王巧華，梅 璐，馬美湖，高 升，李慶旭

利用機(jī)器視覺與近紅外光譜技術(shù)的皮蛋無損檢測與分級

王巧華1,2，梅 璐1，馬美湖2,3，高 升1，李慶旭1

（1. 華中農(nóng)業(yè)大學(xué)工學(xué)院/農(nóng)業(yè)部長江中下游農(nóng)業(yè)裝備重點實驗室，武漢 430070；2. 國家蛋品加工技術(shù)研發(fā)分中心 華中農(nóng)業(yè)大學(xué)，武漢 430070；3. 華中農(nóng)業(yè)大學(xué)食品科學(xué)技術(shù)學(xué)院，武漢 430070）

為了對優(yōu)質(zhì)蛋、次品蛋和劣質(zhì)蛋這3種皮蛋進(jìn)行檢測及分級，該文應(yīng)用機(jī)器視覺結(jié)合近紅外光譜技術(shù)，研究利用皮蛋凝膠品質(zhì)無損檢測的分級方法。首先采集皮蛋透射光圖像，提取18個圖像顏色特征值，然后將所提取的18維特征利用主成分分析（principal component analysis，PCA）進(jìn)行降維，對PCA降維后的3個主成分建立遺傳算法優(yōu)化支持向量機(jī)(genetic algorithm-support vector machine，GA-SVM）分級模型，把皮蛋樣本分為兩大類：可食用蛋（優(yōu)質(zhì)蛋與次品蛋）與不可食用蛋（劣質(zhì)蛋），劣質(zhì)蛋測試集識別率為100%。然后在機(jī)器視覺分類結(jié)果的基礎(chǔ)上，利用近紅外光譜技術(shù)獲取可食用蛋（優(yōu)質(zhì)蛋與次品蛋）的原始光譜，并進(jìn)行多元散射矯正（multiplicative scatter correction，MSC），利用競爭性自適應(yīng)重加權(quán)算法（competitive adaptive reweighted sampling，CARS）降維提取特征波長，基于支持向量機(jī)（support vector machine，SVM)對特征波長變量建立分級模型，區(qū)分出優(yōu)質(zhì)蛋與次品蛋，優(yōu)質(zhì)蛋測試集識別率為96.49%，次品蛋識別率為94.12%。研究結(jié)果表明：基于機(jī)器視覺和近紅外光譜進(jìn)行皮蛋凝膠品質(zhì)無損檢測分級是可行的。

機(jī)器視覺；近紅外光譜；凝膠品質(zhì)；皮蛋；支持向量機(jī)

0 引 言

皮蛋又名松花蛋、彩蛋，是中國特有的一種蛋制品[1-2]。它多以鮮鴨蛋為原料，在堿液中經(jīng)過蛋白質(zhì)變性制作而成。其營養(yǎng)價值豐富，每100 g的可食用皮蛋中含有32 mg的氨基酸，是鮮蛋含量的11倍[3-5]。皮蛋還具有開胃、去火和治瀉痢等功效，深受廣大人民的喜愛，因此皮蛋占據(jù)了中國再制蛋的主要部分，位于蛋制品產(chǎn)量第一位，目前已出口20多個國家和地區(qū)[6-8]。

皮蛋形成過程一般會從溶膠狀態(tài)轉(zhuǎn)化為溶液狀態(tài)（化清），再從溶液狀態(tài)轉(zhuǎn)化為凝膠狀態(tài)（凝固），凝固后腌制液持續(xù)滲透，蛋白質(zhì)分子間的空間結(jié)構(gòu)遭到破壞，已經(jīng)吸附的結(jié)合水又會釋放出來以自由水狀態(tài)存在，凝膠再次液化（稀化），俗稱堿傷。蛋殼氣孔大小，腌制液濃度、腌制溫度等均對皮蛋凝膠品質(zhì)有影響[1,6]。皮蛋凝膠性的強(qiáng)弱是衡量其品質(zhì)的重要指標(biāo)[4]，工廠對皮蛋分級時，一般分為3級。第1級是凝固完整呈凝膠狀態(tài)的皮蛋，屬于優(yōu)質(zhì)蛋。第2級是輕微堿傷蛋，剝開后有黏殼、爛頭、蠟黃等現(xiàn)象，但仍能食用，屬于次品蛋。第3級是水響蛋，蛋內(nèi)全部液化成水，屬于劣質(zhì)蛋，不能食用。

