基于Slim-YOLOv4与嵌入式设备的无人机检测

doi:10.16180/j.cnki.issn1007-7820.2023.05.009

摘要/Abstract

摘要：

为了在资源受限的嵌入式设备上实现对无人机的实时检测,文中提出了一种基于YOLOv4的轻量化检测网络,即Slim-YOLOv4。该网络选用Ghostnet替换原YOLOv4的主干特征提取部分,将特征融合部分中的3×3卷积替换为深度可分离卷积,并对深度可分离卷积中的激活函数进行优化,以减少网络的参数量、计算量,加快网络的收敛。实验结果表明Slim-YOLOv4的准确率达到91.6%,与原YOLOv4相比损失了1.6%,但是原YOLOv4的权重文件高达250 MB。在不影响鲁棒性的前提下,Slim-YOLOv4的权重文件大小仅为42 MB,且优于Faster-RCNN模型的108 MB和Mobilenetv3模型的53 MB。新方法每秒处理的图片数量在PC上达到31.2 frames·s^-1,在嵌入式设备上高达37.6 frames·s^-1,证明可以将其部署到嵌入式设备上对无人机进行实时检测。

关键词: 轻量化网络, YOLOv4, 无人机检测, 嵌入式平台, 深度学习, Ghostnet, ELU, 实时检测

Abstract:

In order to realize real-time detection of UAV on resource-constrained embedded devices, this study proposes a lightweight detection network, Slim-YOLOv4, based on YOLOv4. In this network, Ghostnet is used to replace the main feature extraction part of the original YOLOv4, the 3×3 convolution in the feature fusion part is replaced by the deep detachability convolution and the activation function in the deep detachability convolution is optimized to reduce the number of parameters and computation of the network and accelerate the convergence of the network. The experimental results show that the accuracy of the Slim-YOLOv4 is 91.6%, which is 1.6% lower than that of the original YOLOv4, but the weight file of original YOLOv4 is up to 250 MB. On the premise of not affecting the robustness, the weight file size of Slim-YOLOv4 is only 42 MB, which is better than the 108 MB of Fan-RCNN model and 53 MB of Mobilenetv3 model. The number of images processed per second by the proposed method reaches 31.2 frames·s^-1 on PCS and 37.6 frames·s^-1on embedded devices, proving that the proposed method can be deployed on embedded devices for real-time detection of UAVs.

Key words: lightweight network, YOLOv4, inspection drone, embedded platform, deep learning, Ghostnet, ELU, real-time detection

中图分类号:

TP391

郑玉恒,付东翔. 基于Slim-YOLOv4与嵌入式设备的无人机检测[J]. 电子科技, 2023, 36(5): 55-61.

ZHENG Yuheng,FU Dongxiang. UAV Detection Based on Slim-YOLOv4 with Embedded Device[J]. Electronic Science and Technology, 2023, 36(5): 55-61.

