RGB圖像相融合以檢測水稻萌發(fā)后50天的禾本科雜草,識(shí)別準(zhǔn)確率最高為85%[10]。利用無人機(jī)裝配先進(jìn)傳感器獲取高光譜、高分辨率圖像進(jìn)行光譜分析,通過在某些特定波長作物與雜草反射率的不同來進(jìn)行識(shí)別,彌補(bǔ)了傳統(tǒng)遙感識(shí)別雜草距離遠(yuǎn)、實(shí)時(shí)性差的缺點(diǎn),在農(nóng)業(yè)領(lǐng)域應(yīng)用中具有廣闊前景[11]。
傳統(tǒng)機(jī)器學(xué)習(xí)方法提取雜草的形狀、顏色、紋理[12-14]等特定特征取得了一定效果,但對(duì)于形狀、顏色和紋理差異不明顯的作物與雜草識(shí)別準(zhǔn)確率較低。而深度網(wǎng)絡(luò)模型能夠提取圖像的高層特征,不受人工設(shè)計(jì)特征的影響。如Potena C等[15]提出了一種用于無人地面車輛除草裝置的甜菜雜草分類感知系統(tǒng),結(jié)合RGB和近紅外圖像使用2種不同的卷積神經(jīng)網(wǎng)絡(luò)架構(gòu),淺層網(wǎng)絡(luò)進(jìn)行植被檢測,深層網(wǎng)絡(luò)進(jìn)一步將檢測到的植被分為作物和雜草。他們首先進(jìn)行像素分類,然后通過投票的方式對(duì)植被掩膜中檢測到的斑點(diǎn)進(jìn)行分類,植被識(shí)別的準(zhǔn)確率約97%。Dos A等對(duì)通過對(duì)無人機(jī)獲取的圖像進(jìn)行超像素分割,然后訓(xùn)練卷積神經(jīng)網(wǎng)絡(luò)AlexNet識(shí)別大豆苗和雜草,平均識(shí)別準(zhǔn)確率在99%以上[16]。王璨等利用卷積神經(jīng)網(wǎng)絡(luò)從圖像的高斯金字塔中提取多尺度分層特征,然后與多層感知機(jī)相連接,通過基于像素的分類實(shí)現(xiàn)玉米雜草的識(shí)別,平均識(shí)別準(zhǔn)確率達(dá)98.92%[17]。卷積神經(jīng)網(wǎng)絡(luò)模型在識(shí)別領(lǐng)域效果顯著,其主要問題在于卷積計(jì)算對(duì)硬件資源要求較高,模型占用內(nèi)存大,難以移植到無人機(jī)等小型嵌入式設(shè)備,且模型結(jié)構(gòu)較為復(fù)雜,對(duì)于小樣本數(shù)據(jù)容易造成過擬合[18]。因此,針對(duì)無人機(jī)獲取的大豆苗雜草小樣本數(shù)據(jù),采用了對(duì)硬件資源及樣本數(shù)量要求較小的和積網(wǎng)絡(luò)(sum-product networks)[19]。目前,和積網(wǎng)絡(luò)已成功應(yīng)用于圖像分割,圖像分類,動(dòng)作識(shí)別,語音識(shí)別,目標(biāo)檢測[20-25]等多個(gè)領(lǐng)域。
在大豆種植園中,禾本科雜草和闊葉型雜草嚴(yán)重影響大豆產(chǎn)量,大面積種植時(shí)使用無人機(jī)獲取圖像根據(jù)不同雜草群落噴施農(nóng)藥是有效的防治手段。本文以無人機(jī)獲取的大豆苗雜草小樣本數(shù)據(jù)為研究對(duì)象,在圖像處理階段,基于和積網(wǎng)絡(luò)的上述優(yōu)點(diǎn),提出了一種輕量和積網(wǎng)絡(luò)雜草識(shí)別模型。模型首先使用均值聚類算法提取圖像低層特征,然后將提取的特征下采樣,再將采樣特征以小批量數(shù)據(jù)作為輸入,通過更新網(wǎng)絡(luò)結(jié)構(gòu)與更新網(wǎng)絡(luò)參數(shù)提取高層特征并對(duì)雜草進(jìn)行識(shí)別。在結(jié)構(gòu)學(xué)習(xí)中采用一定方式對(duì)邊緣節(jié)點(diǎn)進(jìn)行裁剪,使模型結(jié)構(gòu)輕量化,以期為輕量和積網(wǎng)絡(luò)模型在無人機(jī)噴灑農(nóng)藥中的雜草識(shí)別提供參考。
無人機(jī)獲取的大豆幼苗期原始圖像如圖1所示,拍攝于2016年1月3日,平均飛行高度約4 m,采集高度變化不大,對(duì)應(yīng)的垂直攝影地面采樣距離小于1 cm。圖像規(guī)格為4 000×3 000像素,JPG格式,采用默認(rèn)工廠配置中的所有參數(shù),沒有使用額外的圖像校正。
圖1 2016年1月3日大豆幼苗期的原始圖像
1.2.1 超像素分割
超像素分割是指將具有相似視覺特征的相鄰像素分割成視覺特征一致的像素塊。它利用像素之間特征的相似性進(jìn)行分組,用少量的超像素代替大量的像素來表達(dá)圖像特征,減少了大量的數(shù)據(jù)冗余,顯著降低了后期圖像處理的復(fù)雜度,是目標(biāo)定位和圖像分割中重要的預(yù)處理環(huán)節(jié)。對(duì)于無人機(jī)獲取的原始圖像使用簡單線性迭代聚類算法(simple linear iterative clustering, SLIC)[26]構(gòu)建圖像數(shù)據(jù)集。
式中d表示第個(gè)像素中心與第個(gè)像素點(diǎn)的顏色距離;d表示第個(gè)聚類中心與第個(gè)像素點(diǎn)的空間距離;D表示所有像素點(diǎn)在CIELAB顏色空間中的距離與以間隔標(biāo)準(zhǔn)化到平面上的距離之和;表示顏色和空間距離的平衡因子,在該文獻(xiàn)中默認(rèn)為10。
