可见光谱图像联合区间的黄瓜白粉病分割与检测  

Segmentation and Detection of Cucumber Powdery Mildew Based on Visible Spectrum and Image Processing

作　　者：白雪冰 余建树 傅泽田[1] 张领先[1] 李鑫星[1] BAI Xue-bing;YU Jian-shu;FU Ze-tian;ZHANG Ling-xian;LI Xin-xing(Beijing Laboratory of Food Quality and Safety,College of Information and Electrical Engineering,China Agricultural University, Beijing 100083,China)

机构地区：[1]中国农业大学信息与电气工程学院食品质量与安全北京实验室

出　　处：《光谱学与光谱分析》2019年第11期3592-3598,共7页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金面上项目(31271618)资助

摘　　要：黄瓜白粉病是黄瓜病害中爆发频率较高的一种,传播速度极快,常常导致产量降低,无法获得预期的经济效益。特别是在病害爆发严重的年份,黄瓜减产量高达20%。提出了一种基于可见光谱图像联合区间的偏最小二乘回归判别模型(SI-PLSR),用于黄瓜白粉病无损检测。采用Canon EOS 800D和Ocean Optics USB2000+光纤光谱仪采集了200个黄瓜白粉病感病叶片的可见光谱图像和反射率曲线。首先,采用基于小波降噪和H分割的分水岭分割算法从实时采集的黄瓜白粉病感病叶片可见光谱图像中提取目标叶片;其次,通过高斯拟合优化的Otsu算法分割目标叶片的可见光谱图像,获取白粉病病斑;然后,对350~1 100nm全波段光谱反射率曲线建立偏最小二乘回归模型并计算交叉验证均方根误差RMSECV,同时将全波段等分为20个子区间,分别建立偏最小二乘回归模型,选取RMSECV小于全波段反射率曲线建模RMSECV的子区间组成联合区间;最后,将光谱联合区间与白粉病病斑分割结果融合建立SI-PLSR模型。从实验结果可知,感病目标叶片的提取成功率高达94.00%,200幅感病叶片可见光谱图像中成功提取188幅,其中157幅目标叶片的完整性参数高于95%,31幅目标叶片完整性参数在90%~95%之间。188幅目标叶片的病斑分割结果显示,平均错分率为5.81%,其中平均False negative为1.55%,平均False positive为4.26%。对20个子区间分别建立偏最小二乘回归模型发现,第5,6,7,11,12,13和19子区间的RMSECV值小于全波段光谱反射率曲线建模的RMSECV值,说明这7个子区间的光谱信息对白粉病的判别有较大的贡献,这与呈现波峰的470~520,530~580和700~780nm波段相对应,因此选取这7个子区间的光谱反射率曲线建立联合区间。对联合区间建立SI-PLSR模型,其主成分数为7,校正集和验证集的相关系数和标准误差分别是0.975 2,0.907 3和0.919 5,1.091。与全波段PLSR模型相比,SI-PLSRPowdery mildew,as a kind of cucumber disease with high outbreak frequency,spreads very fast,often leads to yield reduction and can’t achieve the expected economic benefits.Especially in serious years of disease outbreak,the reduction of cucumber in some areas was as high as 20%.This paper proposed a subinterval interval partial least squares regression(SI-PLSR) based on visible spectrum image for cucumber powdery mildew non-destructive detection.We usedCanon EOS 800Dand Ocean Optics USB2000+optical fiber spectrometer to collect visible spectral images and reflectivity curves of 200cucumber powdery mildew leaves.Firstly,we used wavelet transform and watershed algorithm to extract the target leaves from the real-timevisible spectral images of cucumber powdery mildew leaves.Secondly,The Otsu algorithm optimized by Gauss fitting was used to segment the powdery mildew lesion.Thirdly,we established the PLSR in 350~1 100nm band and calculated the cross validation root-mean-square error(RMSECV).At the other hand,350~1 100nm was divided into 20sub-intervals,and established the PLSRindependently.The sub-intervals of RMSECVsmaller than the full band were selected to form the joint interval.Finally, the SI-PLSR model was established based on powdery mildew lesions images and joint interval.Results show that 188target leaves were extracted from 200susceptible leaves visible spectral images successfully of which 157were more than 95%and 31 were between 90%and 95%.The success rate was 94.00%.The average misclassification rate of powdery mildew was 5.81%. The average false negative was 1.55%and the average false positive was 4.26%.PLSR was established for 20sub-intervals, and the results showed that the RMSECVvalues of the 5,6,7,11,12,13and 19sub-intervals were lower than those of the fullband modeling,indicating that the spectral information of these seven sub-intervals contributed greatly to the identification of powdery mildew,which was relative to the wavebands of 470~520,530~580and 700~780nm showing peaks.Therefore, these

关 键 词：可见光谱 联合区间 偏最小二乘回归模型 计算机视觉 

分 类 号：O433.4[机械工程—光学工程]