雖然鮮蛋已經(jīng)實現(xiàn)了全自動檢測處理，但是目前對于傳統(tǒng)蛋制品皮蛋的分級檢測依然依靠人工，通過燈照、手敲等方法來判斷分級，存在勞動強(qiáng)度大、工作效率低等諸多弊端[9]。近紅外光譜和機(jī)器視覺技術(shù)是2種常用的無損檢測方法，具有快速、高效、無損等優(yōu)點[10-18]，廣泛應(yīng)用于農(nóng)產(chǎn)品品質(zhì)檢測分級中，如核桃[19]、紅提[20]、蘋果[21-22]、紫薯[23]、雞蛋[24]等。前人有檢測皮蛋表面斑點、裂紋、振動的初步研究：樂立強(qiáng)等[2]利用機(jī)器視覺技術(shù)研究了皮蛋表面黑斑形成因素與腌制配方的關(guān)系，確定了最佳配比；Wang[8]針對皮蛋表面有大量灰褐色斑點和大塊黑斑使其表面裂紋不易檢測的情況，搭建了偏振光圖像采集系統(tǒng)，利用皮蛋表殼不同點偏振度的差異性來識別裂紋。Chen等[9]利用加速度傳感器采集皮蛋的振動信號來判斷皮蛋的凝膠狀況，通過不同的振動分析表明：凝膠完整皮蛋的衰減率低于80%。但是Chen的檢測方法具有局限性，對環(huán)境要求較高，只能區(qū)分出不可食用蛋（劣質(zhì)蛋）與可食用蛋（優(yōu)質(zhì)蛋與次品蛋）。目前市場尚無關(guān)于皮蛋凝膠品質(zhì)無損檢測分級的研究報道，因此本文結(jié)合圖像與光譜技術(shù)研究一種皮蛋凝膠品質(zhì)無損檢測分級方法。

1 材料與方法

1.1 材料

試驗采用的皮蛋樣本是市場上常見的以鴨蛋為原料的青殼皮蛋，由湖北神丹健康食品有限公司提供，其中優(yōu)質(zhì)蛋210個，次品蛋191個，劣質(zhì)蛋200個，一共601個樣本，3類樣本如圖1所示。

利用TMS-PRO型質(zhì)構(gòu)儀進(jìn)行TPA(texture profile analysis)試驗，測定皮蛋質(zhì)構(gòu)參數(shù)，試驗步驟為：先在質(zhì)構(gòu)儀上裝好P100/R探頭，然后在凝膠測定程序中設(shè)置好試驗參數(shù)：測試前速度為1mm/s，測試速度為1mm/s，測試后速度為2mm/s，壓縮百分比為40%，測試時間間隔為5s。

表1是不同等級皮蛋質(zhì)構(gòu)參數(shù)的統(tǒng)計結(jié)果，從表中可以看出優(yōu)質(zhì)蛋的彈性、硬度、凝聚性和咀嚼性平均值都大于次品蛋，次品蛋的膠粘性大于優(yōu)質(zhì)蛋，優(yōu)質(zhì)蛋的質(zhì)構(gòu)參數(shù)要優(yōu)于次品蛋[1]（劣質(zhì)蛋為液體，無法進(jìn)行質(zhì)構(gòu)試驗）。

圖1 3類皮蛋樣本圖

表1 兩級皮蛋的質(zhì)構(gòu)參數(shù)

1.2 儀器與設(shè)備

人工照蛋采用滬字牌白熾燈泡（功率為100 W）作為光源來觀察皮蛋整體及四周邊緣對光的透射情況。本文機(jī)器視覺裝置仿照人工檢測皮蛋的方法搭建而成，設(shè)計較大光孔，讓更多的光透射皮蛋以便凸顯皮蛋邊緣，裝置示意圖如圖2所示。機(jī)器視覺試驗用到的儀器與設(shè)備有：丹麥JAI公司的AD-080GE雙通道工業(yè)相機(jī)（鏡頭接口為C接口，靶面尺寸為0.847 cm，分辨率為1024×768像素，幀率為30幀/s）；日本Kowa公司的LM6NC鏡頭（C接口，靶面尺寸為1.27 cm，定焦鏡頭，焦距為6 mm）；Cob款帕燈（功率為24 W，顏色為暖白）。采集近紅外光譜所用儀器為美國賽默飛世爾科技公司的Antaris II傅里葉變換近紅外光譜儀。質(zhì)構(gòu)試驗所用儀器為美國FTC公司的TMS-PRO質(zhì)構(gòu)儀。

1.帕燈 2.皮蛋 3.暗箱 4.相機(jī)及鏡頭 5.計算機(jī)

1.3 方法

1.3.1 皮蛋機(jī)器視覺圖像采集

采集圖像時，將皮蛋放置在透光孔處，打開光源，關(guān)上暗箱，調(diào)整相機(jī)的物距和光圈，當(dāng)物距為0.2m，光圈值為1.8時，圖像最為清晰。固定參數(shù)采集皮蛋圖像。

