基于統(tǒng)計復(fù)雜性測度、多重分形譜等方法的柑橘品質(zhì)分級

曹樂平1, 溫芝元2*

(1.湖南生物機(jī)電職業(yè)技術(shù)學(xué)院科研處,長沙410127;2.湖南農(nóng)業(yè)大學(xué)理學(xué)院,長沙410128)

摘要為精確地度量柑橘品質(zhì)分級,研究了病蟲害為害狀冰糖橙缺陷果實復(fù)雜性測度機(jī)器識別、臍橙果實周長-面積分形維數(shù)與分段色調(diào)單位坐標(biāo)化多重分形譜高度/寬度的形狀和顏色分級及糖酸度無損檢測。對冰糖橙生理性缺硼、銹壁虱、油胞凹陷病3種常見病蟲害果實為害狀缺陷在0°—50°主色調(diào)區(qū)域?qū)嵤╅L度為1°的分段,統(tǒng)計各分段色調(diào)區(qū)間像素分布概率,并計算統(tǒng)計復(fù)雜性測度C(Y)與Shannon信息熵H(Y),以C(Y)與H(Y)為檢索詞計算機(jī)查詢果實病蟲害檢索表來進(jìn)行病蟲害缺陷果機(jī)器識別,平均正確識別率為93.33%。對臍橙果實果梗面與側(cè)面在相垂直的2個投影面上的圖像進(jìn)行去背景與邊界輪廓提取操作,計算邊界輪廓周長-面積分形維數(shù),以此為指標(biāo)檢索果實信息字典進(jìn)行臍橙形狀分級,正確率100%。以臍橙果實相對的2個側(cè)面圖像為研究對象,去其背景,將30°—120°主色調(diào)區(qū)域進(jìn)行30°—50°、50°—70°、70°—90°和90°—120°的區(qū)間分割,生成4幅色調(diào)圖像,計算此圖像多重分形譜質(zhì)心坐標(biāo)、高度與寬度,對該高度與寬度進(jìn)行單位質(zhì)心坐標(biāo)化處理,一方面以單位質(zhì)心坐標(biāo)化多重分形譜高度與寬度為指標(biāo)檢索果實信息字典進(jìn)行臍橙顏色分級,正確率98%;另一方面以單位質(zhì)心坐標(biāo)化多重分形譜高度與寬度為參數(shù)通過糖酸度偏最小二乘模型映射果實糖酸度,糖度與酸度標(biāo)準(zhǔn)差分別在0.77及0.36以內(nèi),與實際值的相關(guān)系數(shù)分別在0.8及0.7以上。試驗結(jié)果表明：統(tǒng)計復(fù)雜性測度、周長-面積分形維數(shù)、單位質(zhì)心坐標(biāo)化多重分形譜高度與寬度較精確地反映了柑橘分級中需識別的冰糖橙果實病蟲害缺陷的特征、臍橙果實形狀與顏色特性及內(nèi)部糖酸度無損檢測映射參數(shù)特點。

關(guān)鍵詞冰糖橙與臍橙; 復(fù)雜性測度; 分形維數(shù); 多重分形譜; 病蟲害缺陷果機(jī)器識別; 形狀與顏色機(jī)器分級; 糖酸度無損檢測

中圖分類號S 126; S 666.4文獻(xiàn)標(biāo)志碼A

Citrus quality grading based on statistical complexity measurement and multifractal spectrum method. Journal of ZhejiangUniversity (Agric. & LifeSci.), 2015,41(3)：309-319

Cao Leping1, Wen Zhiyuan2*(1.ScientificResearchDepartment,HunanBiologicalandElectromechanicalPolytechnic,Changsha410127,China; 2.CollegeofScience,HunanAgriculturalUniversity,Changsha410128,China)

SummaryCitrus quality grading can raise the observability degree and grade degree of citrus, improving the product level and increasing market competitiveness. It can also make huge economic and social benefits and increase farmers income and agricultural productivity so as to promote the sustained and healthy development of the citrus industry.

For the purpose of precise measurement of citrus quality grading, the complexity measurement of Bingtang orange defective fruit damaged by diseases and insect pest patterns were studied by machine recognition, along with the navel orange fruit perimeter-area fractal dimension and the shape, color grading and sugar acid nondestructive detection of section tone unit coordinates multifractal spectrum height and width.

Physiological boron deficiency,Eriophyesoleivorusand rind oil spotting disease were very common in Bingtang orange fruits. The 0°—50° main tone region of these diseases and insect pests damage pattern were augmented into the length of 1°, and pixel distribution probability of each segment tone, complexity measurementC(Y) and Shannon entropyH(Y) were calculated.C(Y) andH(Y) were set as the features to identify fruit diseases and insect pests by machine recognition. The background and extracting boundary contour from the two projection images formed by navel orange fruits’ stalk surface and side perpendicular were removed, and then perimeter-area fractal dimension was calculated. The result was used as index to retrieve fruit information and navel orange shape grading. The two side images where navel oranges were relative were set as the research object, and then the background was extracted, and the main tone region 30°—120°of two images were segmented into 30°—50°, 50°—70°, 70°—90° and 90°—120°, and four tone images were created. Its multifractal spectrum barycentric coordinate, height and width were calculated. The height and width were transformed into the unit barycentric coordinate. On one hand, multifractal spectrum height and width of the unit barycentric coordinate were set as the index and were retrieved in fruit information dictionary to grade navel oranges by color; on the other hand, multifractal spectrum height and width of the unit barycentric coordinate were set as parameters and reflected the degree of fruit sugar and acidity by sugar and acidity partial least square mode.

