基于改进GhostNet V2的轻量化熊蜂图像分类模型

doi:10.3969/j.issn.1004-1524.20240190

浙江农业学报 ›› 2024, Vol. 36 ›› Issue (12): 2832-2845.DOI: 10.3969/j.issn.1004-1524.20240190

基于改进GhostNet V2的轻量化熊蜂图像分类模型

范为培¹^,²^,³(), 于晓明¹, 沈凤龙¹^,^*(), 王亮²^,³, 王星⁴

1.辽东学院信息工程学院,辽宁丹东 118003
2.沈阳化工大学计算机科学与技术学院,辽宁沈阳 110142
3.沈阳化工大学辽宁省化工过程工业智能化技术重点实验室,辽宁沈阳 110142
4.辽东学院农学院,辽宁丹东 118003

收稿日期:2024-03-01 出版日期:2024-12-25 发布日期:2024-12-27
作者简介:范为培(1996—),男,江苏盐城人,硕士,主要研究方向为图像处理与深度学习。E-mail:z2021394@stu.syuct.edu.cn
通讯作者: *沈凤龙,E-mail:shenlu-2000@126.com
基金资助:
2022年辽宁省应用基础研究计划项目(2022JH2/101300149)

A lightweight bumblebee image classification model based on improved GhostNet V2

FAN Weipei¹^,²^,³(), YU Xiaoming¹, SHEN Fenglong¹^,^*(), WANG Liang²^,³, WANG Xing⁴

1. School of Information Engineering, Liaodong University, Dandong 118003, Liaoning, China
2. School of Computer Science and Technology, Shenyang University of Chemical Technology, Shenyang 110142, China
3. Key Laboratory of Chemical Process Industry Intelligent Technology of Liaoning Province, Shenyang University of Chemical Technology, Shenyang 110142, China
4. School of Agriculture, Liaodong University, Dandong 118003, Liaoning, China

Received:2024-03-01 Online:2024-12-25 Published:2024-12-27

摘要/Abstract

摘要：

为准确、快速地实现熊蜂的自动分拣,提出了一种轻量化深度学习熊蜂图像分类模型。首先,采集了地熊蜂的蜂王、雄蜂和工蜂图像1 742张,并通过数据增强构建了包含13 117张熊蜂图像的数据集BumblebeeImage。然后,以GhostNet V2模型为基础,通过多尺度卷积获取输入图像更多感受野下的特征信息,增加两条捷径分支分别将低层与中层、高层的特征融合,将ReLU激活函数替换为SiLU,删减bottleneck层数和通道数,设计了一种轻量化熊蜂图像分类模型GMCFF。结果表明,利用GMCFF模型对BumblebeeImage数据集进行分类的准确率达到了98.40%,较原模型提高了1.53百分点,与ShuffleNetV2和MobileNetV2模型的分类准确率对比也更高,分别提高了1.53百分点和1.15百分点。该模型参数量只有0.73 M,浮点运算量较改进前下降了25.15 M,模型大小仅有3.01 MB,单张熊蜂图像的平均测试时间为17.08 ms,满足轻量化与实时性的要求。

关键词: 熊蜂分类, GhostNet V2, 轻量化, 多尺度卷积, 特征融合

Abstract:

In order to realize automatic sorting of bumblebees accurately and quickly, a lightweight deep learning bumblebee image classification model is proposed. Firstly, 1 742 images of queen, drone and worker bees of ground bumblebee were collected, and a dataset containing 13 117 bumblebee images was constructed by data augmentation. Then, based on the GhostNet V2 model, the feature information of the input image under more receptive fields was obtained by multi-scale convolution, two shortcut branches were added to fuse the features of the low-level and the middle-level, the high-level respectively, the ReLU activation function was replaced by SiLU, and the number of bottleneck layers and channels was cut, and a lightweight bumblebee image classification model GMCFF was designed. The results showed that the classification accuracy of the BumblebeeImage dataset using the GMCFF model reached 98.40%, which was 1.53 percentage points higher than the original model, and the classification accuracy compared with the ShuffleNetV2 and MobileNetV2 model was also higher, increasing by 1.53 and 1.15 percentage points respectively. The number of parameters of the model was only 0.73 M, the amount of floating point operations was decreased by 25.15 M compared with that before the improvement, the model size was only 3.01 MB, and the average test time of a single bumblebee image was 17.08 ms, which met the requirements of lightweight and real-time.

