摘要
目的:利用计算机深度学习实现对中药饮片二维图像的自动化识别的研究具有重要实用价值,可广泛应用于医疗、生产和教学等领域。既往多采用传统的提取图像中的底层特征的方法来进行识别,然而这种方法不能在复杂背景的图像条件下给出鲁棒的识别结果。因此,中药饮片图像识别需要更高级别的图像表达方法。方法:构建包含50种常见中药饮片图像数据库,共2 554张图像,作为模型的训练与测试对象,并运用Softmax损失训练卷积神经网络。结果:卷积神经网络在所有测试的50种中药饮片图像中可以实现70%的平均识别精度。结论:卷积神经网络在多个饮片相互遮挡并带有复杂背景情况下较为理想,未来具有一定应用前景。
It is of great importance that deep learning of computer for the automate identification of the two-dimensional image of Chinese herbal slices is valuable in the application to medicine, production and education. Traditional methodsusually extract low-level image features for the identification, but they cannot give robust recognition results undercomplex backgrounds. Therefore, higher level image representation is necessary in the image identification. A publicChinese herbal medicine database was constructed with 50 common categories and 2,554 images in total, for training andevaluating our recognition model. Then, the softmax loss function was adopted to train the convolutional neural networkmodel. As a result, the convolutional neural network can achieve the average precision of 70% under all the 50 medicineherbal classes. In conclusion, convolutional neural network can obtain good results in image identification with complexbackgrounds and mutually occluded herbal slices, which has promising potential for future applications.
出处
《世界科学技术-中医药现代化》
CSCD
2017年第2期218-222,共5页
Modernization of Traditional Chinese Medicine and Materia Medica-World Science and Technology
基金
北京中医药大学研究生自主课题项目(2016-JYB-XS026):基于卷积神经网络的中药饮片图像识别与检索
主持人:孙鑫
关键词
中药饮片
图像识别
深度学习
卷积神经网络
Chinese herbal slices
image recognition
deep learning
convolutional neural network