The average correct recognition rate of Bingtang orange disease and insect pest defect fruit was 93.33%. The correct rate of navel orange fruit shape grading was 100%. The correct rate of navel orange color grading was 98%. The standard deviations of sugar and acidity in navel orange were within 0.77 and 0.36, separately. And the correlation coefficients with the true value were above 0.8 and 0.7.

The above results show that calculating complexity measurement, perimeter-area fractal dimension, unit barycentric coordinate multifractal spectrum height and width can better reflect the characteristics of Bingtang orange fruits with disease and insect pest defects which need to grade citrus fruit quality, and can also reflect navel oranges fruit shape, color characteristics and internal sugar and acidity level which are nondestructive detection mapping parameter features.

Key wordsBingtang orange and navel orange; complexity measurement; fractal dimension; multifractal spectrum; machine recognition of defective fruits with diseases and insect pests; machine grading for shape and color; nondestructive detection of sugar acidity

復(fù)雜性本身復(fù)雜多樣,至今復(fù)雜性科學(xué)的發(fā)展還處于萌芽階段,缺少對復(fù)雜性統(tǒng)一的定義[1-3].Seth Lloyd總結(jié)的復(fù)雜性就有分形維、重分形、熵、計算復(fù)雜性等31條定義,也正是因為復(fù)雜性定義的多樣性,對不同定義下復(fù)雜性的刻畫方法也就各不相同。如何刻畫復(fù)雜性問題的復(fù)雜性是20世紀(jì)科學(xué)前沿與研究熱點,巖石的海水與油飽和特性[4]、社會公眾部門的業(yè)務(wù)流程問題[5]、軟件調(diào)整的規(guī)模問題[6]就是復(fù)雜性測度的應(yīng)用嘗試。

柑橘作為一生物體,其生長過程受眾多因素的影響使其形狀、顏色、內(nèi)部品質(zhì)及病蟲害為害狀呈現(xiàn)復(fù)雜性,部分學(xué)者對此類問題進(jìn)行了復(fù)雜性量測的研究。用果梗面和側(cè)面柑橘輪廓盒維數(shù)作為形狀特征值,0°—100° 5個等分色調(diào)區(qū)間色調(diào)盒維數(shù)作為顏色特征值進(jìn)行柑橘形狀與顏色分級,平均正確分級率95.83%[7];以臍橙果實病蟲害為害狀輪廓分形維數(shù)作為特征值之一,結(jié)合為害狀紅色、綠色、藍(lán)色3個顏色參數(shù)識別其病蟲害,平均正確識別率為85.51%[8];用病蟲害為害狀多重分形譜特征值識別柑橘果實病蟲害,平均正確識別率分別為83.12%和92.67%[9-10]。形狀與顏色的盒維數(shù)刻畫只計盒子數(shù),未考慮盒子內(nèi)像素數(shù)使其形狀與顏色的描述粗糙,影響分級正確率;以為害狀區(qū)域顏色分量均值和為害狀邊界分形維數(shù)作為病蟲害識別特征形狀表達(dá)較充分,但識別病蟲害的典型小色斑信息因調(diào)和而被弱化,影響病蟲害正確識別率;以為害狀多重分形譜參數(shù)作為病蟲害識別特征,雖結(jié)合了為害狀的形狀與顏色信息,但計算過程較為復(fù)雜,計算機(jī)消耗大。有鑒于此,本文運(yùn)用周長-面積分形維數(shù)、多重分形、熵、統(tǒng)計復(fù)雜性測度理論,分別研究臍橙果實形狀與顏色計算機(jī)分級、糖酸度無損檢測及冰糖橙果實病蟲害計算機(jī)識別的柑橘實際問題。

1試驗設(shè)備、軟件與樣本

1.1試驗設(shè)備與圖像分析軟件

紐荷爾臍橙果實及冰糖橙果實病蟲害圖像拍攝像機(jī)為索尼DSC-H20,焦距38～380 mm,對焦范圍25 mm～∞,鏡頭f=6.3～63 mm,最高分辨率3 648～2 736像素,快門速度1/4～1/1 600 s。計算機(jī)Lenovo PⅣ2.13 GHz CPU,內(nèi)存512 MB。圖像分析軟件Matlab R2010a及ACDSee 10。

1.2試驗樣本

2013年10月上旬與11月中旬先后2次分別在湖南省永州市藍(lán)山縣與懷化市麻陽縣進(jìn)行健康臍橙果實和冰糖橙病蟲害果實采樣采摘,紅黃色、黃色、綠黃色、黃綠色、綠色健康臍橙果實各色學(xué)習(xí)樣本與檢驗樣本均采樣100個;生理性缺硼(physiological deficient boron)[11]、銹壁虱(Eriophyesoleivorus)、油胞凹陷病(rind oil spotting disease)[12]3種常見的病蟲害冰糖橙果實學(xué)習(xí)樣本與檢驗樣本采樣100個。洗凈2類果實樣本的果面,并晾干。用白紙襯底在自然光照條件下拍攝冰糖橙果實病蟲害為害狀圖像及健康臍橙果實果梗面與相對2側(cè)面圖像。所采集的果實圖像用ACDSee 10軟件進(jìn)行512像素×512像素裁切,備后續(xù)圖像分析用。