圖3所示是采集3種凝膠品質(zhì)皮蛋的代表性圖像。優(yōu)質(zhì)蛋和次品蛋大頭或小頭部分透光，而劣質(zhì)蛋不透光或透光面積大。

圖3 皮蛋原始圖像

1.3.2 皮蛋近紅外光譜采集

選擇積分球固體采樣模塊采集皮蛋漫反射光譜。采集光譜時，將樣本豎立放置，分別采集皮蛋大頭、小頭的光譜，然后取平均值作為原始光譜。設(shè)置測量波段范圍10000～4000cm-1，掃描次數(shù)32，分辨率4cm-1。從圖4皮蛋的平均光譜圖中可以看到4個明顯的吸收峰，分別位于4270、4628、5153、6900cm-1附近。

圖4 皮蛋平均光譜

2 結(jié)果與分析

2.1 基于機(jī)器視覺的皮蛋檢測分析

2.1.1 去除圖像背景

首先對原始圖像進(jìn)行灰度化處理，然后提取皮蛋輪廓，在比較了canny、sobel、roberts、prewitt、log等邊緣檢測算子的處理結(jié)果后，本文采用canny算子進(jìn)行邊緣檢測，然后利用凸包算法進(jìn)行橢圓擬合，再進(jìn)行掩膜，得到去除背景的圖像。主要處理過程如圖5所示。

圖5 去除圖像背景處理過程

2.1.2 提取圖像特征值

通過分析皮蛋圖像的特點，發(fā)現(xiàn)選擇顏色特征值作為圖像特征能有效表征不同級別皮蛋之間的差異。RGB、HSV顏色空間是常用的2個顏色空間，而皮蛋在Lab顏色空間測定色度，故選擇了RGB、HSV、Lab這3個顏色空間。RGB顏色空間為圖像中每個像素的RGB分量分配了從0～255范圍內(nèi)的強(qiáng)度值，可以形成16777216（256×256×256）種顏色[25]。HSV空間是以色調(diào)、飽和度和明度所組成的圓錐體坐標(biāo)系描述圖像顏色[26]。Lab空間用數(shù)字化的方法來描述人的視覺感應(yīng)。分量用于表示像素的亮度，從純黑到純白，表示從紅色到綠色的范圍，表示從黃色到藍(lán)色的范圍。Lab顏色空間比人類視覺的色域大，常常被用在顏色識別相關(guān)的算法中[27-28]。

針對皮蛋透射圖像特征，在RGB空間提取紅均值（）、綠均值（）、藍(lán)均值（）和對應(yīng)標(biāo)準(zhǔn)差σ、σ、σ為特征值。把圖像轉(zhuǎn)換到HSV空間，提取色調(diào)均值（）、飽和度均值（）、明度均值（）和對應(yīng)標(biāo)準(zhǔn)差σ、σ、σ為特征值。將圖像轉(zhuǎn)換到Lab空間，提取亮度均值（）、紅度均值（）、黃度均值（）和對應(yīng)標(biāo)準(zhǔn)差σ、σ、σ為特征值。經(jīng)過大量預(yù)試驗對比分析，在3個空間共提取了以上18個顏色特征值。

2.1.3 主成分分析及分類建模

為減少數(shù)據(jù)的維度，將提取的顏色特征值進(jìn)行主成分分析，即把數(shù)據(jù)進(jìn)行標(biāo)準(zhǔn)化處理后計算其相關(guān)系數(shù)矩陣、特征矩陣和各主成分載荷矩陣，最終根據(jù)各主成分累計貢獻(xiàn)率來選擇主成分，一般選取累計貢獻(xiàn)率大于90%的前幾個主成分。主成分分析有利于提高模型的收斂速度和識別率[29-30]。

如表2所示、、、、、、σ、σ、σ、σ、σ、σ等顏色特征載荷大，對第1主成分有較高貢獻(xiàn)率。、、、、、σ、σ對第2主成分貢獻(xiàn)較高，、σ、σ對第3主成分貢獻(xiàn)較高。前3個主成分包含了18個顏色特征信息。

表2 主成分因子載荷矩陣

然后基于機(jī)器視覺技術(shù)對可食用蛋（優(yōu)質(zhì)蛋和次品蛋）與不可食用蛋（劣質(zhì)蛋）進(jìn)行鑒別分類，將18個圖像顏色特征值進(jìn)行主成分分析，以測試集識別率為判斷依據(jù)，提取不同數(shù)目的主成分建立判別模型。從圖6可以看出，當(dāng)主成分?jǐn)?shù)為3時測試集識別率最高，且前3個主成分累計貢獻(xiàn)率達(dá)到97.594%，故選取前3個主成分變量輸入到GA-SVM模型當(dāng)中建立分類模型。