試驗(yàn)環(huán)境為Win10操作系統(tǒng),CPU Intel(R) Core i7-8 700 K @3.60 GHz,16 G運(yùn)行內(nèi)存。對(duì)無人機(jī)獲取的原始圖像使用了scikit-image庫中的SLIC超像素分割算法。為將原始圖像分割成大豆苗、雜草和土壤片段,在100至1 200范圍內(nèi)測試不同值,最終選擇=300將其分割為大小約200×200像素的片段,該參數(shù)的選擇取決于圖像采集的高度和分辨率。在陰天拍攝的圖像中含有較少的陰影,平衡因子的值可以使用較小的值,以使超像素邊界對(duì)圖像元素的邊緣更加敏感。在晴天,陰影對(duì)圖像的影響使得超像素對(duì)圖像的光照過于敏感,易含有在同一超像素中屬于不同類別的元素。因此可以選擇較大的值,以使空間鄰近信息相對(duì)于顏色和光照的相似性具有更大的權(quán)重[16]。
1.2.2 圖像裁剪
該數(shù)據(jù)集已將無人機(jī)獲取的原始圖像分割成15 336張小圖像,分別為3 249張土壤、7 376張大豆苗、3 520張禾本科雜草和1 191張闊葉型雜草。由于在數(shù)據(jù)集中的圖片大小不一,絕大部分圖像的高度和寬度小于256像素。為便于處理,首先將圖像放置在一個(gè)由黑色背景組成的圖像左上角,大小為512×512像素,然后在左上角以256×256像素將其裁剪。處理后的雜草與土壤圖像如圖2所示,從左至右依次為大豆苗,禾本科雜草,闊葉型雜草和土壤。
圖2 裁剪后的雜草與土壤圖像
1.2.3 數(shù)據(jù)標(biāo)準(zhǔn)化
標(biāo)準(zhǔn)化之后的數(shù)據(jù)將被擴(kuò)展到一個(gè)合理的范圍,并轉(zhuǎn)化為一個(gè)無量綱的純數(shù)據(jù)。當(dāng)圖像中的像素包含多個(gè)維度且不穩(wěn)定時(shí),特征標(biāo)準(zhǔn)化公式可以確保每個(gè)維度都是零均值和單位方差。= {(1),(2),…,(i),…,(n)}表示給定的數(shù)據(jù)集,其中(i)表示數(shù)據(jù)集中第個(gè)樣本。數(shù)據(jù)標(biāo)準(zhǔn)化公式[27]可表示為
1.2.4 數(shù)據(jù)降維
和積網(wǎng)絡(luò)是一種新型概率深度網(wǎng)絡(luò)。它可以看成是一個(gè)含有根結(jié)點(diǎn)的廣義有向無環(huán)圖,內(nèi)部節(jié)點(diǎn)由和節(jié)點(diǎn)與積節(jié)點(diǎn)遞歸組成,葉結(jié)點(diǎn)可以是離散或連續(xù)的概率分布[19],其主要有如下定義:
定義三(和積網(wǎng)絡(luò)的計(jì)算):若表示在實(shí)數(shù)變量集上的和積網(wǎng)絡(luò),其參數(shù)為,S為以節(jié)點(diǎn)為根結(jié)點(diǎn)的子網(wǎng)絡(luò)。對(duì)于每一個(gè)隨機(jī)變量服從的分布,以=作為可觀測變量的網(wǎng)絡(luò)輸入,S()表示節(jié)點(diǎn)處的輸出值,S()表示節(jié)點(diǎn)下子節(jié)點(diǎn)輸出值。此時(shí),S()可由式(4)計(jì)算。
定義四(完整性):和積網(wǎng)絡(luò)具有完整性當(dāng)且僅當(dāng)其和節(jié)點(diǎn)的所有子節(jié)點(diǎn)具有相同的作用域。
定義五(可分解性):和積網(wǎng)絡(luò)具有可分解性當(dāng)且僅當(dāng)其積節(jié)點(diǎn)的所有子節(jié)點(diǎn)具有互斥的作用域。
注:X1,X2為布爾變量,,分別對(duì)應(yīng)于上述變量相反的邏輯狀態(tài)。
和積網(wǎng)絡(luò)在表達(dá)能力、推理能力及易處理性上具有深厚的理論支持。它的積節(jié)點(diǎn)可表示提取的特征,和節(jié)點(diǎn)可以表示特征的混合[19]。和積網(wǎng)絡(luò)可以解釋為一種具有非負(fù)參數(shù)的特殊前饋神經(jīng)網(wǎng)絡(luò),其中葉結(jié)點(diǎn)是輸入神經(jīng)元,而和節(jié)點(diǎn)和積節(jié)點(diǎn)是隱藏神經(jīng)元。它也可以作為一種密度估計(jì)器,相對(duì)于傳統(tǒng)概率圖模型,在一些推理任務(wù)如邊際推理,最大可能解釋推理中可達(dá)到精確和快速的推理[30]。和積網(wǎng)絡(luò)的學(xué)習(xí)分為參數(shù)學(xué)習(xí)和結(jié)構(gòu)學(xué)習(xí)。其參數(shù)學(xué)習(xí)一般需要人工預(yù)定義網(wǎng)絡(luò)結(jié)構(gòu),如使用極大似然估計(jì)[19],判別式學(xué)習(xí)[31],坍塌變分推理[32]等。因此近年來提出了幾種自動(dòng)化結(jié)構(gòu)學(xué)習(xí)的方法,能夠在構(gòu)建網(wǎng)絡(luò)的同時(shí)學(xué)習(xí)參數(shù),無需超參數(shù)調(diào)節(jié),顯著降低了人工預(yù)定義的網(wǎng)絡(luò)的成本。如傳統(tǒng)和積網(wǎng)絡(luò)LearnSPN算法以及基于該算法的變體[33-35],它們采用分治策略,分層聚類遞歸劃分?jǐn)?shù)據(jù)集實(shí)例與變量生成網(wǎng)絡(luò)結(jié)構(gòu)。Vergari A等則通過一種混合技術(shù)自頂向下學(xué)習(xí)含有多元葉結(jié)點(diǎn)的網(wǎng)絡(luò)結(jié)構(gòu),進(jìn)一步提高了和積網(wǎng)絡(luò)的表達(dá)能力[36]。
