基于BS-ResNeXt-50的密云地区野生动物图像识别

doi:10.11707/j.1001-7488.LYKX20220378

摘要/Abstract

摘要：

目的: 以卷积神经网络为基础，对现有的网络结构进行改进，实现对红外相机拍摄的野生动物图像进行种类自动识别。方法: 构建从北京市密云区的北京市级雾灵山自然保护区2014—2015年期间采集到的8类2172张野生动物图像数据集，并使用随机增强策略从14个增强方案中随机选择增强方式，为图像数据添加噪声。使用SENet及BlurPool构建了基于ResNeXt-50的改进网络：增强特征提取的SE-ResNeXt-50、维持平移不变性的BP-ResNeXt-50、结合二者的BS-ResNeXt-50，并在自建数据集上测试了不同固定学习率、分段学习率及余弦退火学习率对BS-ResNeXt-50网络准确率的影响。使用VGG16、ResNeXt-50、EfficientNet-B0、InceptionV3、DenseNet-121、BS-ResNeXt-50网络在CCT公开野生动物数据集中常见的16个类别图像上进行训练，并对单一物种的识别准确率进行比较。结果: SE-ResNeXt-50和BP-ResNeXt-50准确率分别达到了75.16%±0.14%和73.74%±0.13%。融合SENet以及BlurPool的改进方案BS-ResNeXt-50在自建数据集上测试的准确率达到78.04%±0.11%，为最优改进方案。使用余弦退火学习率后，BS-ResNeXt-50的准确率提升至81.54%，比固定学习率提升了3.5%；分段学习率准确率达到79.3%，与余弦退火学习率相差2.24%。在CCT数据集中BS-ResNeXt-50的识别准确率可达95.07%，比ResNeXt-50准确率高出1.95%，同时也高于VGG16的85.5%、EfficientNet-B0的90.23%、InceptionV3的91.38%以及DenseNet-121的93.3%准确率，并在各单一类别的预测准确率也均高于上述模型。单一类别的识别中除数量最少的类别外，BS-ResNeXt-50在其他类别识别准确率均高于90%，最高类别准确率达到98.6%。结论: 改进后的BS-ResNeXt-50模型相比ResNeXt-50可以更准确地完成对野生动物图像的识别任务，在不同的野生动物图像数据集上也具有较好的泛化能力。

关键词: 野生动物图像, 物种识别, 深度学习, 卷积神经网络

Abstract:

Objective: In the wild environment, the background of wildlife images captured by camera traps is complex, which poses a challenge for identifying wild animals in images with a large number of images and a wide variety of wildlife species. Based on convolutional neural network, this research aims to improve the existing structure and so as to implement the automatic recognition for wildlife images. Method: In this study, 2 712 wildlife images of 8 categories were taken from Wuling Mountain Beijing Nature Reserve, Miyun Districts, Beijing. The Auto Augment policy was randomly selected from 14 augmentation policies to add noise to the images. SENet and BlurPool were used to construct an improved network based on ResNeXt-50: SE-ResNeXt-50 for enhancement feature extraction, BP-ResNeXt-50 for Shift-invariance maintenance, and BS-ResNeXt-50 for both. The influences of fixed learning rate, segmented learning rate, and cosine annealing learning rate on the accuracy of the BS ResNeXt-50 network were tested on the self-built dataset. VGG16, ResNeXt-50, EfficientNet-B0, InceptionV3, DenseNet-121, and BS-ResNeXt-50 were used to train on 16 common categories of images in CCT public wildlife dataset, and the recognition accuracy of single species was compared.e influences of fixed learning rate, segmented learning rate, and cosine annealing learning rate on the accuracy of the BS ResNeXt-50 network were tested on the self-built dataset. VGG16, ResNeXt-50, EfficientNet-B0, InceptionV3, DenseNet-121, and BS-ResNeXt-50 were used to train on 16 common categories of images in CCT public wildlife dataset, and the recognition accuracy of single species was compared.eXt-50 is used to test influence of different learning rate include fixed and CosineAnnealing learning rate on collected dataset. VGG16, ResNeXt-50, EfficientNet-B0, InceptionV3, DenseNet-121, BS-ResNeXt-50 were used for training on CCT dataset, and the recognition accuracy of single species was compared. on ResNeXt-50: SE-ResNeXt-50 for enhancement feature extraction, BP-ResNeXt-50 for Shift-invariance maintenance, and BS-ResNeXt-50 for both. The influences of fixed learning rate, segmented learning rate, and cosine annealing learning rate on the accuracy of the BS ResNeXt-50 network were tested on the self-built dataset. VGG16, ResNeXt-50, EfficientNet-B0, InceptionV3, DenseNet-121, and BS-ResNeXt-50 were used to train on 16 common categories of images in CCT public wildlife dataset, and the recognition accuracy of single species was compared.eXt-50 is used to test influence of different learning rate include fixed and CosineAnnealing learning rate on collected dataset. VGG16, ResNeXt-50, EfficientNet-B0, InceptionV3, DenseNet-121, BS-ResNeXt-50 were used for training on CCT dataset, and the recognition accuracy of single species was compared. Result: The accuracy of SE-ResNeXt-50 and BP-ResNeXt-50 reached 75.16%±0.14% and 73.74%±0.13%, respectively. The enhanced scheme BS-ResNeXt-50, which integrated SENet and BlurPool, achieved an accuracy of 78.04%±0.11% when tested on a self built dataset, which was the best improved scheme. When the cosine annealing learning rate is used, the accuracy of BS-ResNeXt-50 was improved to 81.54%, which was 3.5% higher than that with the constant learning rate. The step decay learning rate achieved 79.3% accuracy, which was 2.24% less than the cosine annealing learning rate. The classification accuracy of BS-ResNeXt-50 was able to reach 95.07%, which was 1.95% higher than that of ResNeXt-50 on CCT dataset. At the same time, it was also 85.5% higher than that of VGG16, 91.38% higher than that of EfficientNet-B0, 91.38% higher than that of InceptionV3 and 93.3% higher than that of DenseNet-121. The prediction accuracy of each single category was also higher than that of the above model. In the recognition of a single category, except for the least one category, the accuracy of BS-ResNeXt-50 was 90% higher than that in other categories, and the highest category accuracy was 98.6%. Conclusion: The BS-ResNeXt-50 can more accurately complete the recognition task, and also has good generalization ability on different datasets.

