基于迁移学习和数据增强技术的物种识别
|
摘要
为了应对动物保护工作者和普通民众希望对国内珍稀动植物准确识别并加以保护的场景,实验提出一种基于数据增强技术和迁移学习的深度学习方法,并借助App实现实时的珍稀动植物识别。首先使用迁移模型提取基于ImageN-et数据集的标准化瓶颈描述算子和所有卷积层神经网络的权重,然后通过网络爬虫搜集大量野生珍稀动植物图片,并对得到的数据集分别进行静态数据增强和动态数据增强,用来训练迁移模型新的特征表示。实验使用Softmax函数实现多物种分类。实验中,分别采用MobileNet、InpectionV3等不同模型的不同参数进行训练,得到不同配置下模型的准确率并加以比较,最后将其应用于真实自然场景,实现了90%以上的珍稀动植物精准分类,实验证明提出的方法具有很高准确率和良好的运行性能。
animals in China, proposes a deep learning method based on data enhancement technology and migration learning, and realizes real-timerare animal and plant identification by means of App. Firstly, the migration model is used to extract the weight of the standardized bottle-neck description operator and all bottleneck layer activation functions based on the ImageNet dataset. Then, a large number of wild rare ani-mal and plant images are collected through the web crawler, and the data set is separately subjected to static data enhancement and dynam-ic data enhancement. Used to train the new feature representation of the migration model. The experiment uses the Softmax function toachieve multi-species classification. In the experiment, different parameters of different models such as MobileNet and InpectionV3 areused to train the test, and the accuracy of the model under different configurations is tested. Finally, applies to the actual natural scene, re-alizes more than 90% of the rare and accurate real-time classification of rare animals and plants. The proposed method has high accuracyand good running performance.
animals in China, proposes a deep learning method based on data enhancement technology and migration learning, and realizes real-timerare animal and plant identification by means of App. Firstly, the migration model is used to extract the weight of the standardized bottle-neck description operator and all bottleneck layer activation functions based on the ImageNet dataset. Then, a large number of wild rare ani-mal and plant images are collected through the web crawler, and the data set is separately subjected to static data enhancement and dynam-ic data enhancement. Used to train the new feature representation of the migration model. The experiment uses the Softmax function toachieve multi-species classification. In the experiment, different parameters of different models such as MobileNet and InpectionV3 areused to train the test, and the accuracy of the model under different configurations is tested. Finally, applies to the actual natural scene, re-alizes more than 90% of the rare and accurate real-time classification of rare animals and plants. The proposed method has high accuracyand good running performance.
引文
[1]马克平,高吉喜,李琳,魏伟,徐靖,刘忆南,肖戈.中国面向2020后的生物多样性保护[J].全球治理与国家行动,2018(Z1):138-143.
[2]朱虎明,李佩,焦李成,杨淑媛,侯彪.深度神经网络并行化研究综述[J].计算机学报.2018,8:1861-1881.
[3]GE Yun,JIANG Shun-liang,XU Qing-yong,JIANG Chang-long,YE Fa-mao.Exploiting Representations from Pre-Trained Convolutional Neural Networks for High-Resolution Remote Sensing Image Retrieval[J].Multimedia Tools and Applications,2018.
[4]Mark Sandler,Andrew Howard,Menglong Zhu,Andrey Zhmoginov,Liang-Chieh Chen.Inverted Residuals and Linear Bottlenecks:Mobile Networks for Classification,Detection and Segmentation[J].Computer Vision and Pattern Recognition,2018.
[5]黄嘉祺.基于迁移学习的行人检测方法研究[D].南京理工大学,2015.
[6]吴斌,吉佳,孟琳,等.基于迁移学习的唐诗宋词情感分析[J].电子学报,2016,44(11):2780-2787.DOI:10.3969/j.issn.0372-2112.2016.11.030.
[7]程波.基于迁移学习的多模态脑图像分析及其应用研究[D].南京航空航天大学,2015.
[8]花君林.基于Selenium的Python网络爬虫的实现[J].电脑编程技巧与维护.2017,15:30-31+36.
[9]葛芸,江顺亮,叶发茂,许庆勇,唐祎玲.基于ImageNet预训练卷积神经网络的遥感图像检索[J].武汉大学学报(信息科学版),2018,43(01):67-73.
[10]李林,李小舜,吴少智.基于迁移学习和显著性检测的盲道识别[J].计算机工程与应用,2018,54(11):8-14+29.