傳統(tǒng)和積網(wǎng)絡(luò)的LearnSPN算法使用多次掃描的批處理學(xué)習(xí)方式,通過在線期望最大化算法遞歸劃分?jǐn)?shù)據(jù)矩陣的實(shí)例和變量,這樣無需指定集群的數(shù)量,但復(fù)雜的網(wǎng)絡(luò)結(jié)構(gòu)是通過爆炸性地增加和節(jié)點(diǎn)后裔節(jié)點(diǎn)的個(gè)數(shù)實(shí)現(xiàn)的,含有較多冗余節(jié)點(diǎn)且訓(xùn)練時(shí)間較長。參數(shù)學(xué)習(xí)使用極大似然估計(jì),葉結(jié)點(diǎn)全部為單變量分布,較為簡單。但由于網(wǎng)絡(luò)結(jié)構(gòu)復(fù)雜,易將噪聲數(shù)據(jù)的特征也學(xué)習(xí)到模型當(dāng)中,導(dǎo)致模型泛化能力下降,容易過擬合。
1.4.1 特征提取
圖4 K均值聚類
1.4.2 輕量和積網(wǎng)絡(luò)雜草識(shí)別流程
輕量和積網(wǎng)絡(luò)雜草識(shí)別流程如圖5所示,模型包括均值聚類和輕量和積網(wǎng)絡(luò)兩部分。經(jīng)由上文所述方法,利用均值聚類提取輸入圖像的低層特征,然后以小批量學(xué)習(xí)的方式,將下采樣后的數(shù)據(jù)經(jīng)一次傳遞訓(xùn)練輕量和積網(wǎng)絡(luò),提取圖像高層特征并進(jìn)行雜草識(shí)別。輕量和積網(wǎng)絡(luò)由內(nèi)部節(jié)點(diǎn)學(xué)習(xí)輸入特征的聯(lián)合概率分布,由根結(jié)點(diǎn)輸出特征所屬類別的概率。不同種類分別訓(xùn)練,在進(jìn)行雜草識(shí)別時(shí),每一種類別對(duì)應(yīng)于一個(gè)獨(dú)立的網(wǎng)絡(luò)結(jié)構(gòu)。對(duì)于輸入圖像,每個(gè)網(wǎng)絡(luò)結(jié)構(gòu)輸出該類別的概率值,概率最高的即為所屬類別,依據(jù)識(shí)別結(jié)果可供無人機(jī)選擇噴施特定農(nóng)藥。
圖5 雜草識(shí)別流程
更新網(wǎng)絡(luò)結(jié)構(gòu)的方法是將作用域包含2個(gè)變量的子節(jié)點(diǎn)相關(guān)聯(lián)。當(dāng)子節(jié)點(diǎn)的個(gè)數(shù)小于一定數(shù)量時(shí)就創(chuàng)建一個(gè)多元的葉結(jié)點(diǎn),反之則在變量上創(chuàng)建一個(gè)和積混合結(jié)構(gòu)。此過程如圖6所示。
注:x1 ,…, x5表示積節(jié)點(diǎn)的作用域。
圖6a表示一個(gè)含有3個(gè)子節(jié)點(diǎn)的原始積節(jié)點(diǎn),1,…,5表示該節(jié)點(diǎn)下的作用域。關(guān)聯(lián)1與3的方法如圖6b所示,積節(jié)點(diǎn)跟蹤這5個(gè)變量的經(jīng)驗(yàn)均值和經(jīng)驗(yàn)協(xié)方差。當(dāng)1和3的皮爾遜相關(guān)系數(shù)高于一定閾值時(shí),該算法將具有1,32個(gè)子節(jié)點(diǎn)的作用域合并,并將它們轉(zhuǎn)換為一個(gè)多元葉節(jié)點(diǎn),以其統(tǒng)計(jì)的均值和協(xié)方差作為葉結(jié)點(diǎn)參數(shù)。另一種關(guān)聯(lián)1和3的方法是創(chuàng)建一個(gè)混合模型,如圖6c所示,它有2個(gè)組成部分。第一個(gè)部分為包含1和3積節(jié)點(diǎn)的原始子節(jié)點(diǎn)。第二個(gè)組成部分是一個(gè)新的積節(jié)點(diǎn),它再次被初始化為在其作用域上的一個(gè)完全因式分解的分布,然后再將小批數(shù)據(jù)點(diǎn)傳遞給新混合模型以更新其參數(shù)。更新結(jié)構(gòu)的方式可通過設(shè)置子節(jié)點(diǎn)個(gè)數(shù)的閾值來選擇。
矩匹配是一種基于經(jīng)驗(yàn)矩估計(jì)分布參數(shù)的常用方法。例如,在保證一致性的同時(shí),矩匹配被用來估計(jì)混合模型、主題生成模型和隱藏馬爾科夫模型的參數(shù)[38]。矩匹配也可以用于貝葉斯近似一個(gè)難以計(jì)算的后驗(yàn)分布。也就是說,通過計(jì)算這個(gè)分布的矩的子集,選擇一個(gè)與這些矩匹配的易處理分布簇作為近似,如期望傳播[39]。該文提出使用貝葉斯矩匹配更新網(wǎng)絡(luò)中的參數(shù)。貝葉斯學(xué)習(xí)始于一個(gè)在權(quán)重上的先驗(yàn)概率(),然后對(duì)于給定觀測數(shù)據(jù)集,根據(jù)貝葉斯式(7)學(xué)習(xí)計(jì)算后驗(yàn)概率(|)。
