APP下载

基于DeepLearning4J on Spark深度学习方法在药用植物图像识别中应用初探

2018-12-08赵彦辉范欣宁张建逵谢明

中国中医药图书情报 2018年5期
关键词:深度学习

赵彦辉 范欣宁 张建逵 谢明

摘要:目的 基于卷积神经网络,使药用植物的图像识别自动化、智能化,从而提高药用植物识别的效率。方法 使用中国植物图像库搜集药用植物图像,进行灰度化、角度随机旋转等批量处理。采用卷积神经网络图像识别技术,结合DeepLearning4j框架,对模型参数进行优化调试。结果 使用DL4J-example、AlexNet、VGG16模型,小样本单一背景药用植物图像识别正确率分别达到68.00%、70.00%、24.00%,大样本复杂背景药用植物图像识别正确率分别达到48.87%、60.00%、91.35%。 结论 图像识别技术的使用提高了复杂背景下药用植物图像识别的正确率和效率,为药用植物的识别提供了新方法,识别正确率高低与激励函数、迭代次数、卷积层数及参与训练的图像数量直接相关。

关键词:图像识别;卷积神经网络;深度学习;迁移学习;药用植物

DOI: 10.3969/j.issn.2095-5707.2018.05.005

中图分类号:G250.7 文献标识码:A 文章编号:2095-5707(2018)05-0018-05

Discussion on Application of Deep Learning Method in Medical Plant Image Recognition Based on DeepLearning4J on Spark

ZHAO Yan-hui1, FAN Xin-ning2, ZHANG Jian-kui2, XIE Ming2*

(1. Library of Liaoning University of Traditional Chinese Medicine, Shenyang 110032, China; 2. College of Pharmacy, Liaoning University of Traditional Chinese Medicine, Dalian 116600, China)

Abstract: Objective To automate and intelligentize medical plant image recognition based on the convolution neural network. To improve the efficiency of medical plant recognition. Methods Chinese plant image database was used to collect medical plant images, and the batch processing of gray scale and angle random rotation were carried out. Convolution neural network image recognition technology was used and DeepLearning4J frame was combined. Model parameters were optimized and debugged. Results DL4J-example, AlexNet and VGG16 models were used. The accuracy rates of image recognition for small samples of single background medical plants reached 68.00%, 70.00% and 24.00%, respectively, and the accuracy rates of image recognition of medical plants in complex background reached 48.87%, 60.00% and 91.35%, respectively. Conclusion The use of image recognition technology can improve accuracy and efficiency of medical plant image recognition under complex background, and provide a new method for recognition of medical plants. The recognition accuracy rate is directly related to the excitation function, the number of iterations, the number of convolution layers, and the number of images participating in the training.

Key words: image recognition; convolution neural network; deep learning; transfer learning; medical plants

基金項目:2014年辽宁省高等学校创新团队(WT2014004)

第一作者:赵彦辉,E-mail: 644735344@qq.com

*通讯作者:谢明,E-mail: x6m6@163.com

传统的药用植物识别主要依靠专业人员通过眼看、手摸、口尝、鼻闻的方式进行[1],这种方式主要依赖操作者的专业水平和经验,因此不同的操作者识别药用植物有一定的主观性,其正确率有所不同。而图像识别技术能解决人工识别主观性强等问题,可以降低药用植物识别的难度。图像识别是指利用计算机对图像进行处理、分析和理解,以识别各种不同模式的目标和对象的技术,是信息科学和人工智能的重要组成部分[2]。近年来,深度学习已经成为图像识别领域的热点。……

登录APP查看全文

猜你喜欢

深度学习
从合坐走向合学:浅议新学习模式的构建
面向大数据远程开放实验平台构建研究
基于自动智能分类器的图书馆乱架图书检测
搭建深度学习的三级阶梯
有体验的学习才是有意义的学习
电子商务中基于深度学习的虚假交易识别研究
利用网络技术促进学生深度学习的几大策略
MOOC与翻转课堂融合的深度学习场域建构
大数据技术在反恐怖主义中的应用展望
深度学习算法应用于岩石图像处理的可行性研究