基于Xception与迁移学习的中药饮片图像识别研究

Research on traditional Chinese medicine piece image recognition based on Xception and transfer learning

为实现对常用60种中药饮片图像的精准快速识别,构建13 088张常见中药饮片图像数据集,采用迁移学习的方式,以深度学习算法中Xception卷积神经网络模型为基准,对饮片图像进行训练与识别。模型训练的初始学习率设置为0.01,优化器中设置Nesterov动量超参数为0.9,训练次数为100轮,得到在训练集上的分类准确率达到100%,验证集准确率为97.42%,测试集准确率为97.26%,最后结合混淆矩阵这一指标对模型的识别能力进行评估分析。该模型与传统依靠提取中药饮片图像特征的机器学习算法相比,分类效果更好,泛化能力更强。

In order to achieve accurate and fast recognition of images of 60 kinds of the traditional Chinese medicine pieces,a dataset of 13088 images of common traditional Chinese medicine pieces is built,and the transfer learning method is used to train and recognize the images of the traditional Chinese medicine pieces on the basis of the Xception convolutional neural network model in the depth learning algorithm.The initial learning rate for the model training is set as 0.01,the Nesterov momentum hyperparameter in the optimizer is set as 0.9,and the training times are set as 100.The results are obtained as follows.The accuracy of the training set is 100%,the accuracy of the verification set is 97.42%,and the accuracy of the test set is 97.26%.The recognition ability of the model is evaluated and analyzed by combining the confusion matrix.In comparison with the traditional machine learning algorithms that rely on extracting image features of the traditional Chinese medicine pieces,the proposed model has better classification effect and stronger generalization ability.