對(duì)于給定數(shù)據(jù)集= {(1),(2),…,(m),…,(n)}的小批量學(xué)習(xí)方式,可以將貝葉斯公式改寫為公式(8),即將觀測數(shù)據(jù)集均勻劃分成若干個(gè)批次,表示Batch Size的大小。的取值可根據(jù)GPU或CPU硬件的內(nèi)存需求調(diào)節(jié)。
當(dāng)用狄利克雷分布近似一個(gè)難以計(jì)算的分布時(shí),首先計(jì)算的一階矩和二階矩,然后使用式(12)計(jì)算超參數(shù)以使其與分布有相同的矩。
進(jìn)一步說,對(duì)于由狄利克雷分布的乘積得到的聯(lián)合概率分布(),為了計(jì)算超參數(shù)α,可以通過計(jì)算每個(gè)邊緣分布()的一階矩和二階矩,然后通過貝葉斯矩匹配與狄利克雷分布的乘積來近似,這樣做的目的是為了減小直接顯示計(jì)算的復(fù)雜性。
為了簡化結(jié)構(gòu),當(dāng)一個(gè)積節(jié)點(diǎn)最終只有一個(gè)子節(jié)點(diǎn)時(shí),就將其從網(wǎng)絡(luò)中刪除,而使它的子節(jié)點(diǎn)與其父節(jié)點(diǎn)相連。如果一個(gè)和節(jié)點(diǎn)是另一個(gè)和節(jié)點(diǎn)的最后一個(gè)節(jié)點(diǎn),那么就刪除該和節(jié)點(diǎn),并將其所有子節(jié)點(diǎn)提升一個(gè)層。這樣有效地減少了冗余的邊緣分支和子樹,使模型結(jié)構(gòu)更加輕量化。
為驗(yàn)證模型的適用性,將試驗(yàn)數(shù)據(jù)分為平衡數(shù)據(jù)和非平衡數(shù)據(jù)2組,因?yàn)樵趯?shí)際應(yīng)用中獲取的樣本數(shù)據(jù)往往是非平衡的。平衡數(shù)據(jù)是指模型在訓(xùn)練、測試、驗(yàn)證時(shí)每種類別均使用相同的樣本數(shù)量。反之,當(dāng)每種類別的樣本數(shù)量分布不均勻時(shí)則是非平衡數(shù)據(jù)。在平衡數(shù)據(jù)中每個(gè)類別所選的圖像數(shù)量為1 125張,共4 500張圖片。其中3 000張用于訓(xùn)練,500張用于驗(yàn)證,1 000張用于測試。非平衡數(shù)據(jù)是在沒有限制類別平衡的情況下形成的。為保持和文獻(xiàn)[16]中試驗(yàn)的一致性且便于將數(shù)據(jù)集劃分為訓(xùn)練集、驗(yàn)證集和測試集,使用了15 336張圖片中的15 000張,組成較大數(shù)據(jù)集。丟棄其中的336張大豆類圖像,因?yàn)榇蠖诡惖膱D像數(shù)量遠(yuǎn)比其他類別高,對(duì)試驗(yàn)幾乎沒有影響,其他類別的數(shù)據(jù)則全部使用。非平衡數(shù)據(jù)中的每個(gè)類別70%用于訓(xùn)練,10%用于驗(yàn)證,20%用于最終測試。小批量學(xué)習(xí)Batch Size的值一般設(shè)置為1 000以內(nèi),數(shù)值越小學(xué)習(xí)過程收斂越快,但產(chǎn)生的噪聲越多。經(jīng)多次試驗(yàn),根據(jù)硬件設(shè)備的內(nèi)存需求及試驗(yàn)結(jié)果,選擇Batch Size為62。
2次試驗(yàn)使用輕量和積網(wǎng)絡(luò)與傳統(tǒng)和積網(wǎng)絡(luò)算法LearnSPN、卷積神經(jīng)網(wǎng)AlexNet[16]對(duì)大豆苗、闊葉型雜草、禾本科雜草以及土壤的識(shí)別結(jié)果進(jìn)行對(duì)比。不同模型算法下平衡數(shù)據(jù)與非平衡數(shù)據(jù)下的分類混淆矩陣如表1所示。
準(zhǔn)確率計(jì)算公式為TP/(TP+FP),召回率計(jì)算公式為TP/(TP+FN),式中TP表示真正例,F(xiàn)P表示假正例,F(xiàn)N表示假反例。根據(jù)平衡數(shù)據(jù)與非平衡數(shù)據(jù)下的分類混淆矩陣,可得到平衡數(shù)據(jù)與非平衡數(shù)據(jù)下的準(zhǔn)確率與召回率。由表1平衡數(shù)據(jù)中的結(jié)果可知使用均衡的數(shù)據(jù)集對(duì)模型的訓(xùn)練有所幫助,此時(shí)使用和積網(wǎng)絡(luò)模型、卷積神經(jīng)網(wǎng)絡(luò)模型進(jìn)行雜草識(shí)別的準(zhǔn)確率均在96%以上。本文方法較傳統(tǒng)和積網(wǎng)絡(luò)方法LearnSPN的平均分類準(zhǔn)確率提高了2.8個(gè)百分點(diǎn),平均召回率提高了2.9個(gè)百分點(diǎn);較卷積神經(jīng)網(wǎng)絡(luò)的平均分類準(zhǔn)確率提高了0.4個(gè)百分點(diǎn),平均召回率提升了0.5個(gè)百分點(diǎn)??梢钥吹轿锤倪M(jìn)的傳統(tǒng)和積網(wǎng)絡(luò)其識(shí)別效果低于傳統(tǒng)神經(jīng)網(wǎng)絡(luò),主要是由于其冗余的節(jié)點(diǎn)分支對(duì)識(shí)別精度有一定影響。誤分類的圖片主要由于個(gè)別圖像質(zhì)量差、陰影、采集時(shí)不同類別重疊等問題導(dǎo)致。其中闊葉型雜草和大豆苗容易誤分,某些圖像由于以上原因及植物本身特點(diǎn),即使人眼也很難完全正確分類。而土壤類別的識(shí)別準(zhǔn)確率高于其他3種植物類別,主要由于它在顏色空間中的特征值與其他類別差異較大。
