基于模型压缩与感受野增强的下茧实时检测

doi:10.13475/j.fzxb.20210103610

摘要/Abstract

摘要：

针对目前选茧时下茧检测主要依赖人工目测,工作效率低的问题,提出一种基于锚点框参数预置、通道剪枝和嵌入感受野模块改进的轻量化下茧实时检测模型。首先,通过K-means聚类分析出适用于下茧检测的锚点框以预置YOLOv3模型参数;然后,根据预设的剪枝率对稀疏化训练后的模型进行基于批量正则化层缩放因子的模型剪枝,以此压缩模型的大小;最后,在剪枝后的模型中嵌入感受野模块,使模型的感受野变大,增强模型的辨别能力和鲁棒性。实验结果表明:提出的下茧实时检测模型大小为46.90 M,平均检测速度达到50.18 帧/s,平均检测精度为96.80%,较原YOLOv3模型参数压缩了79.96%,平均检测速度提高了60.63%,平均检测精度提高了3.20%。

关键词: 蚕茧, 下茧检测, YOLOv3模型, 聚类分析, 模型压缩, 感受野模块

Abstract:

The current inferior cocoons detection mainly depends on manual visual inspection, which leads to low work efficiency. Based on anchor parameter presetting, channel pruning and embedding receptive field block, an improved lightweight real time inferior cocoons detection model was proposed.Firstly, the model parameters of YOLOv3 were preset through K-means clustering analysis of the anchor suitable for inferior cocoons detection. Then, according to a preset pruning rate, the sparsely trained model was pruned based on batch normalization layer scaling factor to compress the size of the model. Finally, the receptive field block was embedded the pruned model to enlarge the receptive field of the model and enhance the discriminability and robustness of the model. The experimental results show that the proposed inferior cocoons real-time detection model is 46.90 M in size, and the mean average detection speed and precision reach 50.18 frames/s and 96.80% respectively. Compared with the original YOLOv3 model, the parameters are compressed by 79.96%, the mean average detection speed is increased by 60.63%, and the mean average detection precision is increased by 3.20%.

Key words: cocoons, inferior cocoons detection, YOLOv3 model, clustering analysis, model compression, receptive field block

中图分类号:

TP391

张印辉, 杨宏宽, 刘强, 何自芬. 基于模型压缩与感受野增强的下茧实时检测[J]. 纺织学报, 2021, 42(11): 29-38.

ZHANG Yinhui, YANG Hongkuan, LIU Qiang, HE Zifen. Real-time detection of inferior cocoons through model compression and receptive field enhancement[J]. Journal of Textile Research, 2021, 42(11): 29-38.