Key words: bumblebee classification, GhostNet V2, lightweight, multi-scale convolution, feature fusion

中图分类号:

TP391.4

范为培, 于晓明, 沈凤龙, 王亮, 王星. 基于改进GhostNet V2的轻量化熊蜂图像分类模型[J]. 浙江农业学报, 2024, 36(12): 2832-2845.

FAN Weipei, YU Xiaoming, SHEN Fenglong, WANG Liang, WANG Xing. A lightweight bumblebee image classification model based on improved GhostNet V2[J]. Acta Agriculturae Zhejiangensis, 2024, 36(12): 2832-2845.

图/表 22

图1 熊蜂图像采集系统模型图

Fig.1 Model diagram of bumblebee image acquisition system

图2 熊蜂图像采集系统

Fig.2 Bumblebee image acquisition system

图3 熊蜂图像

Fig.3 Bumblebee image

表1 BumblebeeImage数据集分布情况

Table 1 Distribution of BumblebeeImage dataset

熊蜂类别 Bumblebee category	训练集 Training set	验证集 Validation set	测试集 Test set	总计 Total
蜂王Queen bee	3 011	861	430	4 302
雄蜂Drone bee	3 130	895	447	4 472
工蜂Worker bee	3 040	869	434	4 343
总计Total	9 181	2 625	1 311	13 117

图4 图像数据增强示例

Fig.4 Example of image data augmentation

图5 GMCFF网络结构示意图 GV2-bottleneck,GhostNetV2瓶颈层;Conv2d,二维卷积;AvgPool,平均池化;FC,全连接;F,低层特征;B1,捷径分支1;B2,捷径分支2。

Fig.5 Schematic diagram of GMCFF network structure GV2-bottleneck, GhostNetV2 bottleneck; Conv2d, Convolution 2D; AvgPool, Average Pooling; FC, Fully Connected; F, Low-level feature; B1, Shortcut branch 1; B2, Shortcut branch 2.

表2 GMCFF网络结构参数

Table 2 Parameters of GMCFF network structure

输入Input	操作Operator	增大比例Exp	输出通道数Out	通道注意力SE	步长Stride
224×224×3	Conv2d 3×3、5×5、7×7	—	16	—	2
122×122×16	GV2-bneck	16	16	—	1
122×122×16	GV2-bneck	48	24	—	2
56×56×24	GV2-bneck	72	24	—	1
56×56×24	GV2-bneck	72	40	0.25	2
28×28×40	GV2-bneck	120	40	0.25	1
28×28×40	GV2-bneck	240	80	—	2
14×14×80	GV2-bneck	200	80	—	1
14×14×80	GV2-bneck	184	112	0.25	1
14×14×112	GV2-bneck	672	112	0.25	1
14×14×112	GV2-bneck	160	160	0.25	2
7×7×160	Conv2d 1×1	—	160	—	1
7×7×160	AvgPool 7×7	—	480	—	-
1×1×480	Conv2d 1×1	—	480	—	1
1×1×480	FC	—	3	—	-

图6 DFC注意力机制

Fig.6 Decoupled fully connected attention mechanism

图7 GhostNetV1 bottleneck和GhostNetV2 bottleneck DWConv,深度卷积;BN,批量归一化;ReLU,修正线性单元;DFC,解耦全连接;FC,全连接。

Fig.7 GhostNetV1 bottleneck and GhostNetV2 bottleneck DWConv, Depthwise convolution;BN, Batch normalization; ReLU, Rectified linear unit; DFC, Decoupled fully connected; FC, fully connected.

图8 多尺度卷积

Fig.8 Multi-scale convolution

图9 特征融合

Fig.9 Feature fusion

图10 GMCFF模型损失值和准确率曲线

Fig.10 The curve of loss value and accuracy of GMCFF model

表3 数据增强试验对比

Table 3 Comparison of data augmentation experiments %

数据增强 Data augmentation	准确率 Accuracy	精确率 Precision	召回率 Recall	F1分数 F1-score
未使用Unused	82.53	83.99	82.49	82.83
使用Used	98.40	98.42	98.41	98.41