表1 不同模型算法下平衡數(shù)據(jù)與非平衡數(shù)據(jù)的分類混淆矩陣與分析結(jié)果
由表1使用15 000張非平衡數(shù)據(jù)的試驗(yàn)結(jié)果可知,在非平衡數(shù)據(jù)中輕量和積網(wǎng)絡(luò)的平均分類準(zhǔn)確率為99.6%,較平衡數(shù)據(jù)下的平均分類準(zhǔn)確率提高了0.1個(gè)百分點(diǎn),較傳統(tǒng)和積網(wǎng)絡(luò)提高了1.2個(gè)百分點(diǎn)。2組數(shù)據(jù)中的平均識(shí)別準(zhǔn)確率均達(dá)到了99.5%。本文方法在4 500張數(shù)據(jù)時(shí)識(shí)別準(zhǔn)確率已趨于飽和,說明模型對(duì)訓(xùn)練樣本數(shù)的要求較少。在非平衡數(shù)據(jù)中傳統(tǒng)和積網(wǎng)絡(luò)的平均分類準(zhǔn)確率為98.4%,相對(duì)于平衡數(shù)據(jù)提高了1.7個(gè)百分點(diǎn)。平均召回率為98.6%,提高了1.9個(gè)百分點(diǎn)。說明增加訓(xùn)練樣本,對(duì)于和積網(wǎng)絡(luò)的識(shí)別效果有所幫助。卷積神經(jīng)網(wǎng)絡(luò)在非平衡數(shù)據(jù)中的平均分類準(zhǔn)確率和平均召回率為99.5%,較平衡數(shù)據(jù)均提高了0.4個(gè)百分點(diǎn)。可以看到訓(xùn)練數(shù)據(jù)較大時(shí),所有方法的識(shí)別準(zhǔn)確率均有上升,但對(duì)于非平衡數(shù)據(jù),卷積神經(jīng)網(wǎng)絡(luò)更容易受到有更多數(shù)據(jù)的類的影響,相比于其他類別更為敏感,如在闊葉型雜草出現(xiàn)的現(xiàn)象,易誤分為訓(xùn)練樣本較多的類別,這種情況同樣出現(xiàn)在和積網(wǎng)絡(luò)中。
輕量和積網(wǎng)絡(luò)的每個(gè)內(nèi)部節(jié)點(diǎn)在其作用域上均定義了一個(gè)概率分布,其內(nèi)部節(jié)點(diǎn)可看作特征提取器。為了將提取到的特征進(jìn)行可視化,使用文獻(xiàn)[30]中的方法自向下遍歷網(wǎng)絡(luò),使每個(gè)葉結(jié)點(diǎn)根據(jù)自身的分布在實(shí)數(shù)集上生成一些觀測值,相對(duì)于原始RGB圖像,通過作用域函數(shù)對(duì)輸入空間編碼得到的大豆苗、禾本科雜草、闊葉型雜草及土壤的二值圖像和偽彩色圖像如圖7所示。
注:從左至右依次是大豆苗、禾本科雜草、闊葉型雜草和土壤。
卷積神經(jīng)網(wǎng)AlexNet[16]中輸入圖片大小為256×256×3,第一層卷積有96個(gè)11×11×3的卷積核,第二層有256個(gè)5×5×48的卷積核,第三層有384個(gè)3×3×256的卷積核,第四層有384個(gè)3×3×192的卷積核,第五層有256個(gè)3×3×192個(gè)卷積核。前2層全連接層含有4 096個(gè)節(jié)點(diǎn),總參數(shù)量為61 011 208。和積網(wǎng)絡(luò)的參數(shù)量可通過節(jié)點(diǎn)計(jì)數(shù)器累加得到,采用多次訓(xùn)練后的均值。2次試驗(yàn)中輕量和積網(wǎng)絡(luò)的訓(xùn)練時(shí)間、內(nèi)存大小、參數(shù)量對(duì)比LearnSPN,卷積神經(jīng)網(wǎng)絡(luò)AlexNet0性能結(jié)果分析表如表 2所示。
由表2可知,輕量和積網(wǎng)絡(luò)模型的平均參數(shù)量在平衡數(shù)據(jù)下較傳統(tǒng)和積網(wǎng)絡(luò)減少了482 637,約66.70%。在非平衡數(shù)據(jù)下參數(shù)量減少了1 916 572,約66.67%。即2組試驗(yàn)中輕量和積網(wǎng)絡(luò)的平均參數(shù)量為傳統(tǒng)和積網(wǎng)絡(luò)參數(shù)量的33%。內(nèi)存占用在最大時(shí)較傳統(tǒng)和積網(wǎng)絡(luò)減小了549 M,較卷積神經(jīng)網(wǎng)絡(luò)減小了1 072 M。和積網(wǎng)絡(luò)的結(jié)構(gòu)隨著輸入數(shù)據(jù)的不斷增加,其深度逐步加深,網(wǎng)絡(luò)結(jié)構(gòu)不固定。而卷積神經(jīng)網(wǎng)絡(luò)的結(jié)構(gòu)固定,其參數(shù)量與輸入圖片尺寸,卷積核大小、個(gè)數(shù)等有關(guān),與輸入數(shù)據(jù)量無關(guān)。對(duì)于15 000張的數(shù)據(jù),輕量和積網(wǎng)絡(luò)的參數(shù)量仍遠(yuǎn)小于卷積神經(jīng)網(wǎng)絡(luò),說明輕量和積網(wǎng)絡(luò)仍適用于較大數(shù)據(jù)。