[11]Navin Kumar Manaswi.CNN in Keras[M].Apress,2018.
[12]高友文,周本君,胡晓飞.基于数据增强的卷积神经网络图像识别研究[J].计算机技术与发展,2018,(8):62-65.
[13]袁文翠,孔雪.基于TensorFlow深度学习框架的卷积神经网络研究[J].微型电脑应用,2018,2:29-32.
[14]Neil Patrick,Del Gallego,Joel Ilao.Multiple-Image Super-Resolution on Mobile Devices:an Image Warping Approach[J].EURASIPJournal on Image and Video Processing,2017(1).
[15]张坤,章东平,杨力.样本增强的人脸识别算法研究[J].中国计量大学学报,2018,(2):194-199.
[16]葛婷,牟宁,李黎.基于Softmax回归与图割法的脑肿瘤分割算法[J].电子学报,2017,3:644-649.
[17]Matthew D.Zeiler,Rob Fergus.Visualizing and Understanding Convolutional Networks[C].New York University,USA,2014:818-833.
[18]虢齐.基于深度学习的图像生成技术研究与应用[D].电子科技大学,2017.
[19]唐珂.人工智能未来发展前景[J].人民论坛,2018,2:24-25.
[2]朱虎明,李佩,焦李成,杨淑媛,侯彪.深度神经网络并行化研究综述[J].计算机学报.2018,8:1861-1881.
[3]GE Yun,JIANG Shun-liang,XU Qing-yong,JIANG Chang-long,YE Fa-mao.Exploiting Representations from Pre-Trained Convolutional Neural Networks for High-Resolution Remote Sensing Image Retrieval[J].Multimedia Tools and Applications,2018.
[4]Mark Sandler,Andrew Howard,Menglong Zhu,Andrey Zhmoginov,Liang-Chieh Chen.Inverted Residuals and Linear Bottlenecks:Mobile Networks for Classification,Detection and Segmentation[J].Computer Vision and Pattern Recognition,2018.
[5]黄嘉祺.基于迁移学习的行人检测方法研究[D].南京理工大学,2015.
[6]吴斌,吉佳,孟琳,等.基于迁移学习的唐诗宋词情感分析[J].电子学报,2016,44(11):2780-2787.DOI:10.3969/j.issn.0372-2112.2016.11.030.
[7]程波.基于迁移学习的多模态脑图像分析及其应用研究[D].南京航空航天大学,2015.
[8]花君林.基于Selenium的Python网络爬虫的实现[J].电脑编程技巧与维护.2017,15:30-31+36.
[9]葛芸,江顺亮,叶发茂,许庆勇,唐祎玲.基于ImageNet预训练卷积神经网络的遥感图像检索[J].武汉大学学报(信息科学版),2018,43(01):67-73.
[10]李林,李小舜,吴少智.基于迁移学习和显著性检测的盲道识别[J].计算机工程与应用,2018,54(11):8-14+29.
[11]Navin Kumar Manaswi.CNN in Keras[M].Apress,2018.
[12]高友文,周本君,胡晓飞.基于数据增强的卷积神经网络图像识别研究[J].计算机技术与发展,2018,(8):62-65.
[13]袁文翠,孔雪.基于TensorFlow深度学习框架的卷积神经网络研究[J].微型电脑应用,2018,2:29-32.
[14]Neil Patrick,Del Gallego,Joel Ilao.Multiple-Image Super-Resolution on Mobile Devices:an Image Warping Approach[J].EURASIPJournal on Image and Video Processing,2017(1).
[15]张坤,章东平,杨力.样本增强的人脸识别算法研究[J].中国计量大学学报,2018,(2):194-199.
[16]葛婷,牟宁,李黎.基于Softmax回归与图割法的脑肿瘤分割算法[J].电子学报,2017,3:644-649.
[17]Matthew D.Zeiler,Rob Fergus.Visualizing and Understanding Convolutional Networks[C].New York University,USA,2014:818-833.
[18]虢齐.基于深度学习的图像生成技术研究与应用[D].电子科技大学,2017.
[19]唐珂.人工智能未来发展前景[J].人民论坛,2018,2:24-25.
© 2004-2018 中国地质图书馆版权所有 京ICP备05064691号 京公网安备11010802017129号 地址：北京市海淀区学院路29号 邮编：100083 电话：办公室：(+86 10)66554848；文献借阅、咨询服务、科技查新：66554700 |