图11 ReLU和SiLU的函数图像对比

Fig.11 Comparison of function images between ReLU and SiLU

表4 不同激活函数的模型试验对比

Table 4 Comparison of model experiments with different activation functions %

激活函数 Activation function	准确率 Accuracy	精确率 Precision	召回率 Recall	F1分数 F1-score
ReLU	97.48	97.49	97.49	97.49
LeakyReLU	97.48	97.51	97.50	97.50
H-swish	97.48	97.49	97.49	97.49
SiLU	98.40	98.42	98.41	98.41

表5 GMCFF模型消融试验

Table 5 Ablation experiment of GMCFF model

模型 Model	改进方式Improvement mode			准确率 Accuracy/%	F1分数 F1- score/%	参数量 Parameters/ M	浮点运算量 FLOPs/M
模型 Model	多尺度卷积 Multi-scale convolution	特征融合 Feature fusion	架构精简 Streamlined architecture	准确率 Accuracy/%	F1分数 F1- score/%	参数量 Parameters/ M	浮点运算量 FLOPs/M
GhostNet V2	×	×	×	96.87	96.89	4.88	184.50
	√	×	×	98.02	98.03	5.03	229.06
	×	√	×	97.56	97.57	4.90	194.82
	×	×	√	96.49	96.50	0.71	103.54
	√	√	×	97.78	97.80	5.06	239.38
	√	×	√	97.25	97.26	0.82	148.10
	×	√	√	96.57	96.58	0.75	113.86
GMCFF	√	√	√	98.40	98.41	0.73	159.35

图12 中间层特征图

Fig.12 Feature map of the middle layer

表6 不同模型的性能对比

Table 6 Performance comparison of different models

模型 Model	准确率 Accuracy/%	精确率 Precision/%	召回率 Recall/%	F1分数 F1-score/%	参数量 Parameters/M	浮点运算量 FLOPs/M	模型大小 Model size/MB
DenseNet121	97.71	97.74	97.71	97.72	6.96	2 895.99	27.1
RegNet	94.66	94.68	94.69	94.68	2.32	207.34	9.00
ShuffleNetV2	96.87	96.90	96.89	96.89	1.26	151.69	4.95
MobileNetV2	97.25	97.29	97.26	97.26	2.23	326.27	8.73
MobileNetV3_small	96.26	96.36	96.30	96.29	1.52	61.17	5.92
ResNet18	97.03	97.04	97.05	97.04	11.18	1 823.52	42.7
EfficientNetB0	97.64	97.65	97.64	97.65	4.01	411.55	15.5
GhostNet	97.25	97.26	97.27	97.26	3.91	154.58	15.1
GhostNetV2	96.87	96.90	96.90	96.89	4.88	184.50	19.1
GMCFF	98.40	98.42	98.41	98.41	0.73	159.35	3.01

图13 BeeOrWasp数据集图像示例

Fig.13 Example of images from the BeeOrWasp dataset

图14 在BeeOrWasp数据集上损失值和准确率曲线对比图

Fig.14 Curve comparison chart of loss values and accuracy on the BeeOrWasp dataset

表7 在BeeOrWasp数据集上各评价指标对比

Table 7 Comparison of various evaluation indicators on the BeeOrWasp dataset

模型 Model	准确率 Accuracy/%	精确率 Precision/%	召回率 Recall/%	F1分数 F1-score/%	参数量 Parameters/M	浮点运算量 FLOPs/M	模型大小 Model size/MB
GhostNet V2	93.90	93.92	93.92	93.92	4.88	184.50	19.1
GMCFF	93.94	93.97	93.96	93.96	0.73	159.35	3.01

图15 不同熊蜂在实际环境下的分类测试结果 A, 蜂王,分类预测值95.87%;B,雄蜂,分类预测值94.58%;C,工蜂,分类预测值94.49%。

Fig.15 Classification test results of different bumblebees in actual environment A. Queen bee, with a classification prediction value of 95.87%; B. Drone bee, with a classification prediction value of 94.58%; C. Worker bee, with a classification prediction value of 94.49%.