Key words: wildlife images, species recognition, deep learning, convolutional neural network

中图分类号:

TP391.4

齐建东,马鐘添,张德怀,田赟. 基于BS-ResNeXt-50的密云地区野生动物图像识别[J]. 林业科学, 2023, 59(8): 112-122.

Jiandong Qi,Zhongtian Ma,Dehuai Zhang,Yun Tian. Wildlife Image Recognition in Miyun District Based on BS-ResNeXt-50[J]. Scientia Silvae Sinicae, 2023, 59(8): 112-122.

图/表 17

图1

图2

表2

图3

图4

表1

图5

图6

图7

图8

图9

图10

表3

图11

图12

图13

表4

参考文献 0

	何　嘉. 2019. 基于深度学习的野生动物智能检测与识别. 深圳: 深圳大学.
	He J. 2019. Wildlife smart detection and recognition based on deep learning. Shenzhen: Shenzhen University.［in Chinese］
	刘文定, 李安琪, 张军国, 等. 基于 ROI-CNN 的赛罕乌拉国家级自然保护区陆生野生动物自动识别. 北京林业大学学报, 2018, 40 (8): 123- 131.
	Liu W D, Li A Q, Zhang J G, et al. Automatic identification method for terrestrial wildlife in Saihanwula National Nature Reserve in Inner Mongolia of northern China based on ROI-CNN. Journal of Beijing Forestry University, 2018, 40 (8): 123- 131.
	邱志斌, 石大寨, 况燕军, 等. 基于深度迁移学习的输电线路涉鸟故障危害鸟种图像识别. 高电压技术, 2021, 47 (11): 1- 15. doi: 10.13336/j.1003-6520.hve.20210063
	Qiu Z B, Shi D Z, Kuang Y J, et al. Image recognition of harmful bird species related to transmission line outages based on deep transfer learning. High Voltage Engineering, 2021, 47 (11): 1- 15. doi: 10.13336/j.1003-6520.hve.20210063
	束祖飞, 何文通, 李　健, 等. 2021. 广东车八岭国家级自然保护区生物多样性综合监测信息服务平台. 数据与计算发展前沿, 3(6): 127−141.
	Shu Z F, He W T, Li J, et al. A comprehensive monitoring information service platform for biodiversity at Guangdong Chebaling National Nature Reserve. Frontiers of Data and Computing, 3(6): 127−141.［in Chinese］
	王柯力, 袁红春. 2018. 基于迁移学习的水产动物图像识别方法. 计算机应用, 38(5): 1304−1308, 1326.
	Wang K L, Yuan H C. 2018. Aquatic animal image classification method based on transfer learning. Journal of Computer Applications, 38(5): 1304−1308, 1326.［in Chinese］
	杨国国, 鲍一丹, 刘子毅. 基于图像显著性分析与卷积神经网络的茶园害虫定位与识别. 农业工程学报, 2017, 33 (6): 156- 162.
	Yang G G, Bao Y D, Liu Z Y. Location and recognition of tea pests based on image saliency analysis and convolution neural network. Transactions of the Chinese Society of Agricultural Engineering, 2017, 33 (6): 156- 162.
	于莉莉. 2017. 陆生野生动物保护对生物多样性的影响机理及对策. 南京: 南京林业大学.
	Yu L L. 2017. Effects of terrestrial wildlife conservation on biodiversity and countermeasures. Nanjing: Nanjing Forestry University.［in Chinese］
	张　毓, 高雅月, 常峰源, 等. 小样本条件下基于数据扩充和ResNeSt的雪豹识别. 北京林业大学学报, 2021, 43 (10): 89- 99.
	Zhang M, Gao Y Y, Chang F Y, et al. Panthera unica recognition based on data expansion and ResNeSt with few samples. Journal of Beijing Forestry University, 2021, 43 (10): 89- 99.
	Azulay A, Weiss Y. 2018. Why do deep convolutional networks generalize so poorly to small image transformations? arXiv: 1805.12177.
	Cubuk E D, Zoph B, Shlens J, et al. 2020. Randaugment: Practical automated data augmentation with a reduced search space. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 702−703.
	Gomez V A, Salazar A, Vargas F. Towards automatic wild animal monitoring: Identification of animal species in camera-trap images using very deep convolutional neural networks. Ecological Informatics, 2017, 41, 24- 32. doi: 10.1016/j.ecoinf.2017.07.004