2試驗理論與方法

復(fù)雜性問題一般具有非線性、多樣性、多層次性、自相似性等多種復(fù)雜特征[13-15],周長-面積分形維數(shù)、多重分形、信息熵與統(tǒng)計復(fù)雜性測度(statistical complexity measurement)就是解決這類具有復(fù)雜特征問題的理論方法。

2.1周長-面積分形維數(shù)

分形維數(shù)是刻畫非規(guī)則曲線特征差異的重要理論工具,Mandelbrot提出封閉的粗糙曲線周長Q與面積A滿足Q∝A0.5D關(guān)系,將這種關(guān)系應(yīng)用到柑橘果實的分形上有周長-面積分形維數(shù)D[16-17].

(1)

2.2多重分形

多重分形考慮了尺度單元中像素數(shù)使結(jié)果包含了許多被簡單分形所忽略的信息,目前已成為研究分形物質(zhì)的重要手段。定義概率p(δ)的q次方加權(quán)和為一配分函數(shù)χq(δ)[18-21].

(2)

δτ(q),

(3)

式中τ(q)為質(zhì)量指數(shù)。

依據(jù)統(tǒng)計物理方法有

f(α)=αq-τ(q),

(4)

式中:f(α)為多重分形譜;α=dτ(q)/dq為奇異指數(shù)。

2.3統(tǒng)計復(fù)雜性測度

近年來,隨著非線性科學(xué)的發(fā)展與混沌運(yùn)動研究的不斷深入,系統(tǒng)復(fù)雜性的深入研究越來越有必要。復(fù)雜性測度是對對象復(fù)雜程度的客觀度量,大體包含統(tǒng)計復(fù)雜度和算法復(fù)雜度2大類[22-25],本文引入統(tǒng)計復(fù)雜性測度。

在信息熵的基礎(chǔ)上,R. Lòpez-Ruiz等定義統(tǒng)計復(fù)雜性測度

C(Y)=H(Y)B(Y),

(5)

(6)

(7)

顯然,C(Y)反映了系統(tǒng)內(nèi)在排列無序性及系統(tǒng)結(jié)構(gòu)規(guī)則性,可以較方便地表達(dá)復(fù)雜性的程度。

2.4試驗方法

2.4.1形狀分級根據(jù)周長-面積分形維數(shù)理論對臍橙果實按以下步驟進(jìn)行形狀分級.

1)依據(jù)亮度直方圖雙峰分布特性,取谷底亮度作為閾值去除臍橙果實背景,備形狀及顏色分級和糖酸度無損檢測使用。

2)提取臍橙果實邊界,并進(jìn)行邊界跟蹤與細(xì)化。

3)統(tǒng)計果梗面及一個側(cè)面圖像果實邊界像素與果實區(qū)域像素。

4)作邊界像素與區(qū)域像素數(shù)的最小二乘擬合,根據(jù)擬合直線截距計算水果形狀因子β。

5)根據(jù)式(1)計算臍橙果實果梗面、一個側(cè)面的分形維數(shù),以此作為臍橙果實形狀分級的特征值。

6)計算機(jī)檢索臍橙果實信息字典,進(jìn)行形狀定級。

2.4.2顏色分級及糖酸度無損檢測運(yùn)用多重分形理論對臍橙果實實施下列步驟的顏色分級及糖酸度無損檢測.

1)對亮度閾值法去背景的2個側(cè)面臍橙果實圖像進(jìn)行色調(diào)-飽和度-亮度(HSI)色空間轉(zhuǎn)換。

2)對臍橙果實圖像色調(diào)分布范圍[30°,120°]進(jìn)行30°—50°、50°—70°、70°—90°和90°—120°的區(qū)間分割,生成4幅色調(diào)圖。

3)統(tǒng)計δ×δ(δ=21,22,…,29)滑動窗口內(nèi)及整幅色調(diào)圖像像素nij及N,計算像素分布概率Pij(δ)=nij/N(i,j=20,21,…,28)。

5)各色調(diào)區(qū)間多重分形譜位置與形狀均存在差異,為便于比較,計算單位坐標(biāo)化多重分形譜高度η=Δf/f(αc)及單位坐標(biāo)化多重分形譜寬度μ=Δα/αc來作為臍橙果實顏色特征值。

6)計算機(jī)檢索臍橙果實信息字典,進(jìn)行顏色定級與糖酸度映射。

2.4.3臍橙果實信息字典參照《鮮柑橘》(GB/T12947—2008)將臍橙果實外觀品質(zhì)分為優(yōu)等、一等、二等、等外4個等級,果形用果梗面及1個側(cè)面的2個分形維數(shù)作為評價指標(biāo),其整體分布區(qū)間為(1,2),依據(jù)縱橫徑比果形指數(shù)在人工對500個學(xué)習(xí)樣本等級評定的基礎(chǔ)上計算果梗面及側(cè)面分形維數(shù)的級間界點,以此作為分形維數(shù)等級區(qū)間的上下限sub與inf。顏色指標(biāo)依據(jù)《柑橘等級規(guī)格》(NY/T1190—2006)進(jìn)行學(xué)習(xí)樣本著色面積的人工評定,再計算等級間單位坐標(biāo)化多重分形譜高度η及寬度μ來確定infη、infμ、subη、subμ。