此外,使用輕量和積網(wǎng)絡(luò)作為雜草識(shí)別模型,在平衡數(shù)據(jù)下的平均訓(xùn)練時(shí)間較傳統(tǒng)和積網(wǎng)絡(luò)減少了282.15 s,在非平衡數(shù)據(jù)下減少了688.79 s。因?yàn)樾∨康膶W(xué)習(xí)方式只需經(jīng)數(shù)據(jù)的一次傳遞,因此對(duì)于較大數(shù)據(jù)集,其學(xué)習(xí)速率高于批量學(xué)習(xí)。因和積網(wǎng)絡(luò)緊湊的網(wǎng)絡(luò)結(jié)構(gòu),使其在內(nèi)存占用和訓(xùn)練時(shí)間上均優(yōu)于卷積神經(jīng)網(wǎng)絡(luò)。
表2 不同模型算法的性能結(jié)果分析
本文以大豆苗中常見禾本科雜草和闊葉型雜草為研究對(duì)象,提出了一種基于輕量和積網(wǎng)絡(luò)的雜草識(shí)別模型。模型首先使用均值聚類作為低層特征提取器,然后將提取的特征下采樣,再將采樣特征以小批量數(shù)據(jù)作為輸入訓(xùn)練輕量和積網(wǎng)絡(luò)。在結(jié)構(gòu)學(xué)習(xí)中,當(dāng)積節(jié)點(diǎn)作用域內(nèi)的變量個(gè)數(shù)小于一定閾值時(shí),合并積節(jié)點(diǎn)為多元葉節(jié)點(diǎn),否則將積節(jié)點(diǎn)重組為和積混合結(jié)構(gòu),并采用一定方式對(duì)邊緣冗余節(jié)點(diǎn)和子樹進(jìn)行裁剪,這樣有效降低了模型的參數(shù)量和復(fù)雜度,且提出以貝葉斯矩匹配的方法更新網(wǎng)絡(luò)參數(shù),使得模型對(duì)小樣本學(xué)習(xí)的效率更高。在雜草識(shí)別中,由網(wǎng)絡(luò)內(nèi)部節(jié)點(diǎn)學(xué)習(xí)輸入特征的聯(lián)合概率分布,由根結(jié)點(diǎn)輸出特征所屬類別的概率。對(duì)不同類別的雜草,依據(jù)識(shí)別結(jié)果可供無人機(jī)選擇噴施特定農(nóng)藥。試驗(yàn)結(jié)果表明,針對(duì)大豆苗、禾本科雜草和闊葉型雜草以及土壤的小樣本數(shù)據(jù),該模型的平均識(shí)別準(zhǔn)確率達(dá)到了99.5%,參數(shù)量僅為傳統(tǒng)和積網(wǎng)絡(luò)的33%,且遠(yuǎn)小于卷積神經(jīng)網(wǎng)絡(luò),模型占用內(nèi)存最大時(shí)減小了549 M,訓(xùn)練時(shí)間最多減小了688.79 s,更適用于無人機(jī)等小型嵌入式設(shè)備。此外,該模型在訓(xùn)練時(shí)間上也具有明顯優(yōu)勢。
本文不足之處在于前期對(duì)于無人機(jī)圖像獲取的數(shù)據(jù)需要多步處理,數(shù)據(jù)集本身依賴于人工分類,部分圖像有不同類別重疊,在提取特征時(shí)對(duì)此類圖像的誤分類會(huì)增大。但通過調(diào)整分類閾值,整體識(shí)別準(zhǔn)確率可達(dá)到預(yù)期效果。后續(xù)可為深入研究輕量和積網(wǎng)絡(luò)模型在復(fù)雜環(huán)境中的雜草識(shí)別提供參考。
[1] Zhu Jinwen, Wang Jian, Ditommaso A, et al. Weed research status, challenges, and opportunities in China[J]. Crop Protection, 2018: S0261219418300255.
[2] Herrera P, Dorado J, Ribeiro á. A novel approach for weed type classification based on shape descriptors and a fuzzy decision-making method[J]. Sensors, 2014, 14(8): 15304-15324.
[3] 毛文華,王一鳴,張小超,等. 基于機(jī)器視覺的田間雜草識(shí)別技術(shù)研究進(jìn)展[J]. 農(nóng)業(yè)工程學(xué)報(bào),2004,20(5):43-46. Mao Wenhua, Wang Yiming, Zhang Xiaochao, et al. Research advances of weed identification technology using machine vision[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2004, 20(5): 43-46. (in Chinese with English abstract)
[4] Ip R H L, Ang L M, Seng K P, et al. Big data and machine learning for crop protection[J]. Computers & Electronics in Agriculture, 2018, 151: 376-383.