图/表 14

图1

图2

图3

表1

图4

图5

图6

图7

图8

表2

图9

表3

表4

表5

参考文献 20

[1]	Liu W, Anguelov D, Erhan D, et al. SSD: Single shot multibox detector[C]. Amsterdam: Proceedings of the European Conference on Computer Vision, 2016:21-37.
[2]	Molchanov P, Mallya A, Tyree S, et al. Importance estimation for neural network pruning[C]. Long Beach: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019: 11264-11272.
[3]	Chen Z, Zhang L, Cao Z, et al. Distilling the knowledge from handcrafted features for human activity recognition[J]. IEEE Transactions on Industrial Informatics, 2018, 14(10):4334-4342. doi: 10.1109/TII.2018.2789925
[4]	Yin P H, Zhang S, Qi Y, et al. Quantization and training of low bit-width convolutional neural networks for object detection[J]. Mathematica Numerica Sinica, 2019, 37(3):349-359.
[5]	Lee H J, Ullah I, Wan W, et al. Real-time vehicle make and model recognition with the residual SqueezeNet architecture[J]. Sensors, 2019, 19(5):982-997. doi: 10.3390/s19050982
[6]	Howard A, Sandler M, Chu G, et al. Searching for mobilenetv3[C]. Long Beach: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019:1314-1324.
[7]	Zhang X, Zhou X, Lin M, et al. Shufflenet: An extremely efficient convolutional neural network for mobile devices[C]. Salt Lake City: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018:6848-6856.
[8]	Han K, Wang Y, Tian Q, et al. GhostNet: More features from cheap operations[C]. Seattle: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020:1580-1589.
[9]	Li Y, Lü C. SS-YOLO: An object detection algorithm based on YOLOv3 and ShuffleNet[C]. Chongqing: Proceedings of the IEEE Fourth Information Technology, Networking, Electronic and Automation Control Conference, 2020:769-772.
[10]	Farhadi A, Redmon J. YOLOv3: An incremental improvement[C]. Berlin: Proceedings of the Computer Vision and Pattern Recognition, 2018:1-12
[11]	冯媛, 李敬兆. 改进的卷积神经网络行人检测方法[J]. 计算机工程与设计, 2020, 41(5):1452-1457.
	Feng Yuan, Li Jingzhao. Improved convolutional neural network pedestrian detection method[J]. Computer Engineering and Design, 2020, 41(5):1452-1457.
[12]	靳丽蕾, 杨文柱, 王思乐, 等. 一种用于卷积神经网络压缩的混合剪枝方法[J]. 小型微型计算机系统, 2018, 39(12):2596-2601.
	Jin Lilei, Yang Wenzhu, Wang Sile, et al. A hybrid pruning method for convolutional neural network compression[J]. Small Microcomputer Systems, 2018, 39(12):2596-2601.
[13]	成凌飞, 贺扬, 张培玲, 等. 压缩深层神经网络隐藏层维度对其分类性能的影响[J]. 电子科技, 2019, 32(1):72-75.
	Cheng Lingfei, He Yang, Zhang Peiling, et al. Effect of hidden layer dimension on classification performance of compressed deep neural network[J]. Electronic Science and Technology, 2019, 32(1):72-75.
[14]	邵伟平, 王兴, 曹昭睿, 等. 基于MobileNet与YOLOv3的轻量化卷积神经网络设计[J]. 计算机应用, 2020, 40(S1):8-13.
	Shao Weiping, Wang Xing, Cao Zhaorui, et al. Design of lightweight convolutional neural network based on MobileNet and YOLOv3[J]. Computer Applications, 2020, 40(S1):8-13.
[15]	Bochkovskiy A, Wang C Y, Liao H Y M. YOLOv4: Optimal speed and accuracy of object detection[C]. Seattle: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020:1-10.
[16]	Chollet F. Xception: Deep learning with depthwise separable convolutions[C]. Holunono: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017:1251-1258.
[17]	Devi T, Deepa N. A novel intervention method for aspect-based emotion using exponential linear unit activation function in a deep neural network[C]. Madurai: Proceedings of the Fifth International Conference on Intelligent Computing and Control Systems, 2021: 1671-1675.
[18]	Wang C Y, Liao H Y M, Wu Y H, et al. CSPNet: A new backbone that can enhance learning capability of CNN[C]. Seattle: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020:390-391.
[19]	Wang K, Liew J H, Zou Y, et al. Panet: Few-shot image semantic segmentation with prototype alignment[C]. Long Beach: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019:9197-9206.
[20]	He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]. Las Vegas: Proceeding of the IEEE Conference on Computer Vision and Pattern Recognition, 2016:770-778.

输入	操作	扩张通道	输出通道	SE	步长
224×224×3	Conv2d 3×3	-	16	-	2
112×112×16	G-bneck	16	16	-	1
112×112×16	G-bneck	48	24	-	2
56×56×24	G-bneck	72	24	-	1
56×56×24	G-bneck	72	40	1	2
28×28×40	G-bneck	120	40	1	1
28×28×40	G-bneck	240	80	-	2
14×14×80	G-bneck	200	80	-	1
14×14×80	G-bneck	184	80	-	1
14×14×80	G-bneck	184	80	-	1
14×14×80	G-bneck	480	112	1	1
14×14×112	G-bneck	672	112	1	1
14×14×112	G-bneck	672	160	1	2
7×7×160	G-bneck	960	160	-	1
7×7×160	G-bneck	960	160	1	1
7×7×160	G-bneck	960	160	-	1
7×7×160	G-bneck	960	160	1	1
7×7×160	Conv2d 1×1	-	960	-	1
7×7×960	AvgPool 7×7	-	-	-	-
1×1×960	Conv2d 1×1	-	1 280	-	1
1×1×1 280	FC	-	1 000	-	-

模型	准确率 /%	召回率 /%	Weights /MB	FPS /frames·s^-1
Faster-RCNN	90.7	88.3	108	3.1
YOLOv4	93.2	90.5	250	5.8
Gh-YOLOv4	91.5	90.1	150	15.4
Slim-YOLOv4	91.6	86.9	42	31.2

主干网络	准确率 /%	召回率 /%	F1 /%	Weights /MB	FPS /frames·s^-1
Mobilenetv3	89.8	90.6	0.90	53	25.54
Ghostnet	91.6	86.9	0.89	42	31.20

类别	Faster R-CNN	YOLOV4	Slim-YOLOv4
Cat	98.2	97.6	96.3
Person	90.3	93.2	90.8
Aero	95.3	98.5	97.8
Car	88.4	92.6	90.0
Bus	97.2	97.9	94.4
Dog	95.6	96.2	95.5
Motorbike	92.1	94.5	90.6
Train	98.5	98.7	96.0
Horse	92.8	96.5	93.6
Table	70.8	86.3	83.0
Sheep	83.9	88.4	86.4
Tv	83.3	94.6	90.9
Boat	86.9	91.2	86.4
Chair	79.6	90.6	86.9
Bird	93.2	97.4	92.3
Bicycle	90.8	93.6	87.8
Bottle	72.4	80.3	78.8
Sofa	88.8	94.8	88.6
Plant	74.7	79.8	76.8
Cow	98.6	98.1	95.3
mAP/%	88.6	93.0	89.9