	Guo Y, Rothfus T A, Ashour A S, et al. Varied channels region proposal and classification network for wildlife image classification under complex environment. IET Image Processing, 2019, 14 (4): 585- 591.
	He K, Zhang X, Ren S, et al. 2016. Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770−778.
	Hu J, Shen L, Sun G. 2018. Squeeze-and-excitation networks. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 7132−7141.
	Huang G, Liu Z, Van D M L, et al. 2017. Densely connected convolutional networks. Proceedings of the IEEE conference on computer vision and pattern recognition, 4700−4708.
	Krizhevsky A, Sutskever I, Hinton G E. 2012. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25.
	Loshchilov I, Hutter F. 2016. Sgdr: Stochastic gradient descent with warm restarts. arXiv: 1608.03983.
	Norouzzadeh M S, Nguyen A, Kosmala M, et al. Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning. Proceedings of the National Academy of Sciences, 2018, 115 (25): E5716- E5725.
	Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 2015, 28, 91- 99.
	Simonyan K, Zisserman A. 2014. Very deep convolutional networks for large-scale image recognition. arXiv: 1409.1556.
	Szegedy C, Liu W, Jia Y, et al. 2015. Going deeper with convolutions. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 1−9.
	Szegedy C, Vanhoucke V, Ioffe S, et al. 2016. Rethinking the inception architecture for computer vision. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2818−2826.
	Tan M, Le Q. 2019. Efficientnet: Rethinking model scaling for convolutional neural networks. International Conference on Machine Learning, 6105−6114.
	Timm M, Maji S, Fuller T. 2018. Large-scale ecological analyses of animals in the wild using computer vision. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 1896−1898.
	Willi M, Pitman R T, Cardoso A W, et al. Identifying animal species in camera trap images using deep learning and citizen science. Methods in Ecology and Evolution, 2019, 10 (1): 80- 91. doi: 10.1111/2041-210X.13099
	Xie J, Li A, Zhang J, et al. An integrated wildlife recognition model based on multi-branch aggregation and squeeze-and-excitation network. Applied Sciences, 2019, 9 (14): 2794. doi: 10.3390/app9142794
	Xie S, Girshick R, Dollár P, et al. 2017. Aggregated residual transformations for deep neural networks. Proceedings of The IEEE Conference on Computer Vision and Pattern Recognition, 1492−1500.
	Zhang R. Making convolutional networks shift-invariant again. International Conference on Machine Learning, 2019, 97, 7324- 7334.

物种 Species	图像数量 Image number
负鼠 Opossum	16479
郊狼 Coyote	16306
兔 Rabbit	12315
鹿 Deer	12191
浣熊 Raccoon	10888
鸟 Bird	9501
山猫 Bobcat	7597
猫 Cat	5163
松鼠 Squirrel	4430
啮齿动物（不含松鼠） Rodent (Squirrel exclusion)	4273
奶牛 Cow	3626
狗 Dog	3608
狐狸 Fox	2574
臭鼬 Skunk	1892

物种 Species	图像数量 Image number
狍 Roe Deer	500
鸟 Bird	500
山羊 Goat	332
松鼠 Squirrel	215
猪獾 Hog Badger	190
兔 Rabbit	155
豹猫 Leopard Cat	145
野猪 Boar	135