圖6 基于機(jī)器視覺的不同主成分?jǐn)?shù)訓(xùn)練集與測試集識別率

將得到的主成分輸入到GA-SVM模型中進(jìn)行訓(xùn)練。建立GA-SVM模型時，采用RBF核函數(shù)和遺傳算法全局尋優(yōu)。若對3種等級皮蛋進(jìn)行三分類時，把601個樣本隨機(jī)分為訓(xùn)練集421個，測試集180個，測試集識別率僅為75.00%。從圖3可以看出可食用的優(yōu)質(zhì)蛋、次品蛋與不可食用的劣質(zhì)蛋圖像差異大，故對只包含優(yōu)質(zhì)蛋和劣質(zhì)蛋的試驗樣本分類，訓(xùn)練結(jié)果如表3所示，把410個樣本隨機(jī)分為訓(xùn)練集287個，測試集123個，測試集識別率為97.56%。把只包含次品蛋和劣質(zhì)蛋的391個試驗樣本，隨機(jī)分為訓(xùn)練集274個，測試集117個，測試集識別率為93.16%，均高于90%。試驗結(jié)果表明機(jī)器視覺技術(shù)可以將皮蛋分為可食用蛋（優(yōu)質(zhì)蛋和次品蛋）與不可食用蛋（劣質(zhì)蛋）。

表3 基于機(jī)器視覺的皮蛋分級結(jié)果

2.2 基于近紅外光譜的皮蛋分級檢測分析

為了消除儀器噪聲、環(huán)境背景等因素的影響，本研究分別使用多元散射矯正（multiplicative scatter correction，MSC）、標(biāo)準(zhǔn)正交變換（standard normal vafiate，SNV）、自動標(biāo)尺放大（autoscale）、Savitzky-Golay卷積平滑處理（SG）等方法對原始光譜信息進(jìn)行預(yù)處理。譜區(qū)范圍為4 000～8 000 cm-1[31]。因全光譜數(shù)據(jù)量大，存在無關(guān)干擾信息，所以對比了競爭性自適應(yīng)重加權(quán)（competitive adaptive reweighted sampling，CARS）、連續(xù)投影算法（successive projections algorithm，SPA）和無信息變量消除（uninformative variable elimination，UVE）等常用的算法后，擇優(yōu)選取CARS算法提取特征波長，其中設(shè)置蒙特卡洛采樣次數(shù)為50，交叉驗證分組數(shù)為5，提取的最大主成分?jǐn)?shù)為10。當(dāng)采樣次數(shù)為23時，誤差最小。利用SVM模型對CARS算法選取的特征波長變量建立分級模型，以確定最優(yōu)預(yù)處理方法和判別模型。

對3種等級皮蛋進(jìn)行三分類時（如表4所示），把601個樣本隨機(jī)分為訓(xùn)練集421個，測試集180個，測試集識別率僅為78.33%。從圖4可以看出，次品蛋與劣質(zhì)蛋的平均光譜信息基本重合，所以三分類時識別率較低。而優(yōu)質(zhì)蛋的平均光譜在4500～5100cm-1、5400～10000cm-1波段與兩者差異明顯。又因基于機(jī)器視覺技術(shù)能將皮蛋分為可食用蛋（優(yōu)質(zhì)蛋和次品蛋）與不可食用蛋（劣質(zhì)蛋），卻難以進(jìn)一步區(qū)分優(yōu)質(zhì)蛋與次品蛋。故可以利用近紅外光譜技術(shù)對所有優(yōu)質(zhì)蛋與次品蛋進(jìn)行分級。

把剔除劣質(zhì)蛋后的401個可食用蛋樣本隨機(jī)分為訓(xùn)練集293個，測試集108個。首先獲取其原始光譜，然后進(jìn)行多元散射矯正。利用CARS算法提取了分布在4 420～5 152 cm-1和5 538～7 042 cm-1波段范圍內(nèi)的75個特征波長，輸入到SVM模型當(dāng)中進(jìn)行訓(xùn)練，采用RBF核函數(shù)和交叉驗證法尋找的最優(yōu)參數(shù)為256，最優(yōu)參數(shù)為0.004 5，測試集識別率達(dá)到95.37%。說明近紅外光譜可以對優(yōu)質(zhì)蛋與次品蛋進(jìn)行分類。

表4 基于近紅外光譜的皮蛋分級結(jié)果

2.3 基于機(jī)器視覺和近紅外光譜的皮蛋綜合分級檢測分析

利用機(jī)器視覺技術(shù)和近紅外光譜技術(shù)單獨對3種等級的皮蛋分級，測試集識別率都低于80%。為了提高分級識別率，本研究嘗試將圖像特征信息和光譜特征信息進(jìn)行特征層信息融合，建立基于機(jī)器視覺和近紅外光譜融合技術(shù)的SVM判別模型，但發(fā)現(xiàn)信息融合模型測試集識別率為77.38%，不能滿足實際生產(chǎn)需要。

故本文提出分步檢測法，先利用機(jī)器視覺技術(shù)把皮蛋分為可食用蛋（優(yōu)質(zhì)蛋和次品蛋）與不可食用蛋（劣質(zhì)蛋），再基于近紅外光譜對可食用蛋進(jìn)行分類，將優(yōu)質(zhì)蛋與次品蛋識別分開。

