温室黄瓜光谱成像识别方法研究.docVIP

  温室黄瓜光谱成像识别方法研究# 袁挺，纪超，陈英，李伟，张俊雄** （中国农业大学 工学院） 5 10 15 摘要：提出了一种基于统计方差结合人工神经网络的光谱选择方法对黄瓜敏感波段进行分析 验证，并将选定的光谱组合作为温室黄瓜识别中光谱图像获取的参考依据，处理结果表明利 用所摄敏感波段的图像信息可有效地解决近色系目标与背景的区分问题。综合比较黄瓜作物 （果实、叶、花）在不同光谱域的分光反射特性差异，利用方差分解的思想获取果实信息的 敏感波段，在敏感区域内进行主成分分析，以前 4 个主成分作为网络输入，作物器官类别作 为输出，建立 3 层 LMBP 神经网络验证模型。160 个样本数据按比例分为建模集和预测集， 模型对建模集 120 个样本的正确判别率为 100%，对预测集 40 个样本的验证准确率为 95%。 说明敏感波段的选择能较好地反映黄瓜作物不同器官间的特性差异，为光谱成像技术的进一 步研究应用提供了理论根据。 关键词：农业工程；光谱成像；方差分析；神经网络；黄瓜 中图分类号：S126  Cucumber Recognition in Greenhouse Imaging  Based on Spectral 20 25 30 35 Yuan Ting, Ji Chao, Chen Ying, Li Wei, Zhang Junxiong (College of Engineering, China Agricultural University) Abstract: A new spectrum selected method was developed to analyze and verify sensitive bands of cucumber based on statistical variance analysis and artificial neural network. Then the selected spectrum composition was used as reference basis for spectral image acquisition in greenhouse cucumber recognition, and the results of image processing indicated that the images within sensitive bands were captured to cope with the similar-color segmentation problem under complex environment effectively. By comparing the spectral reflectance difference of cucumber plant (fruit, leaf and flower) from visible to infrared region (350 nm-1200 nm), sensitive bands of fruit information were obtained by statistical variance analysis. After that, principal component analysis compressed the sensitive bands into several new variables that were the linear combination of original spectral data. In order to set up the three layer verifying model of back propagation artificial neural network (BP-ANN), the first four principal components were applied as inputs of BP-ANN, and the values of type of cucumber organs were applied as outputs. In this model, the trained network arrives at a 100% identification rate for 120 training samples as well as a 95% identification rate for 40 test samples. It proved that the selected spectrum composition could better re