臍橙果實糖酸度按照食品衛(wèi)生檢測方法理論部分總則(GB/T5009.1—2003)與食品中總酸的測定方法(GB/T12456—1990)分別用WYT-4型上海有限公司生產(chǎn)的手持糖度計及PHS-2F型南京庚辰科學(xué)儀器公司生產(chǎn)的數(shù)字pH計,對圖像采集后的學(xué)習(xí)樣本與檢測樣本取果肉榨汁攪拌均勻進(jìn)行逐個檢測,將學(xué)習(xí)樣本檢測結(jié)果與臍橙果實4個色調(diào)區(qū)間單位坐標(biāo)化多重分形譜高度及單位坐標(biāo)化多重分形譜寬度進(jìn)行偏最小二乘回歸,建立檢測樣本糖酸度無損檢測模型,檢測樣本糖酸度用于模型評價。

(8)

(9)

用果梗面分形維數(shù)、側(cè)面分形維數(shù)、側(cè)面單位坐標(biāo)化多重分形譜高度與寬度、糖度和酸度6個指標(biāo),建立臍橙果實信息字典,進(jìn)行果實機(jī)器等級查詢定級與糖酸度無損檢測(表1),t為1,2,3和4時分別對應(yīng)優(yōu)等果、一等果、二等果及等外果,高一等級形狀與顏色等級區(qū)間上限sub對應(yīng)低一等級形狀與顏色等級區(qū)間下限inf,保證整體等級區(qū)間連續(xù)不間斷。

表1　臍橙果實信息字典

2.4.4病蟲害缺陷果識別根據(jù)統(tǒng)計復(fù)雜性測度理論對冰糖橙果實病蟲害為害狀缺陷進(jìn)行以下步驟的機(jī)器識別：1)在設(shè)置亮度閾值去除冰糖橙果實背景的基礎(chǔ)上進(jìn)行彩色圖像(RGB空間)至(HSI空間)轉(zhuǎn)換。2)改進(jìn)型分水嶺算法進(jìn)行果實病蟲害為害狀邊界提取,對過分割的區(qū)域?qū)嵭袇^(qū)域連通合并。3)根據(jù)為害狀邊界提取病蟲害為害狀,統(tǒng)計其像素M。4)對病蟲害為害狀色調(diào)進(jìn)行長度為1°的區(qū)間分割,統(tǒng)計各分割區(qū)間像素mk,計算各分割區(qū)間像素分布概率ρk(yk)=mk/M。5)根據(jù)式(7)、(6)、(5)分別計算B(Y)、H(Y)、C(Y),依據(jù)病蟲害果實學(xué)習(xí)樣本確定柑橘生理性缺硼、銹壁虱、油胞凹陷病3種常見病蟲害的為害狀缺陷C(Y)、H(Y)范圍,按檢測樣本C(Y)和H(Y)值計算機(jī)查找其所處范圍,進(jìn)而確定哪類病蟲害缺陷果實,如表2所示,對于超檢索范圍的情形依據(jù)數(shù)值與檢索范圍最近的原則進(jìn)行判別。

表2　冰糖橙果實病蟲害檢索表

3結(jié)果與分析

3.1結(jié)果

依據(jù)臍橙果實形狀、顏色特征值計算機(jī)查詢信息字典,進(jìn)行果實形狀、顏色單獨定級與糖酸度無損檢測,特征值超出形狀與顏色檢索范圍的,以與哪個等級區(qū)間距離最近評定為哪級的最短距離原則實施等級界定;顏色特征值跨2個等級的以屬于哪級指標(biāo)數(shù)多界定為哪級為原則,若出現(xiàn)屬于2個等級指標(biāo)數(shù)一致的情形則以2個等級中低的等級進(jìn)行評定。對比人工等級評定結(jié)果,計算機(jī)形狀分級正確率100%,顏色分級正確率98%,5種顏色共500個檢驗樣本糖度與酸度無損檢測相對誤差范圍分別為-19.79%～30.90%和-19.37%～24.38%。

根據(jù)復(fù)雜性測度C(Y)、Shannon信息熵H(Y)2個特征識別參數(shù)值計算機(jī)查詢冰糖橙果實病蟲害檢索表,生理性缺硼、銹壁虱、油胞凹陷病3種病蟲害為害狀缺陷果各100個檢驗樣本的正確識別率分別為93%、95%和92%,3種病蟲害為害狀缺陷果平均正確識別率為93.33%。