基于機(jī)器視覺技術(shù)建立分類模型，把601個皮蛋樣本隨機(jī)分為訓(xùn)練集433個，測試集168個，測試集中優(yōu)質(zhì)蛋57個，次品蛋51個，劣質(zhì)蛋60個。利用機(jī)器視覺技術(shù)將劣質(zhì)蛋分出，結(jié)果如表5所示，60個不可食用蛋（劣質(zhì)蛋）和103個可食用蛋（優(yōu)質(zhì)蛋和次品蛋）判斷正確，劣質(zhì)蛋識別率為100%?；诮t外光譜技術(shù)對可食用蛋（優(yōu)質(zhì)蛋和次品蛋）分類，結(jié)果如表5所示，55個優(yōu)質(zhì)蛋和48個次品蛋判斷正確，優(yōu)質(zhì)蛋識別率為96.49%，次品蛋識別率為94.12%。測試集總體識別率為96.38%。

表5 機(jī)器視覺結(jié)合近紅外光譜技術(shù)的分類結(jié)果

對比了不同分級方法對3種等級皮蛋的分級結(jié)果（如表6所示），發(fā)現(xiàn)只有將機(jī)器視覺與近紅外光譜技術(shù)分步綜合起來才能實現(xiàn)皮蛋分級，3種等級皮蛋的識別率均高于90%，其分級準(zhǔn)確率高于其他方法。

表6 不同分級方法對3種等級皮蛋分類的結(jié)果

3 結(jié) 論

本文利用機(jī)器視覺和近紅外光譜技術(shù)對3種凝膠品質(zhì)的皮蛋檢測分級進(jìn)行了研究，通過試驗發(fā)現(xiàn)單獨使用機(jī)器視覺技術(shù)和近紅外光譜技術(shù)或融合2種技術(shù)進(jìn)行分級，識別率均低于80%。但是分步綜合2種方法可實現(xiàn)較為理想的結(jié)果，先利用機(jī)器視覺技術(shù)將劣質(zhì)蛋分出，劣質(zhì)蛋識別率為100%，然后利用近紅外技術(shù)將機(jī)器視覺無法分開的優(yōu)質(zhì)蛋與次品蛋進(jìn)一步鑒別區(qū)分，優(yōu)質(zhì)蛋識別率為96.49%，次品蛋識別率為94.12%?？傮w識別率為96.38%。研究結(jié)果表明機(jī)器視覺綜合近紅外光譜能夠?qū)崿F(xiàn)皮蛋凝膠品質(zhì)無損檢測分級，可望解決當(dāng)前皮蛋品質(zhì)檢測分級難題。

[1] 郭伶，李慧. 溫度對加鋅工藝皮蛋質(zhì)量的影響[J]. 中國家禽，2015，37(2)：42－44. Guo Ling, Li Hui. Effect of temperature on the quality of zinc-containing alkaline-preserved eggs[J]. China Poultry, 2015, 37(2): 42－44. (in Chinese with English abstract)

[2] 樂立強(qiáng)，吳玲，孫靜，等. 基于機(jī)器視覺的皮蛋斑點檢測和最佳生產(chǎn)配方研究[J]. 食品科學(xué)，2012，33(6)：203－208. Le Liqiang, Wu Ling, Sun Jing, et al. Detection of spots on surface of preserved egg and optimal preserved formula based on machine vision[J]. Food Science, 2012, 33(6): 203－208. (in Chinese with English abstract)

[3] 馬磊，嚴(yán)文靜，趙見營，等. 納米SiO2及TiO2改性復(fù)合涂膜提高松花蛋的保鮮效果[J]. 農(nóng)業(yè)工程學(xué)報，2015，31(18)：269－280. Ma Lei, Yan Wenjing, Zhao Jianying, et al. Preserved effect of nano-SiO2and nano-TiO2[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2015, 31(18): 269－280. (in Chinese with English abstract)

[4] 尹月玲，劉瑤，章建浩，等. 納米SiO2復(fù)合涂膜材料包裝松花蛋的保鮮效果[J]. 農(nóng)業(yè)工程學(xué)報，2012，28(25)：281－287. Yin Yueling, Liu Yao, Zhang Jianhao, et al. Effect of composite conting with nano-silicon dioxide on fresh-keeping of preserved eggs[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(25): 281－287. (in Chinese with English abstract)

[5] 毛昶懿. 皮蛋抗腫瘤功效及其機(jī)制的研究[D]. 武漢：華中農(nóng)業(yè)大學(xué)，2018. Mao Changyi. Antitumor Effect and Mechanism of Preserved Egg[D]. Wuhan: Huazhong Agricultural University, 2018. (in Chinese with English abstract)

