Abstract: Aiming at the problem of low manual classification labor productivity and classification accuracy of betel nut, a betel nut classification model based on convolutional neural network was constructed. A betel nut image feature extraction network composed of two convolution blocks was designed, each with two convolutional layers and one pooling layer respectively. The second convolutional block connected a multilayer perceptron composed of flat layer, hidden layer and output layer. The feature extraction network and perceptron formed a shallow convolutional neural network (SCNN). A comparative analysis of network training, validation and testing was completed through six optimization methods such as SGD, Adam and etc., and manual adjustment of the learning rate. Adam was selected as the optimization method for SCNN and the learning rate was set to 0.001. The SCNN based on Adam optimization was applied to the betel nut classification, the classification accuracy reached 98.05%,  and the AUC value of the network ROC curve was equal to 1. The grading of betel nut could be completed quickly and more accurately by using the SCNN model, which provides the basis for the betel nut machine grading research.

Key words: betel nut, convolutional neural network, Adam optimizer, grading, ROC curve

摘要： 针对槟榔的人工分级劳动生产率和分级准确率低的问题，构建基于卷积神经网络的槟榔分级模型。设计1个由两个卷积块构成的槟榔图像特征提取网络，每个卷积块有2个卷积层和1个池化层。第2个卷积块联接1个由扁平层、隐含层和输出层构成的多层感知器，特征提取网络和感知器构成1个浅层卷积神经网络(SCNN)。通过SGD、Adam等6种优化方法和学习率的人工调整完成网络训练、验证和测试的对比分析，选择Adam为SCNN的优化方法并将学习率设置为0.001。将基于Adam优化的SCNN应用于槟榔分级，分级准确率达到98.05%，网络ROC曲线的AUC值等于1。利用SCNN模型能够快速、准确地完成对槟榔的分级，为槟榔机器分级研究提供基础。

关键词: 槟榔, 卷积神经网络, Adam优化器, 分级, ROC曲线