3.2分析

3.2.1形狀分級分析按形狀分級試驗方法的步驟計算健康果實互相垂直的果梗面及側(cè)面2個投影面分形維數(shù),D1∈[1.010 2,1.017 0],D2∈[1.018 9,1.025 8],標(biāo)準(zhǔn)差s1=0.000 9,s2=0.001 3。臍橙果實果梗面投影接近于圓,分形維數(shù)較小,果實側(cè)面投影接近橢圓,分形維數(shù)較大。按周長-面積分形維數(shù)形狀等級區(qū)間劃分,優(yōu)等、一等、二等及等外4個等級界定清楚無誤.圖1為100個檢測樣本計算形狀分級情況,①、②、③、④、⑤和Ⅰ、Ⅱ、Ⅲ、Ⅳ、Ⅴ分別為果梗面及與側(cè)面形狀等級區(qū)間界點inf和sub,優(yōu)等為①、②水平線與Ⅰ、Ⅱ鉛垂線圍成的矩形區(qū)域,一等、二等及等外分別為②、③水平線與Ⅱ、Ⅲ鉛垂線,③、④水平線與Ⅲ、Ⅳ鉛垂線,④、⑤水平線與Ⅳ、Ⅴ鉛垂線圍成的“7”字形區(qū)域,圖中優(yōu)等果7個、一等果18個、二等果39個、等外果36個。2個互相垂直的投影面果實輪廓分形維數(shù)較準(zhǔn)確地刻畫了果實準(zhǔn)橢球體的立體形狀特征,檢測樣本計算機(jī)形狀分級未出現(xiàn)誤判,較文獻(xiàn)[8]正確分級率高,果實形狀表達(dá)較以幾何參數(shù)進(jìn)行形狀度量的文獻(xiàn)[26-27]及以傅里葉描述子刻畫形狀的文獻(xiàn)[28]精確與全面,正確分級率高。

3.2.2顏色分級分析臍橙果實顏色表征按其試驗方法的步驟分析果實多重分形特征,如圖2所示,在δ=26～29范圍內(nèi)配分函數(shù)呈近似放射狀,該尺度區(qū)間可認(rèn)為具有標(biāo)度不變性,在該區(qū)間研究其多重分形譜特性才有理論依據(jù)。圖3為紅黃、黃色、綠黃、黃綠、綠色各1個檢測樣本在30°—50°、50°—70°、70°—90°和90°—120° 4個色調(diào)區(qū)間的多重分形譜線。

圖1　檢驗樣本分形維數(shù)分布 Fig.1　Fractal dimension of test samples

圖2　黃色果實標(biāo)度 Fig.2　Yellow fruit scales

A:紅黃果;B:黃果;C:綠黃果;D:黃綠果;E:綠果. A: Red-yellow fruit; B: Yellow fruit; C: Green-yellow fruit; D: Yellow-green fruit; E: Green fruit. 圖3　多重分形譜 Fig.3　Multifractal spectra

由圖3可知,分形譜雖有形狀的不同,但位置也存在明顯差異,對分形譜高度與寬度分別以質(zhì)心縱橫坐標(biāo)進(jìn)行單位坐標(biāo)化處理,一方面降低了顏色表征數(shù)據(jù)維數(shù),另一方面避免了數(shù)據(jù)交叉與重疊,充分反映與表達(dá)了等級間臍橙果實色澤的差異。圖4為圖3單位坐標(biāo)化多重分形譜高度與寬度分布。臍橙果實以4個色調(diào)區(qū)間單位坐標(biāo)化多重分形譜高度與寬度8個顏色特征參數(shù)進(jìn)行顏色分級,較以0°—100° 5等分色調(diào)區(qū)域分形維數(shù)為顏色分級參數(shù)的文獻(xiàn)[1]分級正確率高,較以紅色、綠色或藍(lán)色均值為顏色特征值進(jìn)行果實色澤分級的文獻(xiàn)[29-31]顏色描述徹底,分級精度有明顯改進(jìn)。

A: 30°—50°; B: 50°—70°; C: 70°—90°; D: 90°—120°. 圖4　多重分形譜高度與寬度分布 Fig.4　Height and width distribution of multifractal spectra

3.2.3糖酸度無損檢測分析根據(jù)糖酸度偏最小二乘無損檢測模型,用30°—50°、50°—70°、70°—90°和90°—120° 4個色調(diào)區(qū)間的單位坐標(biāo)化多重分形譜高度與寬度映射臍橙果實糖度與酸度,糖度、有效酸度預(yù)測標(biāo)準(zhǔn)差分別為0.765 2和0.358 7,5種顏色共500個檢驗樣本糖度與酸度無損檢測值與實際值相關(guān)系數(shù)r分別在0.8和0.7以上.圖5表明了5種顏色各20個檢測樣本糖酸度無損檢測與理化檢測的相關(guān)程度。以單位坐標(biāo)化多重分形譜高度與寬度8個顏色特征參數(shù)無損檢測臍橙果實糖度與酸度,其相對誤差較以0°—120° 6等分色調(diào)區(qū)域分形維數(shù)為參數(shù)的文獻(xiàn)[16]低,糖度與酸度無損檢測值與實際值相關(guān)系數(shù)高。

3.2.4病蟲害缺陷果識別分析以改進(jìn)型分水嶺算法進(jìn)行臍橙病蟲害為害狀邊緣檢測存在過分割現(xiàn)象,在此基礎(chǔ)上進(jìn)行檢測區(qū)域聯(lián)通與合并生成病蟲害為害狀邊界輪廓,依據(jù)此輪廓提取臍橙病蟲害為害狀,使后續(xù)圖像分析與特征參數(shù)提取不受果面其他區(qū)域的影響。圖6為3種病蟲害缺陷果各1個樣本病蟲害邊界輪廓、病蟲害為害狀情況。