[6] 周贊. 不同腌制方法和不同金屬鹽類對皮蛋質(zhì)量的影響[D].長沙：湖南農(nóng)業(yè)大學(xué)，2012. Zhou Zan. The Effect of Different Pickled Methods and Metal Salts on the Quality of Preserved Eggs[D]. Changsha: Hunan Agricultural University, 2012. (in Chinese with English abstract)

[7] 劉龍. 皮蛋溏心和蛋殼裂紋檢測技術(shù)研究[D]. 武漢：華中農(nóng)業(yè)大學(xué)，2012.Liu Long. Research of Detection Technology on Soft-yolk and Cracked Preserved Egg[D]. Wuhan: Huazhong Agricultural University, 2012. (in Chinese with English abstract)

[8] Wang Fang, Wen Youxian, Tan Zuojun, et al. Nondestructive detecting cracks of preserved eggshell using polarization technology and cluster analysis[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2014, 30(9): 249－255. (in English with Chinese abstract)王芳，文友先，譚佐軍，等. 基于偏振光和聚類分析的皮蛋殼裂紋無損檢測(英文)[J]. 農(nóng)業(yè)工程學(xué)報，2014，30(9)：249－255.

[9] Chen Yuchuan, Hu Miaolin, Cheng Chingwei. Applying non-destructive techniques to inspect preserved egg products by decay rates[J]. Food Eng, 2011, 104: 30－35.

[10] 李龍，彭彥昆，李永玉. 蘋果內(nèi)外品質(zhì)在線無損檢測分級系統(tǒng)設(shè)計與試驗[J]. 農(nóng)業(yè)工程學(xué)報，2018，34(9)：267－275. Li Long, Peng Yankun, Li Yongyu. Design and experiment on grading system for online non-destructive detection of internal and external quality of apple[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(9): 267－275. (in Chinese with English abstract)

[11] 劉燕德，肖懷春，鄧清，等. 柑桔黃龍病近紅外光譜無損檢測[J]. 農(nóng)業(yè)工程學(xué)報，2016，32(14)：202－208. Liu Yande, Xiao Huaichun, Deng Qing, et al. Nondestructive detection of citrus greening by near infrared spectroscopy[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(14): 202－208. (in Chinese with English abstract)

[12] 黃星奕，錢媚，徐富斌. 基于機(jī)器視覺和近紅外光譜技術(shù)的杏干品質(zhì)無損檢測[J]. 農(nóng)業(yè)工程學(xué)報，2012，28(7)：260－265. Huang Xingyi, Qian Mei, Xu Fubin, et al. Nondestructive detection of dried apricots quality based on machine vision and near-infrared diffuse reflectance spectroscopy technology[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(7): 260－265. (in Chinese with English abstract)

[13] 彭彥昆，趙芳，李龍，等. 利用近紅外光譜與PCA-SVM識別熱損傷番茄種子[J]. 農(nóng)業(yè)工程學(xué)報，2018，34(5)：159－165. Peng Yankun, Zhao Fang, Li Long, et al. Discrimination of heat-damaged tomato seeds based on near infrared spectroscopy and PCA-SVM method[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(5): 159－165. (in Chinese with English abstract)

[14] 陶海龍. 基于機(jī)器視覺和近紅外光譜的馬鈴薯分級檢測方法[D]. 武漢：華中農(nóng)業(yè)大學(xué)，2014.Tao Hailong. The Method for Potato Testing and Grading Based on Machine Vision and Near Infrared Spectrum[D]. Wuhan: Huazhong Agricultural University, 2014. (in Chinese with English abstract)

[15] 李文采，李家鵬，田寒友，等. 基于RGB顏色空間的冷凍豬肉儲藏時間機(jī)器視覺判定[J]. 農(nóng)業(yè)工程學(xué)報，2019，35(3)：294－300. Li Wencai, Li Jiapeng, Tian Hanyou, et al. Determination of storage time for chilled pork by using RGB color space method based on machine vision[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(3): 294－300. (in Chinese with English abstract)

[16] 宋彥，謝漢壘，寧井銘，等. 基于機(jī)器視覺形狀特征參數(shù)的祁門紅茶等級識別[J]. 農(nóng)業(yè)工程學(xué)報，2018，34(23)：279－286. Song Yan, Xie Hanlei, Ning Jingming, et al. Grading Keemun black tea based on shape feature parameters of machine vision[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(23): 279－286. (in Chinese with English abstract)

[17] 張偉，屠康，劉鵬，等. 基于機(jī)器視覺與敲擊振動融合的鴨蛋孵化特性檢測[J]. 農(nóng)業(yè)機(jī)械學(xué)報，2012，43(2)：140－145. Zhang Wei, Tu Kang, Liu Peng, et al. Early fertility detection of hatching duck egg based on fusion between computer vision and impact excitation[J]. Transactions of the Chinese Society for Agricultural Machinery, 2012, 43(2): 140－145. (in Chinese with English abstract)