参考文献 25

[1]	周峰, 姚丽媛, 石涵, 等. 传粉熊蜂访花行为的研究进展[J]. 昆虫学报, 2023, 66(3):419-438.
	ZHOU F, YAO L Y, SHI H, et al. Research progress in foraging behavior of pollinating bumblebees[J]. Acta Entomologica Sinica, 2023, 66(3):419-438. (in Chinese with English abstract)
[2]	THENMOZHI K, SRINⅣASULU REDDY U. Crop pest classification based on deep convolutional neural network and transfer learning[J]. Computers and Electronics in Agriculture, 2019, 164:104906.
[3]	DE NART D, COSTA C, DI PRISCO G, et al. Image recognition using convolutional neural networks for classification of honey bee subspecies[J]. Apidologie, 2022, 53(1):5.
[6]	甘雨, 郭庆文, 王春桃, 等. 基于改进EfficientNet模型的作物害虫识别[J]. 农业工程学报, 2022, 38(1):203-211.
	GAN Y, GUO Q W, WANG C T, et al. Recognizing crop pests using an improved EfficientNet model[J]. Transactions of the Chinese Society of Agricultural Engineering, 2022, 38(1):203-211. (in Chinese with English abstract)
[4]	陈彦彤, 陈伟楠, 张献中, 等. 基于深度卷积神经网络的蝇类面部识别[J]. 光学精密工程, 2020, 28(7):1558-1567.
	CHEN Y T, CHEN W N, ZHANG X Z, et al. Fly facial recognition based on deep convolutional neural network[J]. Optics and Precision Engineering, 2020, 28(7):1558-1567. (in Chinese with English abstract)
[7]	阮炬全, 刘朔. 基于高阶残差和注意力机制的轻量型作物害虫识别[J]. 计算机系统应用, 2023, 32(3):104-115.
	RUAN J Q, LIU S. Lightweight recognition of crop pests based on high-order residual and attention mechanism[J]. Computer Systems and Applications, 2023, 32(3):104-115. (in Chinese with English abstract)
[5]	陈俭. 基于卷积神经网络和度量学习的害虫检测方法研究[D]. 杭州: 浙江大学, 2021:1-30.
	CHEN J. Research on pest detection method based on convolutional neural network and metric learning[D]. Hangzhou: Zhejiang University, 2021:1-30. (in Chinese with English abstract)
[8]	彭红星, 徐慧明, 刘华鼐. 融合双分支特征和注意力机制的葡萄病虫害识别模型[J]. 农业工程学报, 2022, 38(10):156-165.
	PENG H X, XU H M, LIU H N. Model for identifying grape pests and diseases based on two-branch feature fusion and attention mechanism[J]. Transactions of the Chinese Society of Agricultural Engineering, 2022, 38(10):156-165. (in Chinese with English abstract)
[9]	MARTINEAU C, CONTE D, RAVEAUX R, et al. A survey on image-based insect classification[J]. Pattern Recognition, 2017, 65:273-284.
[10]	TANG Y, HAN K, GUO J, et al. GhostNetV2: enhance cheap operation with long-range attention[C]// NeurIPS 2022:36th Conference on Neural Information Processing Systems. New Orleans, USA: ACM Press, 2022:64879.
[11]	HAN K, WANG Y H, TIAN Q, et al. GhostNet:more features from cheap operations[C]// 020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 13-19, 2020, Seattle, WA, USA. IEEE, 2020:1577-1586.
[12]	于明, 李若曦, 阎刚, 等. 基于颜色掩膜网络和自注意力机制的叶片病害识别方法[J]. 农业机械学报, 2022, 53(8):337-344.
	YU M, LI R X, YAN G, et al. Crop diseases recognition method via fusion color mask and self-attention mechanism[J]. Transactions of the Chinese Society for Agricultural Machinery, 2022, 53(8):337-344. (in Chinese with English abstract)
[13]	SANDLER M, HOWARD A, ZHU M L, et al. MobileNetV2:inverted residuals and linear bottlenecks[C]// 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. June 18-23, 2018, Salt Lake City, UT, USA. IEEE, 2018:4510-4520.
[14]	张善文, 邵彧, 齐国红, 等. 基于多尺度注意力卷积网络的作物害虫检测[J]. 江苏农业学报, 2021, 37(3):579-588.
