基于高光谱成像的苹果水心病无损检测

Nondestructive Detection of Apple Watercore Based on Hyperspectral Imaging

以240个“秦冠”苹果水心病果和好果为试材，采集9001700nm的近红外波段高光谱图像，选取高光谱图像中的苹果区域作为感兴趣区域（RDI）并计算其平均光谱，分别采用4种特征选择方法和3种核函数支持向量机（SVMD）分类器对水心病果进行判别，以探讨利用近红外高光谱成像技术判别苹果水心痛的可行性。结果表明：基于卡方检验和支持向量机递归消除（SVM-RFE）2种特征选择法优于基于F检验和决策树的方法。4种特征选择的3种核函数支持向量机（SVM）分类器在1～200个波段下对水心病果的判别正确率分别为：48．6％～70．2％、48．6％～72．0％、33．3％～71．8％、47．2％～70．8％。基于SVM-RFE检验的特征选择下，SVM对水心病果的正确识别率达到72．O％，为该试验选出的最优方法。

In order to evaluate the ability of near infrared hyperspectral imaging to detect watercore in apple fruits, hyperspectral images of 240 apples cv.‘Qinguan＇ including sound fruit and watercore fruit were collected by near infrared hyperspectral camera （900-1 700 nm）. The apple regions of hyperspectral images were extracted as region of interest （ROD in which its average spectrum was calculated. To recognize watercore fruits, 4 kinds of feature selection methods and 3 kinds of kernel function of support vector machine （SVM） classifier were adopted. The results showed that 2 kinds of feature selection which based on ehi-square test and support vector machine reeursive feature elimination （SVM-RFE） were superior to the methods of F classic test and decision tree. The accurate rate of watercore distinguish of 4 kinds of feature selection with 3 kinds of kernel function of SVM classifier at 1-200 wavebands was 48. 6%-70. 2% ,48. 6%- 72. 0 G, 33. 3 % -71.8 % and 47. 2 %- 70. 8 %, respectively;moreover, the accurate rate of watercore distinguish based on SVM-RFE,which was the best method,reached the highest level of 72. 0%.