一种改进CenterNet的轻量化目标检测算法

doi:10.19665/j.issn1001-2400.2022.05.016

摘要/Abstract

摘要：

CenterNet算法结构复杂,导致其参数量大、计算复杂度高和检测速度较慢。针对这一问题,提出了一种CenterNet-encoder算法。该算法使用深度为104层的沙漏网络作为backbone,并将其中的残差模块替换为fire模块来减少算法的参数量,提高算法的计算速度;另外,在backbone和head之间加入了编码层,在不损失分辨率的同时增大了感受野,减少了内存的占用,让输出囊括更多尺度的信息;最后,使用均方误差损失进行边界框的回归,加快算法的收敛,进一步提升了算法的检测精度。CenterNet-encoder算法最终在MS-COCO test-dev数据集上的平均检测精度为40.5%,参数量为47×10⁶。在AMD5900X/32GB/RTX3090环境配置下,检测速度达到了18帧/s。实验结果表明,CenterNet-encoder算法虽然牺牲了一定的精度,但参数量比原算法下降了约77.6%,同时检测速度提升了约69.3%。与其他轻量化目标检测算法相比,在参数量、推理时间和检测精度上也有一定的优势。

关键词: fire模块, 轻量化, 编码层, 目标检测

Abstract:

Due to the complex structure,there are a large number of parameters in the CenterNet,which leads to a high computational complexity and long inference time.To solve this problem,a CenterNet-encoder algorithm is proposed for lightweight object detection.First,the fire module is used in the backbone to reduce the number of parameters and increase the calculation speed.Then,an encoding layer is utilized between the backbone and the head,which can increase the receptive field and obtain more accurate corners and center points for heatmaps.Finally,MSE loss is employed in bounding box regression,which accelerates the convergence and further improves the performance.The proposed algorithm achieves 40.5 AP on the MS-COCO test-dev benchmark with 47M parameters.Under the AMD5900X\32GB\RTX3090 environment configuration,the detection speed reaches 18FPS.Experimental results show that the performance of the proposed method is better than other lightweight methods in the number of parameters,inference time and detection accuracy.Although the precision of the proposed method is slightly lower than that of the CenterNet,the number of parameters is reduced by 77.6%,and the inference speed is increased by 69.3%.

Key words: fire module, lightweight, encoder, object detection

中图分类号:

TP391

李悦言, 程培涛, 杜淑幸. 一种改进CenterNet的轻量化目标检测算法[J]. 西安电子科技大学学报, 2022, 49(5): 137-144.

LI Yueyan, CHENG Peitao, DU Shuxing. Lightweight object detection algorithm based on the improved CenterNet[J]. Journal of Xidian University, 2022, 49(5): 137-144.