物种 Species	模型准确率Model accuracy（%）
物种 Species	ResNeXt-50	SE-ResNeXt-50	BP-ResNeXt-50	BS-ResNeXt-50
猪獾 Hog badger	65.2	68.9	67.4	73.2
鸟 Bird	81.5	84.7	85.3	88.7
野猪 Boar	66.4	71.3	67.1	72.5
豹猫 Leopard cat	63.7	65.4	63.7	70.2
狍 Roe deer	86.6	91.3	89.2	93.2
山羊 Goat	78.9	80.5	80.2	83.6
兔子 Rabbit	66.8	68.5	69.1	70.4
松鼠 Squirrel	66.2	70.7	67.9	72.5

物种 Species	模型准确率 Model accuracy（%）
物种 Species	VGG16	EfficientNet-B0	InceptionV3	DenseNet-121	ResNeXt-50	BS-ResNeXt-50
鸟 Bird	89.0	93.8	96.2	95.9	96.4	96.5
山猫 Bobcat	84.3	87.3	88.8	91.2	91.0	95.0
猫 Cat	86.8	92.5	93.2	95.6	94.0	97.6
奶牛 Cow	92.3	95.5	96.5	96.7	97.3	97.3
郊狼 Coyote	86.4	91.4	92.5	93.7	92.3	95.5
鹿 Deer	90.6	95.2	98.1	96.3	98.0	98.6
狗 Dog	80.1	88.5	90.2	93.0	92.9	93.3
狐狸 Fox	85.5	89.1	90.7	90.2	88.7	92.8
负鼠 Opossum	90.2	86.4	85.7	94.9	93.0	96.1
兔 Rabbit	83.0	90.4	92.3	93.7	93.2	94.5
浣熊 Raccoon	85.7	91.2	90.2	92.8	92.5	95.0
啮齿动物（不含松鼠） Rodent (Squirrel exclusion)	85.9	88.2	90.5	94.8	94.3	95.6
臭鼬 Skunk	78.8	83.3	84.0	85.0	86.9	88.9
松鼠 Squirrel	78.4	90.5	90.5	93.1	93.2	94.4

[1]	苗宇杰,祝诗平,普京,李俊贤,马羚凯,黄华. 基于卷积神经网络的家具木材图像种类识别[J]. 林业科学, 2023, 59(8): 133-140.
[2]	毛英伍,郭颖,张王菲,苏勇,关塬. 联合LiDAR、高光谱数据及3D-CNN方法的树种分类[J]. 林业科学, 2023, 59(3): 73-83.
[3]	冯慧敏,金崑. 基于卷积神经网络的雄性海南长臂猿声纹识别[J]. 林业科学, 2023, 59(1): 119-127.
[4]	齐建东,谭新新. 长白山红松阔叶林的净碳交换变化及基于时间卷积神经网络的模拟[J]. 林业科学, 2022, 58(2): 1-12.
[5]	李贾,兰岚,张佐忠,袁文涛,高德民,宗树琴,业巧林. 基于深度学习的森林可燃物含水率反演技术[J]. 林业科学, 2022, 58(10): 47-58.
[6]	何拓,刘守佳,陆杨,张永刚,焦立超,殷亚方. iWood: 基于卷积神经网络的濒危珍贵树种木材自动识别系统[J]. 林业科学, 2021, 57(9): 152-159.
[7]	赵子宇,杨霄霞,郭慧,葛浙东,周玉成. 基于卷积神经网络模型的木材宏、微观辨识方法[J]. 林业科学, 2021, 57(6): 134-143.
[8]	周焱,刘文萍,骆有庆,宗世祥. 基于深度学习的小目标受灾树木检测方法[J]. 林业科学, 2021, 57(3): 98-107.
[9]	宿恒硕,吕军,丁志平,唐彦杰,陈旭东,周强,张哲宇,姚青. 基于改进残差神经网络的木材识别算法[J]. 林业科学, 2021, 57(12): 147-154.
[10]	刘璇昕,孙钰,崔剑,蒋琦,陈志泊,骆有庆. 钻蛀性害虫取食声音的人工智能早期识别[J]. 林业科学, 2021, 57(10): 93-101.
[11]	谢锦莹,丁丽霞,王志辉,刘丽娟. 基于FCN与面向对象的滨海湿地植被分类[J]. 林业科学, 2020, 56(8): 98-106.
[12]	郭颖,李增元,陈尔学,张旭,赵磊,陈艳,王雅慧. 一种改进的高空间分辨率遥感影像森林类型深度学习精细分类方法:双支FCN-8s[J]. 林业科学, 2020, 56(3): 48-60.
[13]	邵明伟,董军宇. 基于深度学习的木材优选锯视觉检测算法[J]. 林业科学, 2020, 56(12): 123-129.
[14]	赵霖,张晓丽,吴艳双,张斌. 面向机载高光谱数据的3D-CNN亚热带森林树种分类[J]. 林业科学, 2020, 56(11): 97-107.
[15]	张广群,李英杰,汪杭军,周厚奎. 基于微调CaffeNet的林业图像分类[J]. 林业科学, 2020, 56(10): 121-128.