基于高光谱成像的苹果病害无损检测方法  

Nondestructive Detection Method of Hyperspectral Imaging for Apple Disease

作　　者：刘思伽[1] 田有文[1] 冯迪[1] 张芳[1] 崔博[1] 

机构地区：[1]沈阳农业大学信息与电气工程学院,沈阳110161

出　　处：《沈阳农业大学学报》2016年第5期634-640,共7页Journal of Shenyang Agricultural University

基　　金：辽宁省大型仪器设备共享服务项目(LNDY201501003)

摘　　要：苹果果实易发生病害,传统的苹果病害的检测不适应苹果分级在线检测的要求。为了实现病害苹果快速、有效的在线检测,采用高光谱成像技术对寒富苹果的炭疽病、苦痘病、黑腐病和褐斑病的病害果进行无损检测研究。根据正常区域与病害区域光谱相对反射率差异,提出改进流形距离方法。综合计算病害与正常区域,病害与果梗/花萼区域,正常与果梗/花萼区域的光谱相对反射率的总改进流行距离L值,从而从全波段中选择了3个特征波段,分别为700,765,904nm。对700nm特征波段下的图像进行阈值分割,以此获得掩膜图像,并对掩膜后的图像二次阈值分割提取感兴趣区域。将3个特征波段下对应的光谱相对反射率分别组合,作为BP神经网络的输入矢量,检测苹果是否为病害果。结果表明:选择700nm与904nm波段下的光谱相对反射率为最佳组合,病害果的检测率达96.25%。说明高光谱成像技术所获得的2个特征波段可以有效对苹果病害进行检测,为开发多光谱成像的苹果品质在线检测和分级系统提供参考。Disease is easy to occur in apple fruit. Traditional detection of apple disease is not adapted to the requirement of apple grading on-line detection. In order to achieve the fast, effective online detection for the disease apple, hyperspectral imaging was adopted to study the nondestructive detection of the anthracnose, bitter pox disease and black fruit rot and leaf spot disease in Hanfu apple. According to the relative reflectance spectrum difference between disease area and normal area, the improved manifold distance method was proposed. The total improved manifold distance L value was comprehensive calculated by the relative reflectance spectra of the disease and normal area, disease with stem/calyx area, normal and stem/calyx area. So three feature wavelengths were selected respectively from the whole band wavelength, 700, 765, 904 nm. In order to get the mask image, the image of the characteristic wave band at 700 nm was threshold segmented. The interested area was extracted after secondary threshold segmentation of the mask image. The relative reflectance spectra of the three characteristic wave bands were combined, respectively, as the BP neural network input vector, to detect whether apple fruit was diseased. Finally, the relative reflectance spectra under 700 nm to 904 nm band were selected as the best combination by comparing the detection results. A recognition rate of the normal apples and diseased apples respectively were 96.25%. Results showed that the two characteristics of band obtained by hyperspectral imaging technology can effectively detect disease for apple and provide the reference for the development of multispectral imaging of apple’s quality detection and classification system.

关 键 词：改进流形距离 苹果病害 高光谱成像 无损检测 BP神经网络 

分 类 号：S436.611[农业科学—农业昆虫与害虫防治] TP391.4[农业科学—植物保护]