基于CNN神经网络的小麦不完善粒高光谱检测

发布时间：2018-06-28 12:48

本文选题：小麦 + 不完善粒　；参考：《食品科学》2017年24期

【摘要】：利用高光谱成像技术对小麦不完善粒进行无损检测。以932个小麦为样本,其中正常粒样本486个、破损粒样本170个、虫蚀粒样本149个及黑胚粒样本127个为研究对象,通过高光谱图像采集系统采集样本的光谱信息,然后从每个样本的116个波段中选取30个波段,建立基于深度学习的卷积神经网络(convolutional neural networks,CNN)模型。实验中的CNN采用2个卷积层,第1层采用大小为3×3的32个卷积核,第2层采用大小为5×5的64个卷积核,池化层采用最大池,激活函数采用修正线性单元,为避免过拟合,在全连接层后面接入dropout层,参数设置为0.5,其他卷积参数均为默认值,得到校正集总识别率为100.00%,测试集总识别率为99.98%。最后,以支持向量机(support vector machine,SVM)为基线模型进行对比,从116个波段中选取90个波段进行建模,测试集总识别率为94.73%。通过实验对比可以看出,CNN模型比SVM模型识别率高。研究表明CNN模型能够实现对小麦不完善粒的准确、快速、无损检测。
[Abstract]:The nondestructive testing of wheat imperfect grains was carried out by using hyperspectral imaging technique. The spectral information of 932 wheat samples was collected by hyperspectral image acquisition system, including 486 normal grain samples, 170 damaged grain samples, 149 wormwood samples and 127 black embryo samples. Then, 30 bands are selected from 116 bands of each sample, and a convolution neural network (convolutional neural) model based on deep learning is established. In the experiment, CNN uses 2 convolution layers, the first layer uses 32 convolution cores with size 3 脳 3, the second layer uses 64 convolution cores with size 5 脳 5, the pool layer adopts the largest pool, and the activation function adopts modified linear unit to avoid overfitting. When the dropout layer is connected behind the full connection layer, the parameters are set to 0.5, the other convolution parameters are all the default values, the total recognition rate of the correction set is 100.00000 and the total recognition rate of the test set is 99.98. Finally, the support vector machine (support vector machine) is used as the baseline model, 90 bands are selected from 116 bands to model, and the total recognition rate of the test set is 94.73. Through the comparison of experiments, we can see that CNN model has higher recognition rate than SVM model. The results show that CNN model can be used for accurate, fast and nondestructive detection of wheat imperfect grains.
【作者单位】：北京工商大学计算机与信息工程学院食品安全大数据技术北京市重点实验室;
【基金】：土壤植物机器系统技术国家重点实验室开放课题(2014-SKL-05) 北京工商大学两科基金培育项目(LKJJ2015-22)
【分类号】：TP183;TP391.41

【相似文献】