图/表 15

图1

图2

图3

图4

图5

图6

表1

表2

图7

表3

表4

图8

图9

表5

图10

参考文献 21

[1]	封槐松, 李建琴. 新中国60年蚕桑业发展历程与特点[J]. 中国蚕业, 2014, 35(3):1-10.
	FENG Huaisong, LI Jianqin. New China 60 years sericulture development course and characteristic[J]. Chinese Sericulture, 2014, 35(3):1-10.
[2]	封槐松, 李建琴. 我国蚕桑产业发展“十二五”回顾与“十三五”展望[J]. 中国蚕业, 2016, 37(1):4-10.
	FENG Huaisong, LI Jianqin. China sericulture industry development "twelfth five-year" review and "thirteenth five-year" outlook[J]. Chinese Sericulture, 2016, 37(1):4-10.
[3]	MUSAYEV E. Optoelectronic nondestructive testing techniquesof cocoons properties and applications[J]. NDT & International, 2005, 38(1):59-68.
[4]	PRASOBHKUMAR P P, FRANCIS C R, GORTHI S S. Automated quality assessment of cocoons using a smart camera based system[J]. Engineering in Agriculture, Environment and Food, 2018, 11(4):202-210. doi: 10.1016/j.eaef.2018.05.002
[5]	周志宇, 刘喜昂, 杨东鹤. 机器视觉在蚕茧表面积测量中的应用[J]. 纺织学报, 2006, 27(12):29-31.
	ZHOU Zhiyu, LIU Xi'ang, YANG Donghe. Application of machine vision in measuring cocoons surface area[J]. Journals of Textile Research, 2006, 27(12):29-31.
[6]	甘勇, 孔庆华, 韦荔甫. 基于数字图像处理的蚕茧表面积计算方法[J]. 光学技术, 2007, 33(1):27-28.
	GAN Yong, KONG Qinghua, WEI Lifu. A method for calculating cocoons surface area based on digital image processing[J]. Optical Technology, 2007, 33(1):27-28.
[7]	黄凌霞, 吴迪, 金航峰, 等. 基于变量选择的蚕茧茧层量可见-近红外光谱无损检测[J]. 农业工程学报, 2010, 26(2):231-236.
	HUANG Lingxia, WU Di, JIN Hangfeng, et al. The cocoons layer quantity based on variable selection can be seen-NIR spectrum nondestructive testing[J]. Journal of Agricultural Engineering, 2010, 26(2):231-236.
[8]	代芬, 吴玲, 叶观燕, 等. 基于近红外漫透射光谱信息的蚕茧雌雄检测[J]. 农业机械学报, 2015, 46(12):280-284.
	DAI Fen, WU Ling, YE Guanyan, et al. Sex detection of silkworm cocoons based on near-infrared diffuse transmission spectral information[J]. Journal of Agricultural Machinery, 2015, 46(12):280-284.
[9]	王超. 基于机器视觉的蚕茧图像识别研究[D]. 柳州:广西科技大学, 2019:24-29.
	WANG Chao. Research on cocoon image recognition based on machine vision[D]. Liuzhou:Guangxi University of Science and Technology, 2019:24-29.
[10]	刘莫尘, 许荣浩, 闫筱, 等. 基于FCM及HSV模型的方格蔟黄斑茧检测与剔除技术[J]. 农业机械学报, 2018, 49(7):31-38.
	LIU Mocheng, XU Ronghao, YAN Xiao, et al. Based on FCM and HSV model, detection and screening technology of grid cluster macular cocoons[J]. Journal of Agricultural Machinery, 2018, 49(7):31-38.
[11]	刘莫尘, 许荣浩, 李法德, 等. 基于颜色与面积特征的方格簇蚕茧分割定位算法与试验[J]. 农业机械学报, 2018, 49(3):43-50.
	LIU Mocheng, XU Ronghao, LI Fade, et al. Segmentation and location algorithm of square silkworm cocoons based on color and area characteristics and experiment[J]. Journal of Agricultural Machinery, 2018, 49(3):43-50.
[12]	REDMON J FARHADI A, YOLOv3: an incremental improvement[EB/OL]. (2018-04-08)[2021-01-10]https://arXiv:1804.02767 .
[13]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]// Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition(CVPR). Las Vegas: IEEE, 2016: 779-788.
[14]	REDMON J, FARHADI A. YOLO9000: better, faster, stronger[C]// Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition(CVPR). Hawaii: IEEE, 2017: 6517-6525.
[15]	HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]// Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition(CVPR). Las Vegas: IEEE, 2016: 770-778.
[16]	IOFFE S, SZEGEDY C. Batch normalization: accelerating deep network training by reducing internal covariate shift[C]// Proceedings of International Conference on Machine Learning(ICML). Lille: IMIS, 2015: 448-456.
[17]	LIN T Y, DOLLAR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]// Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition(CVPR). Hawaii: IEEE, 2017: 936-944.
[18]	LIU Zhuang, LI Jianguo, SHEN Zhiqiang, et al. Learning efficient convolutional networks through network slimming[C]// Proceedings of 2017 IEEE International Conference on Computer Vision(ICCV). Venice: IEEE, 2017: 2755-2736.
[19]	LIU Songtao, HUANG Di, WANG Yunhong. Receptive field block net for accurate and fast object detec-tion [C]//Proceedings of European Conference on Computer Vision(ECCV). Munich: Springer, 2018: 404-419.
[20]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector [C]//Proceedings of European Conference on Computer Vision(ECCV). Amsterdam: Springer, 2016: 21-37.
[21]	BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv4: optimal speed and accuracy of object detec-tion[EB/OL]. (2020-04-23)[2021-01-10]https://arxiv.org/abs/2004.10934 .

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

参数名称	数值
giou损失的系数(giou)	1.582
分类损失的系数(cls)	27.76
分类损失函数中正样本的权重(cls_pw)	1.446
有无物体损失的系数(obj)	21.35
有无物体损失函数中正样本的权重(obj_pw)	3.941
标签与锚点框的iou阈值(iou_t)	0.263 5
学习率(lr0)	0.002 324
余弦退火超参数(lrf)	0.000 1
学习率动量(momentum)	0.97
权重衰减系数(weight_decay)	0.000 456 9

锚点框数量	锚点框尺寸	模型大小/M	平均检测速度/(帧·s^-1)	mAP/%
3	(38,28),(31,35),(31,41)	234.00	31.14	96.10
6	(26,32),(33,26),(28,42), (40,30),(33,38),(40,34)	234.00	30.68	96.00
9	(27,29),(24,33),(33,26), (26,40),(30,35),(42,28, (38,33),(30,42),(35,39)	234.00	30.46	95.90
12	(28,27),(34,24),(26,33),(27,39), (35,31),(41,27),(31,37),(26,38), (44,30),(33,41),(41,34),(37,38)	235.00	30.38	95.20

剪枝率/%	模型大小/M	平均检测速度/ (帧·s^-1)	mAP/%
10.00	205.00	30.96	95.99
20.00	177.00	33.82	95.99
30.00	151.00	37.36	95.99
40.00	127.00	39.93	95.99
50.00	104.00	42.23	95.99
60.00	83.60	44.53	95.98
70.00	65.00	47.39	95.97
80.00	46.90	51.05	95.95

RFB添加的位置	模型大小/M	平均检测速度/ (帧·s^-1)	mAP/%
4层后	46.90	47.52	96.70
11层后	46.90	49.87	96.60
36层后	46.90	50.18	96.80
61层后	46.90	50.04	96.20
74层后	46.90	49.57	96.10
86层后	46.90	48.47	96.70
98层后	46.90	48.77	96.10

检测模型	模型大小/M	平均检测速度/ (帧·s^-1)	mAP/%
YOLOv3	234.00	31.24	93.80
YOLOv4	245.00	26.38	96.78
SSD	92.10	41.79	96.64
YOLOv3-MC-RFB	46.90	50.18	96.80