[18] 仇遜超. 機(jī)器視覺和近紅外光譜對紅松籽品質(zhì)檢測方法的研究[D]. 哈爾濱：東北林業(yè)大學(xué)，2017. Qiu Xunchao. Study on Quality Detection Method of Korean Pine Seed Using Machine Vision and Near Infrared Spectrum[D]. Harbin: Northeast Forestry University, 2017. (in Chinese with English abstract)

[19] 馬文強(qiáng)，張漫，李忠新. 基于近紅外光譜的核桃仁品種快速分類方法[J]. 農(nóng)業(yè)機(jī)械學(xué)報，2015，46(S1)：128－133. Ma Wenqiang, Zhang Man, Li Zhongxin. Rapid classification method of walnut kernel varieties based on near-infrared diffuse reflectance spectra[J]. Transactions of the Chinese Society for Agricultural Machinery, 2015, 46(S1): 128－133. (in Chinese with English abstract)

[20] 高升，王巧華，李慶旭，等. 基于近紅外光譜的紅提維生素C含量、糖度及總酸含量無損檢測方法[J]. 分析化學(xué)，2019，47(6)：941－949. Gao Sheng, Wang Qiaohua, Li Qingxu, et al. Non-destructive detection of vitamin C, sugar content and total acidity of red globe grape based on near-infrared spectroscopy[J]. Chinese Journal of Analytical Chemistry, 2019, 47(6): 941－949. (in Chinese with English abstract)

[21] Roberto Moscetti, Flavio Raponi, Serena Ferri, et al. Real-time monitoring of organic apple (var.) during hot-air drying using near-infrared spectroscopy[J]. Journal of Food Engineering, 2018, 222: 139－150.

[22] 黃辰，費繼友. 基于圖像特征融合的蘋果在線分級方法[J]. 農(nóng)業(yè)工程學(xué)報，2017，33(1)：285－291. Huang Chen, Fei Jiyou. Online apple grading based on decision fusion of image features[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(1): 285－291. (in Chinese with English abstract)

[23] 田瀟瑜，黃星奕，白竣文，等. 基于近紅外光譜技術(shù)的紫薯貯藏期間花青素含量檢測[J]. 農(nóng)業(yè)機(jī)械學(xué)報，2019，50(2)：350－355. Tian Xiaoyu, Huang Xingyi, Bai Junwen, et al. Detection of anthocyanin content of purple sweet potato during storage period based on near infrared spectroscopy[J]. Transactions of the Chinese Society for Agricultural Machinery, 2019, 50(2): 350－355. (in Chinese with English abstract)

[24] Siswantoro J, Hilman M Y, Widiasri M. Computer vision system for egg volume prediction using backpropagation neural network[J]. IOP Conference Series: Materials Science and Engineering, 2017, 273(1):012002.

[25] 孫世鵬. 基于機(jī)器視覺和近紅外高光譜的冬棗檢測方法研究[D]. 楊凌：西北農(nóng)林科技大學(xué)，2017. Sun Shipeng. Detection of Winter Jujube Fruits Based on Machine Vision and Near-infrared Hyperspectral Image[D]. Yangling: Northwest A&F University, 2017. (in Chinese with English abstract)

[26] Bertels L, Smets B, Wolfs D. Dynamic water surface detection algorithm applied on proba-v multispectral data[J]. Remote Sensing, 2016, 8(12): 1010.

[27] 徐海霞. 基于機(jī)器視覺和電子鼻技術(shù)的菠菜新鮮度無損檢測研究[D]. 鎮(zhèn)江：江蘇大學(xué)，2016. Xu Haixia. Study on Nondestructive Detection of Freshness of Post-harvest Spinach Based on Machine Vision and Electronic Nose[D]. Zhenjiang: Jiangsu University, 2016. (in Chinese with English abstract)

[28] 林顥，宋奔騰，金鴻娟，等. 基于嗅覺可視化與圖像處理的食醋醋齡檢測[J]. 農(nóng)業(yè)機(jī)械學(xué)報，2017，48(1)：275－280. Lin Hao, Song Benteng, Jin Hongjuan, et al. Age discrimination of vinegar based on artificial olfaction visualization and image processing[J]. Transactions of the Chinese Society for Agricultural Machinery, 2017, 48(1): 275－280. (in Chinese with English abstract)

[29] Shima B, Sharifuddin M Z, Mehrdad G, et al. Classification of cow milk using artificial neural network developed from the spectral data of single- and three-detector spectrophotometers[J]. Food Chemistry, 2019, 294: 309-315.

[30] Yang Junjie, Pan Jian, Zhu Shuangjie. Application of pca and slda methods for the classification and differentiation of cooked pork from Chinese indigenous pig breeds and a hybrid pig breed[J]. International Journal of Food Properties, 2014, 17(7):1518-1528.