考察色調(diào)0°—50°之間像素,各病蟲害為害狀在該區(qū)域像素占0°—360°色調(diào)像素比均在95%以上,該色調(diào)區(qū)間像素分布不失一般性,同時也減輕了計算工作量.表3給出了檢測樣本0°—50°色調(diào)區(qū)間像素分布概率范圍及復(fù)雜性測度C(Y)、Shannon信息熵H(Y)均值.從中可以看出,不同病蟲害果面留下的為害狀缺陷其C(Y)、H(Y)2個參數(shù)有所不同,較以果面病蟲害為害狀缺陷紅色、綠色、藍(lán)色分量和為害狀邊界分形維數(shù)為特征值的文獻(xiàn)[8]平均識別正確率高,較以果面病蟲害為害狀缺陷多重分形譜高度、寬度為參數(shù)的文獻(xiàn)[9],及以果面病蟲害為害狀缺陷傅里葉變換幅度譜圖多重分形譜的高度、寬度和質(zhì)心坐標(biāo)作為特征值的文獻(xiàn)[10]識別方法簡單,平均識別正確率稍有提高,說明能用此方法識別病蟲害缺陷果。

圖5　檢驗樣本糖酸度預(yù)測值與實際值相關(guān)性 Fig.5　Relationship between prediction and true values sugar content or valid acidity in test samples

A:銹壁虱樣本;B:銹壁虱邊界輪廓;C:銹壁虱為害狀;D:生理性缺硼樣本;E:生理性缺硼邊界輪廓;F:生理性缺硼為害狀;G:油胞凹陷病樣本;H：油胞凹陷病邊界輪廓;I:油胞凹陷病為害狀。 　　A：E. oleivorus sample; B： E. oleivorus boundary; C： E. oleivorus damage pattern; D： Physiological boron deficiency sample; E： Physiological boron deficiency boundary; F： Physiological boron deficiency damage pattern; G： Rind oil spotting disease sample; H： Rind oil spotting disease boundary; I： Rind oil spotting disease damage pattern。 圖6　冰糖橙病蟲害為害狀 Fig.6　Damage patterns of diseases and insect pests in Bingtang orange

檢測樣本Testsamples銹壁虱E.oleivorus生理性缺硼Physiologicalborondeficiency油胞凹陷病Rindoilspottingdiseaseρ(y)[0,0.1282][0,0.1270][0,0.2357]C(Y)7.19556.83877.4759H(Y)4.23204.32614.1143

4討論與結(jié)論

4.1臍橙果實呈準(zhǔn)橢球狀,用其投影面輪廓的縱徑與橫徑比表達(dá)果實典型形狀特征,僅僅是刻畫了在輪廓曲線為橢圓的假設(shè)前提下長軸與短軸的長度比,其實質(zhì)是估計果實縱橫向大小尺寸關(guān)系。若用傅里葉描述子的前幾個諧波分量度量果實輪廓形狀,雖形狀表達(dá)細(xì)膩程度有所提高,但也僅是在三角度、方形度等形狀規(guī)則程度上的改進(jìn),果實輪廓的局部與整體彎曲信息表達(dá)粗糙不徹底。通過互相垂直的果梗面及側(cè)面投影分別提取果實在該2個投影面上輪廓的周長-面積分形維數(shù),一方面單個投影面上果實輪廓曲線形狀表達(dá)完整,覆蓋了輪廓曲線的各個部分;另一方面準(zhǔn)橢球狀果實立體形狀得以刻畫,且形狀描述數(shù)據(jù)僅2維,減少了計算機(jī)消耗。

4.2臍橙果實生長受光照、氣候條件等眾多因素的影響,果面著色并非均勻一致,存在色差。以果面顏色分量均值為指標(biāo)反映果實顏色特征,整體顏色特征得到表達(dá),但影響果實等級的小色塊、小疤痕被忽略。將果實主要色調(diào)區(qū)間等分,用各等分區(qū)間色調(diào)分形維數(shù)度量果實顏色,考慮了果面著色不均勻的情況,在色澤的描述上較顏色分量均值法精細(xì),但還缺少各色調(diào)的分布信息。用臍橙果實相對的2個側(cè)面30°—120°的4個色調(diào)區(qū)間單位坐標(biāo)化多重分形譜高度、寬度作為果實顏色特征值,一方面考慮了絕大部分果面的著色,避免了因采集果實多幅圖像而出現(xiàn)部分果面重復(fù)采圖與重復(fù)計算的現(xiàn)象,另一方面分段色調(diào)多重分形過程中像素的累計信息與分布信息得到丈量,覆蓋了影響果實等級的小色塊、小疤痕情況,顏色度量更為精細(xì)徹底。

4.3果實糖酸度無損檢測技術(shù)較多,近紅外光譜、激光、X射線及高光譜圖像技術(shù)都可應(yīng)用于果實內(nèi)部品質(zhì)的無損檢測,且有一定的精度,但機(jī)器視覺技術(shù)設(shè)備簡單、成本低、數(shù)據(jù)量少、計算機(jī)消耗小,不失為果實內(nèi)部品質(zhì)無損檢測的通用方法之一。果面分段色調(diào)單位坐標(biāo)化多重分形譜高度與寬度映射臍橙果實糖酸度因果面顏色刻畫精細(xì)標(biāo)準(zhǔn)差較低,分別在0.77和0.36以內(nèi),與實際值間的相關(guān)系數(shù)較高,分別在0.80和0.70以上,能基本確定果實糖酸度,表明該方法可用于臍橙果實糖酸度無損檢測。

