基于卷积神经网络与迁移学习的玉米籽粒图像分类识别

Classification and Recognition of Corn Kernel Image Based on Convolution Neural Network and Transfer Learning

种子是农业生产最基本,最主要的生产资料。为实现玉米种子的快速鉴定识别与保护,本文提出基于卷积神经网络(Convolution Neural Network,CNN)与迁移学习相结合的玉米种子籽粒图像分类识别方法,可将预训练的CNN模型参数迁移到玉米籽粒图像分类识别任务中。实验采集了6种玉米种子籽粒图像双面图像共1976张,包括16DX531,京粘1号,科诺58,铁研,小金黄,郑单958,建立胚面,胚乳面和双面混合的3种数据集。按照7∶2∶1的比例随机划分训练集,验证集和测试集,并对训练集图像作数据增强处理。基于4种CNN模型Xception,ResNet50V2,MobileNetV2,DenseNet121进行参数迁移学习。实验结果表明Xception与胚乳数据集建模方法优于其他方法。Xception--胚乳模型训练集与验证集平均识别准确率分别为95.55%和95.97%,测试集准确率为92.78%。基于卷积神经网络与迁移学习相结合的玉米籽粒图像识别方法切实可行。

Seeds are the most basic and main means of production in agricultural production.In order to realize the rapid identification,recognition and protection of corn seeds,a corn kernel image classification and recognition method based on convolution neural network(CNN)and transfer learning was proposed in this paper,it can transfer the pre-trained CNN model parameters to the corn kernel image classification and recognition task.A total of 1976 double-sided images of six kinds of corn grain images were collected,including 16DX531,Jingjian 1,Kenuo 58,Tieyan,Xiaojinhuang and Zhengdan 958.Three data sets of embryo surface,endosperm and double-sided mixture were established.The training set,verification set and test set are randomly divided according to the ratio of 7∶2∶1,and the images of the training set were enhanced.Parameter migration learning was based on four CNN models:Xception,ResNet50V2,MobileNetV2 and DenseNet121.The experimental results indicated that Xception and endosperm dataset modeling method were better than other methods.The average recognition accuracy of Xception-endosperm model training set and validation set was 95.55%and 95.97%,respectively,and that of test set was 92.78%.The corn kernel image recognition method based on convolution neural network combined with migration learning was feasible.