[5] 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)
[6] 黃玉祥,楊青. 精細(xì)農(nóng)業(yè)的環(huán)境效應(yīng)[J]. 農(nóng)業(yè)工程學(xué)報(bào),2009,25(增刊2):250-254. Huang Yuxiang, Yang Qing. Impact of precision agriculture on environment[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2009, 25(Supp. 2): 250-254. (in Chinese with English abstract)
[7] Torressánchez J, Lópezgranados F, Castro A I D. Configuration and specifications of an unmanned aerial vehicle (UAV) for early site specific weed management[J]. Plos One, 2013, 8(3): e58210.
[8] Castro A I D, Jurado-Expósito M, López-Granados F. Airborne multi-spectral imagery for mapping cruciferous weeds in cereal and legume crops[J]. Precision Agriculture, 2012, 13(3): 302-321.
[9] Ishida T, Kurihara J, Viray F A, et al. A novel approach for vegetation classification using UAV-based hyperspectral imaging[J]. Computers & Electronics in Agriculture, 2018, 144: 80-85.
[10] Barrero O, Perdomo S A . RGB and multispectral UAV image fusion for gramineae weed detection in rice fields[J]. Precision Agriculture, 2018, 19(5): 809-822.
[11] Du M, Noguchi N. Monitoring of wheat growth status and mapping of wheat yield’s within-field spatial variations using color images acquired from UAV-camera system[J]. Remote Sensing, 2017, 9(3): 289.
[12] 李先鋒,朱偉興,紀(jì)濱,等. 基于圖像處理和蟻群優(yōu)化的形狀特征選擇與雜草識(shí)別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2010,26(10):178-182. Li Xianfeng, Zhu Weixing, Ji Bin, et al. Shape feature selection and weed recognition based on image processing and ant colony optimization[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2010, 26(10): 178-182. (in Chinese with English abstract)
[13] 毛罕平,胡波,張艷誠,等. 雜草識(shí)別中顏色特征和閾值分割算法的優(yōu)化[J]. 農(nóng)業(yè)工程學(xué)報(bào),2007,23(9):154-158. Mao Hanping, Hu Bo, Zhang Yancheng, et al. Optimization of color index and threshold segmentation in weed recognition[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2007, 23(9): 154-158(in Chinese with English abstract)
[14] 曹晶晶,王一鳴,毛文華,等. 基于紋理和位置特征的麥田雜草識(shí)別方法[J]. 農(nóng)業(yè)機(jī)械學(xué)報(bào),2007,38(4):107-110. Cao Jingjing, Wang Yiming, Mao Wenhua, et al. Weed detection method in wheat field based on texture and position features[J]. Transactions of the Chinese Society for Agricultural Machinery, 2007, 38(4): 107-110. (in Chinese with English abstract)
[15] Potena C, Nardi D, Pretto A. Fast and accurate crop and weed identification with summarized train sets for precision agriculture[C]//International Conference on Intelligent Autonomous Systems. Springer, Cham, 2011: 105-121.
[16] Dos A, Ferreira S, Matte D, et al. Weed detection in soybean crops using convnets[J]. Computers & Electronics in Agriculture, 2017, 143: 314-324.
[17] 王璨,武新慧,李志偉. 基于卷積神經(jīng)網(wǎng)絡(luò)提取多尺度分層特征識(shí)別玉米雜草[J]. 農(nóng)業(yè)工程學(xué)報(bào),2018,34(5):144-151. Wang Can, Wu Xinhui, Li Zhiwei. Recognition of maize and weed based on multi-scale hierarchical features extracted by convolutional neural network[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(5): 144-151. (in Chinese with English abstract)
[18] Li Huiming, Fan Xitian, Li Jiao, et al. A high performance FPGA-based accelerator for large-scale convolutional neural networks[C]// International Conference on Field Programmable Logic & Applications. IEEE, 2016: 1-9.
[19] Poon H, Domingos P. Sum-product networks: A new deep architecture[J]. Research Gate, 2012, 21(5): 689-690.
[20] Rathke F, Desana M, Schn?rr C. Locally adaptive probabilistic models for global segmentation of pathological OCT scans[C]// International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2017: 177-184.
[21] Sguerra B M, Cozman F G. Image classification using sum-product networks for autonomous flight of micro aerial vehicles[C]//2016 5th Brazilian Conference on Intelligent Systems (BRACIS). IEEE, 2016: 139-144.
[22] Amer M R, Todorovic S. Sum Product Networks for Activity Recognition[M]. IEEE Computer Society, 2016.
[23] Peharz R, Kapeller G, Mowlaee P, et al. Modeling speech with sum-product networks: application to bandwidth extension[C]// IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2014: 3699-3703.
[24] Butz C J, Santos A E D, Oliveira J S, et al. Efficient examination of soil bacteria using probabilistic graphical models[C]// International Conference on Industrial, Engineering and Other Applications of Applied Intelligent Systems. Springer, Cham, 2018: 315-326.
[25] Gens R, Domingos P. Learning the structure of sum-product networks[C]// International Conference on Machine Learning. ICML, 2013: 873-880.
[26] Achanta R, Shaji A, Smith K, et al. SLIC superpixels compared to state-of-the-art superpixel methods[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012, 34(11): 2274-2282.
[27] Coates A , Ng A Y . Learning feature representations with K-means[J]. Lecture Notes in Computer Science, 2012, 7700: 561-580.
[28] Coates A, Ng A, Lee H. An analysis of single-layer networks in unsupervised feature learning[C]//Proceedings of the fourteenth International Conference on Artificial Intelligence and Statistics. ICAISC, 2011: 215-223.
[29] Peharz R, Tschiatschek S, Pernkopf F, et al. On theoretical properties of sum-product networks[J]. Journal of Machine Learning Research, 2015,5: 744-752.
[30] Vergari A, Mauro N D, Esposito F. Visualizing and understanding sum-product networks[J]. Machine Learning, 2018,7: 1-23.
