基于RELIEF算法和极限学习机的苹果轻微损伤高光谱检测方法  

Detection method of slight bruises of apples based on hyperspectral imaging and RELIEF-extreme learning machine

作　　者：张萌 李光辉[1,2] ZHANG Meng;LI Guanghui(School of IoT Engineering,Jiangnan University,Wuxi 214000,Jiangsu,China;Engineering Research Center of IoT Technology Applications of Ministry of Education,Wuxi 214000,Jiangsu,China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214000 [2]物联网技术应用教育部工程技术研究中心,江苏无锡214000

出　　处：《浙江大学学报（农业与生命科学版）》2019年第1期126-134,共9页Journal of Zhejiang University:Agriculture and Life Sciences

基　　金：国家自然科学基金(61472368)

摘　　要：采用高光谱成像技术(400～1 000 nm)对苹果轻微损伤进行快速识别及无损检测。采集苹果正常及不同损伤时间的高光谱图像,选择图像中合适的区域作为感兴趣区域并提取平均光谱反射率及图像熵信息,将采集的样本按2∶1的比例分为训练集和测试集。使用RELIEF算法基于光谱平均反射率及图像熵信息提取了8个特征波段(17、30、35、51、61、66、94和120),分别基于全波段和特征波段进行极限学习机(extreme learning machine, ELM)建模分析,并与支持向量机(support vector machine, SVM)和K-均值聚类算法进行比较。结果表明,基于全波段的ELM模型最终测试集识别率为94.44%,基于特征波段的RELIEF-极限学习机(Re-ELM)模型识别率为96.67%,基于特征波段的Re-SVM及Re-K均值模型的最终测试集识别率分别为92.22%和91.67%,证实了Re-ELM是一种更为有效的苹果损伤分类判别方法。在此基础上,基于图像处理技术和特征波段提出了一种苹果轻微损伤高光谱检测算法,使用该算法针对特征波段进行独立成分分析(independent component analysis, ICA)变换,选取ICA第3成分图像进行自适应阈值分割,从而获得损伤图像。对全部高光谱图像进行检测表明,该算法的最终识别率超过94%,说明该算法能够较为有效地识别苹果损伤区域。In order to realize the rapid and nondestructive recognition of slight bruises of apples,a hyperspectral imaging technique(400-1 000 nm)was used.Hyperspectral images of sound and different damage time of Fuji apples were collected,and the average spectral reflectance and entropy were extracted from the region of interest(ROI)of the image.All the samples were divided into training set and test set(2∶1).The characteristic wave-bands extracted based on the spectral average reflectance and entropy using RELIEF algorithm were 17,30,35,51,61,66,94 and 120,respectively.Then,based on full wavebands and characteristic wavebands,an extreme learning machine(ELM)model was built,as comparison with support vector machine(SVM)and K-mean algorithm.The results showed that the recognition accuracy of ELM model for the test set based on the full wave-bands was 94.44%,and the accuracy of the Re-ELM model based on the characteristic wavebands was 96.67%,and the accuracy of the Re-SVM and Re-K mean models for the characteristic wavebands were 92.22%and 91.67%,respectively,which demonstrated that the Re-ELM was a more effective method for the bruise apple classification.Subsequently,an apple damage detection algorithm based on the characteristic wavebands and image processing was proposed,which performed an independent component algorithm(ICA)transformation of the characteristic wavebands,and selected the third component image of the ICA transformation,and used adaptive threshold segmentation to obtain the bruise area on apples.The final detection accuracy of apple damage detection algorithm based on the image processing technology was over 94%,which indicates that the algorithm is efficient for identifying slight bruises of apples.

关 键 词：苹果损伤 高光谱成像 无损检测 极限学习机 独立成分分析 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]