4.4冰糖橙果實病蟲害眾多,雖各病蟲害為害狀具備典型特征,但如何用較少的數(shù)據(jù)以較全面地反映果實病蟲害為害狀這一缺陷典型特征沒有定論。病蟲害為害狀分形維數(shù)反映的是形狀信息,若結(jié)合為害狀的顏色(顏色分量均值)來識別病蟲害缺陷果,會因均值計算中為害點狀或線狀典型特征,被調(diào)和近乎忽略而收不到應(yīng)有的效果。綜合了果實病蟲害為害狀缺陷分段色調(diào)像素累計信息與形狀信息的多重分形譜方法、復(fù)雜性測度法不失為有效方法,尤其是復(fù)雜性測度法因色調(diào)分段細(xì)微(僅為1°),點狀、線狀缺陷被保留未被大色塊調(diào)和而展現(xiàn)出優(yōu)勢。

總之,柑橘形狀與顏色分級、糖酸度無損檢測、病蟲害果實缺陷識別等問題,因面對的是柑橘生物體而呈現(xiàn)出復(fù)雜性,用統(tǒng)計復(fù)雜性測度、分形維數(shù)、多重分形譜方法進(jìn)行復(fù)雜性問題的復(fù)雜性量測,較精確地反映了冰糖橙果實病蟲害缺陷典型特征、臍橙果實形狀、顏色特征和糖酸度無損檢測的映射參數(shù)特點。

參考文獻(xiàn)(References)：

[1]Clark J B, Jacques D R. Practical measurement of complexity in dynamic systems.ProcediaComputerScience, 2012,79(8):14-21.

[2]Muhammad N A. Complexity measurement for dealing with class imbalance in classification modeling. Wellington: Massey University, 2012.

[3]Hengameh R. Model complexity measurement. Sweden: Linkoping University, 2011.

[4]Santos C A, Urdaneta V, Jaimes G,etal. Ultrasonic spectral and complexity measurements on brine and oil saturated rocks.RockMechanicsandRockEngineering, 2010,43(3):351-359.

[5]Kumar A, Kumar A. Reengineering of module for public sector & complexity measurement.GlobalJournalofComputerScienceandTechnology, 2011,11(16):1-4.

[6]Lavazza L, Robiolo G. Functional complexity measurement: Proposals and evaluations//Lavazza L.TheSixthInternationalConferenceonSoftwareEngineeringAdvances. Nice, France: IARIA. 2011:257-261.

[7]Wen Z Y, Shen L M, Jing H P,etal. Color and shape grading of citrus fruit based on machine vision with fractal dimension//Tong Xiangrong.The2010 3rdInternationalCongressonImageandSignalProcessing. Yantai, Shandong: IEEE Xplore, 2010:898-903.

[8]溫芝元,曹樂平.基于補(bǔ)償模糊神經(jīng)網(wǎng)絡(luò)的臍橙不同病蟲害圖像識別.農(nóng)業(yè)工程學(xué)報,2012,28(11):152-157.

Wen Z Y, Cao L P. Image recognition of navel orange diseases and insect pests based on compensatory fuzzy neural networks.TransactionsoftheCSAE, 2012,28(11):152-157. (in Chinese with English abstract)

[9]溫芝元,曹樂平.柑橘病蟲害的機(jī)器識別.工程數(shù)學(xué)學(xué)報,2012,29(5):641-646.

Wen Z Y, Cao L P. Machine identification of disease and insect pest of citrus fruit.ChineseJournalofEngineeringMathematics, 2012,29(5):641-646. (in Chinese with English abstract)

[10]溫芝元,曹樂平.椪柑果實病蟲害的傅里葉頻譜重分形圖像識別.農(nóng)業(yè)工程學(xué)報,2013,29(23):159-165.

Wen Z Y, Cao L P. Citrus fruits diseases and insect pest recognition based on multifractal analysis of Fourier transform spectra.TransactionsoftheCSAE, 2013,29(23):159-165. (in Chinese with English abstract)

[11]盛鷗,嚴(yán)翔, 彭抒昂,等.紐荷爾臍橙果實發(fā)育期葉片不同形態(tài)硼含量與缺硼的關(guān)系.園藝學(xué)報,2007,34(5):1103-1110.

Sheng O, Yan X, Peng S A,etal. Relationship between boron deficiency occurrence and changes in contents of different boron forms in leaves of Newhall navel orange during fruit development.ActaHorticulturaeSinica, 2007,34(5):1103-1110. (in Chinese with English abstract)

[12]涂俊凡,吳黎明,彭抒昂. 柑橘油胞凹陷病組織結(jié)構(gòu)特征的研究.湖北農(nóng)業(yè)科學(xué),2010,49(7):1634-1637.

Tu J F, Wu L M, Peng S A. Histological characterization of rind oil spotting disease in citrus.HubeiAgriculturalSciences, 2010,49(7):1634-1637. (in Chinese with English abstract)

[13]路綱,周明天,唐勇,等.任意圖支配集精確算法回顧.計算機(jī)學(xué)報,2010,33(6):1073-1084.