[31] Gens R, Domingos P. Discriminative learning of sum-product networks[C]//Advances in Neural Information Processing Systems. NIPS, 2012: 3239-3247.
[32] Zhao Han, Adel T, Gordon G, et al. Collapsed variational inference for sum-product networks[C]// International Conference on International Conference on Machine Learning. JMLR. org, 2016: 1310-1318.
[33] Rooshenas A , Lowd D . Learning sum-product networks with direct and indirect variable interactions[C]// International Conference on International Conference on Machine Learning. JMLR. org, 2014: 710-718.
[34] Adel T, Balduzzi D, Ghodsi A. Learning the structure of sum-product networks via an SVD-based algorithm[C]// Conference on Uncertainty in Artificial Intelligence. UAI, 2015: 32-41.
[35] Molina A, Vergari A, Di Mauro N, et al. Mixed sum-product networks: A deep architecture for hybrid domains[C]// Proceedings of the AAAI Conference on Artificial Intelligence. AAAI, 2018: 3828-3835.
[36] Vergari A, Mauro N D, Esposito F. Simplifying, regularizing and strengthening sum-product network structure learning[C]// Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, Cham, 2015: 343-358.
[37] Kodinariya T M, Makwana P R. Review on determining number of cluster in K-means clustering[J]. International Journal, 2013, 1(6): 90-95.
[38] Anandkumar A, Hsu D, Kakade S M. A method of moments for mixture models and hidden markov models [C]//Conference on Learning Theory. COLT, 2012: 33. 1-33. 34.
[39] Minka T, Lafferty J. Expectation-propagation for the generative aspect model[C]//Proceedings of the Eighteenth conference on Uncertainty in artificial intelligence. Morgan Kaufmann Publishers Inc. , 2002: 352-359.
Soybean field weed recognition based on light sum-product networks and UAV remote sensing images
Wang Shengsheng1, Wang Shun2, Zhang Hang2, Wen Changji3
(1.,,130012,; 2.,,130012,; 3.,,130118,)
In weed control, using unmanned aerial vehicle (UAV) to obtain images, spraying specific pesticides according to different weed communities is an effective means of prevention and control. Sum-product networks is suitable for small embedded devices such as UAV. But it has many parameters, long training time, and more redundant nodes and subtrees in the image classification task, so that the recognition accuracy is not high. In response to these problems, this paper improved the learning process of traditional sum-product networks and used a mini-batch learning method to construct a network model through one pass of data. Its lightweight structure required less hardware resources and was more suitable for small embedded devices such as drones. It had reference significance for the subsequent spraying of pesticides by drones. For the input image, the light sum-product networks weed recognition model first used-means clustering as the low-level feature extractor to obtain the feature dictionary, then downsampled the extracted features, and took the sampling features into mini-batches of data as input to train the light sum-product networks.Each category corresponds to an independent network structure, and the high-level features were extracted by internal nodes in the network structure. The probability values of the corresponding categories were output by the root nodes to identify weeds. The network structure was updated by comparing the correlation coefficients between variables. Bayesian moment matching was used to update the network parameters. To simplify the structure, when a product node had only one child, it was removed from the network, and its child nodes were connected to its parent node. Similarly, if a sum node was the last node of another sum node, then the child node was deleted and all its child nodes were promoted one layer up. This effectively reduced redundant edge branches and made the model structure lighter. Using this method, the average classification accuracy of soybean seedlings, grass weeds, broadleaf weeds and soils in UAV images was 99.5%, and the average sensitivity was 99.6%. And the model parameter quantity was only 33% of the traditional sum-product networks. The parameter quantity would increase with the input of the data flow. The amount of parameters was still much smaller than traditional convolutional neural networks AlexNet when using the larger data sets to construct the light sum-product networks. It showed that the model was suitable for larger data sets. The memory usage was reduced by 549 M compared to the traditional sum-product networks and was reduced by 1 072 M compared to the convolutional neural networks. The maximum average training time was reduced by 688.79 s compared to the traditional sum-product networks, which was much less than the convolutional neural networks. The experimental results showed that using the light sum-product network as the weed recognition model, the model parameters were less, the memory requirements were lower, and the training time was shorter without loss of precision. The shortcoming was that the data acquired by the UAV image in the previous stage needed to be processed in multiple steps. The data set itself relied on manual classification. Some images had different categories of overlap, and the misclassification of such images would increase when the features were extracted. However, by adjusting the classification threshold, the overall classification can achieve the desired results. The research can provide a reference for the use of light sum-product networks in weed recognition of UAV spraying pesticides.
unmanned aerial vehicle; remote sensing; recognition; sum-product networks; structure learning; parameter learning; weed
2018-11-14
2019-02-11
吉林省科技發(fā)展計(jì)劃項(xiàng)目(20190302117GX,20180101334JC, 20180101041JC);吉林省教育廳科研規(guī)劃重點(diǎn)課題(2016186)
王生生,教授,博士生導(dǎo)師,主要從事機(jī)器視覺、農(nóng)業(yè)信息化方面的研究。Email:wss@jlu.edu.cn
10.11975/j.issn.1002-6819.2019.06.010
TP391.41
A
1002-6819(2019)-06-0081-09
王生生,王 順,張 航,溫長吉. 基于輕量和積網(wǎng)絡(luò)及無人機(jī)遙感圖像的大豆田雜草識(shí)別[J]. 農(nóng)業(yè)工程學(xué)報(bào),2019,35(6):81-89. doi:10.11975/j.issn.1002-6819.2019.06.010 http://www.tcsae.org
Wang Shengsheng, Wang Shun, Zhang Hang, Wen Changji. Soybean field weed recognition based on light sum-product networks and UAV remote sensing images[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(6): 81-89. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2019.06.010 http://www.tcsae.org
農(nóng)業(yè)工程學(xué)報(bào)2019年6期