應用非破壞性影像處理技術於蓮霧品質檢測

Applying Non-destructive Image Processing Technique on Wax Apple Quality Inspection

胡恆驥(Heng-Ji Hu)

指導教授 ： 蔡正發

國立屏東科技大學/管理學院/資訊管理系所/碩士(2016年)

摘要

消費者於購買農產品時，常需要仰賴所謂專業人員（產農、經銷商、零售商…）的協助，¬而專業人員對於農產品的分類，往往決定了售價之高低；久而久之，於消費市場上，價格便成了判斷農產品品質好壞的最重要依據之一。
因此本研究藉由影像辨識方法建立分類模型，提供蓮霧品質分類結果，做為消費者選購時之參考依據。
以屏東縣內埔鄉冬季生產之黑珍珠蓮霧為例，請種植蓮霧多年之專家，將蓮霧品質依五大分類表徵：賣相、糖度、脆度、口感、成熟度等進行感官品評，後利用eHSI色彩空間配合門檻值進行蓮霧影像外觀紋理分割，以灰階共生矩陣提取特徵值，最終以支持向量機進行蓮霧品質辨識，經綜合評比各項統計數值後，尋找最適分類方法。
經實驗發現，將專家分類之結果以五等第區分，並與影像辨識結果相互比較，各表徵分類正確率約介於35%~45%之間，倘增加分類容許誤差值±1，正確率更可達70%~80%。若再將影像辨識結果與糖度計分類結果進行比較，在容許誤差值±1的情況下，分類準確率更可高達97.93%。

關鍵字

eHSI ； 灰階共生矩陣 ； 支持向量機

並列摘要

While purchasing agricultural products, consumers usually need opinions from professionals, such as production farmer, dealer and retailer. And the ranking category of agricultural products made by professionals will determine its selling price. Thus, the price has become one of the most important factors to inspect the quality of agricultural products in consumer market.
The aim of this research is to establish a classification model based on image recognition to assist wax apple buyers in quality inspection.
Use wax apples from Neipu township in Pingtung county as experimental subjects. The wax apples are classified by experts according to its five characters: (1) look (2) brix (3) crunchiness (4) texture and (5) maturity. We then manipulate Hue-Saturation-Intensity (eHSI) color space with different thresholds to proceed features segmentation with the exterior images of the wax apples, and extract eigenvalues based on gray level co-concurrent matrix. Finally, we inspect the wax apples by the trained support vector machine (SVM) to evaluate the applicability of the five characterization in classification.
The results of our experiments indicates our method has significant accuracy in brix classification. We have ranked the apples by its characters separately into five groups. The accuracy is between 35-45% for individual character classification. The accuracy can reach 70-80% if classification tolerance values ±1. Moreover, if we compare the classification result of brix spindle with our method, the accuracy attains to 97.93%.

並列關鍵字

eHSI ； gray co-occurrence matrix ； SVM

參考文獻

[12] 謝清祿、鄭聖夫、林達德（1997），應用影像紋理分析及類神經網路辨識甘藍種苗之生長階段．農業機械學刊，6（2），1-13。

[7] 陳志杰（2005），應用影像處理於蓮霧外部品質檢測之應用研究（碩士論文），國立中興大學生物產業機電工程學系。

[11] 錢均亮（2011），創新色彩空間的研究與應用（博士論文）．國立中央大學資訊工程學系。

[13] 鍾穎慧（2011），不同影像尺寸與不同特徵表達對影像辨識之影響（碩士論文）．國立中央大學資訊管理學系。

[16] Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine learning, 20(3), 273-297.