基于卷积神经网络干瘪核桃X射线图像判别

Discrimination of shriveled walnut X‑ray image based on convolution neural network

核桃内部品质良莠不齐会使其市场利润降低,现有检测方式不仅劳动成本高,效率低,且无法对核桃内部干瘪程度进行判别,因此,迫切需要一种无损、快速、准确的检测方式及判别方法。采用X射线技术获取核桃内部图像,采用图像处理软件Photoshop对核桃与核桃仁投影面积进行比值计算,确定3类不同干瘪程度的核桃,分别为内部存在略干瘪、过干瘪的核桃与正常核桃,采用3种核桃构建干瘪核桃数据集。基于卷积神经网络(CNN)结构,利用Alexnet、视觉几何群网络(VGG16)、MobileNetV2与残差网络(ResNet50)分别构建核桃内部干瘪程度判别模型。根据3种模型对干瘪核桃数据集分类的预测损失值、预测准确率、测试准确率与Epoch均次用时进行性能分析确定最优模型,并进行参数优化。结果表明,MobileNetV2模型在学习率为0.0001,批处理为32时,网络性能最佳,预测准确率达98.65%,测试准确率为93.40%。

The difference of internal quality of walnut will reduce its market profit.The existing detection methods have high labor cost and low efficiency,as well as impossible to discriminate the shriveled walnuts with different degrees.Therefore,a non‑destructive,rapid and accurate detection method and a discriminated method are urgently needed to detect internal shriveled walnuts.The internal images of walnuts are obtained by using X‑ray technology,and the ratio of the projection area between the walnut and walnut kernel is calculated by employing Photoshop image processing software,three categories of walnuts with different degrees of shriveling are identified,which are slightly shriveled,overly shriveled and normal walnuts,respectively.A shriveled walnut dataset is constructed by using these three types of walnuts.Based on the convolutional neural network(CNN)structure,the discrimination models of walnut internal shriveling degree are constructed by using Alexnet,VGG16,MobileNetV2 and ResNet50.The optimal model is determined through performance analysis based on the prediction loss value,prediction accuracy rate,test accuracy rate and Epoch average duration of the three models on the shriveled walnut dataset,followed by parameter optimization.The results show that the MobileNetV2 model achieves the best network performance with a learning rate of 0.0001 and a batch size of 32,and with a prediction accuracy of 98.65%and a test accuracy of 93.40%.