	ZHANG S W, SHAO Y, QI G H, et al. Crop pest detection based on multi-scale convolutional network with attention[J]. Jiangsu Journal of Agricultural Sciences, 2021, 37(3):579-588. (in Chinese with English abstract)
[15]	尹群杰, 杨文柱, 冉梦影, 等. 结合多特征融合与残差空洞卷积的小目标检测[J]. 计算机工程与设计, 2022, 43(9):2622-2630.
	YIN Q J, YANG W Z, RAN M Y, et al. Small object detection using multi-feature fusion and residual dilated convolution[J]. Computer Engineering and Design, 2022, 43(9):2622-2630. (in Chinese with English abstract)
[16]	王继霄, 李阳, 王家宝, 等. 基于SqueezeNet的轻量级图像融合方法[J]. 计算机应用, 2020, 40(3):837-841.
	WANG J X, LI Y, WANG J B, et al. Light-weight image fusion method based on SqueezeNet[J]. Journal of Computer Applications, 2020, 40(3):837-841. (in Chinese with English abstract)
[17]	LIU W, WU G, REN F. Deep multi-branch fusion residual network for insect pest recognition[J]. IEEE Transactions on Cognitive and Developmental Systems, 2020, 13(3):705-716.
[18]	翟肇裕, 曹益飞, 徐焕良, 等. 农作物病虫害识别关键技术研究综述[J]. 农业机械学报, 2021, 52(7):1-18.
	ZHAI Z Y, CAO Y F, XU H L, et al. Review of key techniques for crop disease and pest detection[J]. Transactions of the Chinese Society for Agricultural Machinery, 2021, 52(7):1-18. (in Chinese with English abstract)
[19]	徐静萍, 王芳. 基于改进的S-ReLU激活函数的图像分类方法[J]. 科学技术与工程, 2022, 22(29):12963-12968.
	XU J P, WANG F. Image classification method based on improved S-ReLU activation function[J]. Science Technology and Engineering, 2022, 22(29):12963-12968. (in Chinese with English abstract)
[20]	李好, 邱卫根, 张立臣. 改进ShuffleNet V2的轻量级农作物病害识别方法[J]. 计算机工程与应用, 2022, 58(12):260-268.
	LI H, QIU W G, ZHANG L C. Improved ShuffleNet V2 for lightweight crop disease identification[J]. Computer Engineering and Applications, 2022, 58(12):260-268. (in Chinese with English abstract)
[21]	HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[J]. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). July 21-26, 2017, Honolulu, HI, USA. IEEE, 2017:2261-2269.
[22]	RADOSAVOVIC I, KOSARAJU R P, GIRSHICK R, et al. Designing network design spaces[J]. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 13-19, 2020, Seattle, WA, USA. IEEE, 2020:10425-10433.
[23]	MA N N, ZHANG X Y, ZHENG H T, et al. ShuffleNet V2:practical guidelines for efficient CNN architecture design[M]// Lecture notes in computer science. Cham: Springer International Publishing, 2018:122-138.
[24]	HOWARD A, SANDLER M, CHEN B, et al. Searching for MobileNetV3[C]// 2019 IEEE/CVF International Conference on Computer Vision (ICCV). October 27-November 2, 2019, Seoul, Korea (South). IEEE. 2019:1314-1324.
[25]	TAN M, LE Q. Efficientnet:Rethinking model scaling for convolutional neural networks[C]// Proceedings of the 36th International Conference on Machine Learning (ICML). Long Beach, California,USA:PMLR,s 2019(97):6105-6114.

基于改进GhostNet V2的轻量化熊蜂图像分类模型

A lightweight bumblebee image classification model based on improved GhostNet V2

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 22

参考文献 25

相关文章 4

编辑推荐

Metrics

本文评价

[1]	李大华, 孔舒, 李栋, 于晓. 基于改进SSD模型的柑橘叶片病害轻量化检测模型[J]. 浙江农业学报, 2024, 36(3): 662-670.
[2]	张宁, 吴华瑞, 韩笑, 缪祎晟. 基于多尺度和注意力机制的番茄病害识别方法[J]. 浙江农业学报, 2021, 33(7): 1329-1338.
[3]	张满囤, 单新媛, 于洋, 米娜, 阎刚, 郭迎春. 基于小波变换和改进KPCA的奶牛个体识别研究[J]. 浙江农业学报, 2017, 29(12): 2000-2008.
[4]	杨春合，江朝晖*，杨宝华，陈祎琼，刘连忠. 基于移动互联网图像处理模式的作物叶片含水量检测研究[J]. 浙江农业学报, 2015, 27(10): 1835-.