Lu G, Zhou M T, Tang Y,etal. A survey on exact algorithms for dominating set related problems in arbitrary graphs.ChineseJournalofComputers, 2010,33(6):1073-1084. (in Chinese with English abstract)

[14]饒運(yùn)清,Efstathiou Janet.基于信息熵的制造系統(tǒng)復(fù)雜性測度及其在調(diào)度中的應(yīng)用.機(jī)械工程學(xué)報,2006,42(7):8-13.

Rao Y Q, Efstathiou J. Entropy-based measurement of manufacturing system complexity and its application in scheduling.ChineseJournalofMechanicalEngineering, 2006,42(7):8-13. (in Chinese with English abstract)

[15]何清華,羅嵐,陸云波,等.基于TO視角的項目復(fù)雜性測度研究.管理工程學(xué)報,2013,27(1):127-133.

He Q H, Luo L, Lu Y B,etal. Investigating project measurement complexity from TO perspectives.JournalofIndustrialEngineering/EngineeringManagement, 2013,27(1):127-133. (in Chinese with English abstract)

[16]曹樂平,溫芝元,沈陸明.基于色調(diào)分形維數(shù)的柑橘糖度和有效酸度檢測.農(nóng)業(yè)機(jī)械學(xué)報,2010,41(3):143-148.

Cao L P, Wen Z Y, Shen L M. Sugar content and the valid acidity test of the citrus based on the fractal dimensions of hue.TransactionsoftheCSAM, 2010,41(3):143-148. (in Chinese with English abstract)

[17]曹樂平.基于周長面積分形維數(shù)的柑橘品種機(jī)器識別.農(nóng)業(yè)工程學(xué)報,2010,26(2):351-355.

Cao L P. Machine recognition of citrus variety based on the fractal dimensions of perimeter-area.TransactionsoftheCSAE, 2010,26(2):351-355. (in Chinese with English abstract)

[18]Thiem S, Schreiber M. Partitioning schemes and non-integer box sizes for the box-counting algorithm in multifractal analysis.TheEuropeanPhysicalJournalB, 2013,86(48):1-10.

[19]Yi M, Kim K, Chang K H. Multifractal measures of ions absorbed on a charged lipid membrane.JournaloftheKoreanPhysicalSociety, 2013,62(5):845-848.

[20]Wang G, Ren D. Classification of surface electromyographic signals by means of multifractal singularity spectrum.Medical&BiologicalEngineering&Computing, 2013,51(3):277-284.

[21]Vasiljevic J, Reljin B, Sopta J,etal. Application of multifractal analysis on microscopic images in the classification of metastatic bone disease.BiomedMicrodevices, 2012,14(3):541-548.

[22]Zhang Z. Manufacturing complexity and its measurement based on entropy models.TheInternationalJournalofAdvancedManufacturingTechnology, 2012, 62(9/10/11/12):867-873.

[23]符玲,何正友,錢清泉.超高壓輸電線路的故障暫態(tài)特征提取及故障類型判斷.中國電機(jī)工程學(xué)報,2010,30(22):100-106.

Fu L, He Z Y, Qian Q Q. Feature extraction of fault transient and fault type determination for EHV transmission lines.ProceedingsoftheCSEE, 2010,30(22):100-106. (in Chinese with English abstract)

[24]劉樹成,魏巍,閆清東,等基于復(fù)雜性測度的變矩器流場仿真模型可信度分析.農(nóng)業(yè)機(jī)械學(xué)報,2012,43(10):19-24.

Liu S C, Wei W, Yan Q D,etal. Credibility analysis of simulation model for torque converters based on complexity measurement.TransactionsoftheCSAM, 2012,43(10):19-24. (in Chinese with English abstract)

[25]吳曉燕,吳靜,陳永興,等.基于復(fù)雜性測度的仿真模型驗證. 空軍工程大學(xué)學(xué)報:自然科學(xué)版,2013,14(2):28-31,36.

Wu X Y, Wu J, Chen Y X,etal. Validation of simulation models based on complexity measure.JournalofAirForceEngineeringUniversity:NaturalScienceEdition, 2013,14(2):28-31,36. (in Chinese with English abstract)

[26]Whitelock D P, Brusewitz G H, Stone M L. Apple shape and rolling orientation.AppliedEngineeringinAgriculture, 2006,22(1):87-94.

[27]RashidiM, Seyfi K. Classification of fruit shape in cantaloupe using the analysis of geometrical attributes.WorldJournalofAgriculturalSciences, 2007,3(6):735-740.

[28]應(yīng)義斌.水果形狀的傅里葉描述子研究.生物數(shù)學(xué)學(xué)報,2001,16(2):234-240.

Ying Y B. Fourier descriptor of fruit shape.JournalofBiomathematics, 2001,16(2):234-240. (in Chinese with English abstract)

[29]Blasco J, Aleixos N, Molto E. Machine vision system for automatic quality grading of fruit.BiosystemsEngineering, 2003,85(4):415-423.

[30]Alfatni M S M, Shariff A R M, Shafri H Z M,etal. Oil palm fruit bunch grading system using red,green and blue digital number.JournalofAppliedSciences, 2008,8(8):1444-1452.