图/表 10

图1

图2

图3

图4

图5

表1

表2

表3

图6

图7

参考文献 23

[1]	黄健, 张钢. 深度卷积神经网络的目标检测算法综述[J]. 计算机工程与应用, 2020, 56(17):12-23. doi: 10.3778/j.issn.1002-8331.2005-0021
	HUANG Jian, ZHANG Gang. Survey of Object Detection Algorithms for Deep Convolutional Neural Networks[J]. Computer Engineering and Applications, 2020, 56(17):12-23. doi: 10.3778/j.issn.1002-8331.2005-0021
[2]	HOSANG J, BENENSON R, DOLLÁRP, et al. What Makes for Effective Detection Proposals?[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 38(4):814-830. doi: 10.1109/TPAMI.2015.2465908
[3]	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2014:580-587.
[4]	GIRSHICK R. Fast R-CNN[C]// Proceedings of the IEEE International Conference on Computer Vision. Piscataway: IEEE, 2015:1440-1448.
[5]	DAI J, LI Y, HE K, et al. R-FCN:Object Detection via Region-Based Fully Convolutional Networks[C]// Advances in Neural Information Processing Systems. New York: ACM, 2016:379-387.
[6]	REN S, HE K, GIRSHICKR, et al. Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 39(6):1137-1149. doi: 10.1109/TPAMI.2016.2577031
[7]	REDMON J, DIVVALA S, GIRSHICKR, et al. You Only Look Once:Unified,Real-Time Object Detection[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2016:779-788.
[8]	LIU W, ANGUELOV D, ERHAN D, et al. SSD:Single Shot MultiBox Detector[C]// Proceeding of the European Conference on Computer Vision. Heidelberg: Springer, 2016:21-37.
[9]	TIAN Z, SHEN C, CHEN H, et al. FCOS:Fully Convolutional One-Stage Object Detection[C]// Proceedings of the IEEE International Conference on Computer Vision. Piscataway: IEEE, 2019:9627-9636.
[10]	LAW H, DENG J. CornerNet:Detecting Objects as Paired Keypoints[C]// Proceedings of the European Conference on Computer Vision. Heidelberg: Springer, 2018:734-750.
[11]	DUAN K, BAI S, XIE L, et al. CenterNet:Keypoint Triplets for Object Detection[C]// Proceedings of the IEEE International Conference on Computer Vision. Piscataway: IEEE, 2019:6569-6578.
[12]	IANDOLA F N, HAN S, MOSKEWICZ M W, et al. SqueezeNet:AlexNet-Level Accuracy with 50x Fewer Parameters and <0.5 MB Model Size (2016)[J/OL].[2016-02-24]. .
[13]	HOWARD A G, ZHU M, CHEN B, et al. MobileNets:Efficient Convolutional Neural Networks for Mobile Vision Applications (2017)[J/OL].[2017-04-17]. https://arxiv.org/pdf/1704.04861.pdf .
[14]	LAW H, TENG Y, RUSSAKOVSKYO, et al. CornerNet-Lite:Efficient Keypoint Based Object Detection (2019) [EB/OL].[2019-04-18]. https://arxiv.org/pdf/1904.08900.pdf .
[15]	CHEN Q, WANG Y, YANG T, et al. You Only Look One-Level Feature[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2021:13039-13048.
[16]	REDMON J, FARHADI A. Yolov3:An Incremental Improvement (2018)[J/OL].[2018-04-08]. https://arxiv.org/pdf/1804.02767.pdf .
[17]	LIN T Y, MAIRE M, BELONGIE S, et al. Microsoft COCO:Common Objects in Context[C]// Proceedings of the European Conference on Computer Vision. Heidelberg: Springer, 2014:740-755.
[18]	HOIEM D, CHODPATHUMWAN Y, DAI Q. Diagnosing Error in Object Detectors[C]// European Conference on Computer Vision. Heidelberg: Springer, 2012:340-353.
[19]	NEWELL A, YANG K, DENG J. Stacked Hourglass Networks for Human Pose Estimation[C]// Proceedings of the European Conference on Computer Vision. Heidelberg: Springer, 2016:483-499.
[20]	BODLA N, SINGH B, CHELLAPPA R, et al. Soft-NMS-Improving Object Detection with One Line of Code[C]// Proceedings of the IEEE International Conference on Computer Vision. Piscataway: IEEE, 2017:5561-5569.
[21]	ZOU Z, SHI Z, GUO Y, et al. Object Detection in 20 years:A Survey (2019)[J/OL].[2019-05-13]. https://arxiv.org/pdf/1905.05055.pdf .
[22]	OKSUZ K, CAM B C, KALKAN S, et al. Imbalance Problems in Object Detection:A Review[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(10):3388-3415. doi: 10.1109/TPAMI.2020.2981890
[23]	陆峰, 刘华海, 黄长缨, 刘财喜, 等. 基于深度学习的目标检测技术综述[J]. 计算机系统应用, 2021, 30(3):1-13.
	LU Feng, LIU Huahai, HUANG Changying, et al. Overview on Deep Learning-Based Object Detection[J]. Computer Systems Applications, 2021, 30(3):1-13.

算法	Backbone	参数量	速度/(ms·image^-1)	AP	AP₅₀	AP₇₅	AP_S	AP_M	AP_L
YOLOv3	DarkNet-53	61.9×10⁶	51	33.0	57.9	34.4	18.3	35.4	41.9
CornerNet_Squeeze	Hourglass-104	31.7×10⁶	33	34.4	49.2	36.2	14.8	36.9	49.5
CenterNet-encoder	Hourglass-104	47.0×10⁶	57	40.5	58.1	42.8	20.5	44.1	56.0

算法	Backbone	参数量	速度/(ms·image^-1)	AP	AP₅₀	AP₇₅	AP_S	AP_M	AP_L
CornerNet	Hourglass-52	99.0×10⁶	125	37.8	53.7	40.1	17.0	39.0	50.5
CornerNet	Hourglass-104	201.0×10⁶	153	40.5	56.4	43.1	19.4	42.7	53.9
CenterNet	Hourglass-52	104.8×10⁶	166	41.6	59.4	44.2	22.5	43.1	54.1
CenterNet	Hourglass-104	210.0×10³	186	44.9	62.4	48.1	25.6	47.4	57.4
CenterNet-encoder	Hourglass-104	47.0×10³	57	40.5	58.4	43.0	20.7	44.5	56.3

偏移损失函数	AP	AP₅₀	AP₇₅	AP_S	AP_M	AP_L
SmoothL₁loss	39.4	57.3	41.7	19.5	42.6	55.4
MSEloss	40.5	58.4	43.0	20.7	44.5	56.3