[31] 于晶，溫亞東，王蘿萍，等. 近紅外光譜定性分析中的特征波長篩選研究[J]. 光譜學(xué)與光譜分析，2013，33(11)：2973－2977. Yu Jing, Wen Yadong, Wang Luoping, et al. Study of selecting characteristic wavelengths in qualitative analysis of near infrared spectroscopy[J]. Spectroscopy and Spectral Analysis, 2013, 33(11): 2973-2977.

Nondestructive testing and grading of preserved duck eggs based on machine vision and near-infrared spectroscopy

Wang Qiaohua1,2, Mei Lu1, Ma Meihu2,3, Gao Sheng1, Li Qinxu1

（1.,,,430070,;2.,,430070,;3,,430070,)

Preserved duck eggs are made from fresh duck eggs. The protein was denatured in the alkali liquor, and the preserved egg was divided into 3 grades according to the quality of the gel. The first grade eggs are the preserved duck egg that are solidified in a gel state and belong to the high quality eggs. The second grade eggs are the slight alkali-damaged. After peeling, there are cases of sticky shell, rotten head, sallow, etc. But they can still be eaten, which are the inferior eggs. The third grade eggs are the water-sounding egg. They are liquefied into water in eggs. They cannot be eaten, which are bad eggs. Over the years, the classification of preserved duck eggs in the egg industry is entirely dependent on manual work, which is cumbersome and inefficient, and the market urgently needs relevant detection technology.The quality classification of preserved duck egg gel was studied based on machine vision and near-infrared spectroscopy. It was found that the use of machine vision and near-infrared spectroscopy alone could not accurately classify preserved duck eggs, the grading accuracy was 75% and 78.33%, respectively. But the combination of the two technology could achieve the classification of preserved duck eggs. Firstly, the transmitted light images of preserved duck eggs were collected by using industrial camera. The MATLAB was used to extract 18 image color feature values in RGB、HSV and CIELab or Lab color eigenvalues. Then the extracted 18-dimensional features were reduced by principal component analysis (PCA). The 601 samples were randomly divided into 433 training sets and 168 test sets. And the genetic algorithm-support vector machine (GA-SVM) classification model was built for the three principal components of the PCA. The preserved duck egg samples were divided into edible eggs (high quality eggs and inferior eggs) and inedible eggs (bad eggs). The 60 inedible eggs (bad eggs) in the test set were all judged correctly. The 103 of the 108 edible eggs (high quality eggs and inferior eggs) were judged correctly and 5 were misjudged. Test set recognition rate of bad eggs was 100%. Then the near-infrared spectroscopy technique was used to obtain the original spectrum of edible eggs (high quality eggs and inferior eggs). Multiplicative scatter correction (MSC) was performed, and the characteristic wavelength was extracted by using competitive adaptive re-weighted sampling (CARS) which extracted 75 data in the range of 4 420-51 52 cm-1and 5 538-7 042 cm-1characteristic bands. The 401 samples were randomly divided into 293 training sets and 108 test sets. Based on the support vector machine (SVM), a hierarchical model was established for the characteristic wavelength variable, and edible high quality eggs and inferior eggs were separated. The test set selected 108 edible eggs, including 57 high quality eggs and 51 inferior eggs. 54 high quality eggs and 49 inferior eggs were judged correctly. The recognition rate of high quality egg test set was 96.49%, and that of inferior egg was 94.12%. The results showed that it was feasible to perform non-destructive classification of preserved duck egg based on machine vision and near-infrared spectroscopy. In practical applications, machine vision technology can be used to separate the inferior eggs, and then the near-infrared spectroscopy technique is used to separate the high-quality eggs from the defective eggs. The result provides a reference for realizing the online non-destructive testing of preserved duck eggs.

machine vision; near-infrared spectroscopy; gel quality; preserved duck egg; support vector machine

王巧華，梅 璐，馬美湖，高 升，李慶旭. 利用機(jī)器視覺與近紅外光譜技術(shù)的皮蛋無損檢測與分級[J]. 農(nóng)業(yè)工程學(xué)報，2019，35(24)：314－321. doi：10.11975/j.issn.1002-6819.2019.24.037 http://www.tcsae.org

Wang Qiaohua, Mei Lu, Ma Meihu, Gao Sheng, Li Qinxu. Nondestructive testing and grading of preserved duck eggs based on machine vision and near-infrared spectroscopy[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(24): 314－321. (in Chinese with English abstract) doi：10.11975/j.issn.1002-6819.2019.24.037 http://www.tcsae.org

2019-08-19

2019-12-02

國家自然科學(xué)基金項目（31871863）；公益性行業(yè)（農(nóng)業(yè)）科研專項（201303084）

王巧華，教授，博士生導(dǎo)師，博士，研究方向為農(nóng)畜禽產(chǎn)品無損檢測。Email：wqh@mail.hzau.edu.cn

10.11975/j.issn.1002-6819.2019.24.037

TS253.7

A

1002-6819(2019)-24-0314-08