一种融合MeanShift聚类分析和卷积神经网络的Vibe++背景分割方法

doi:10.11959/j.issn.1000-0801.2021046

电信科学 ›› 2021, Vol. 37 ›› Issue (3): 133-145.doi: 10.11959/j.issn.1000-0801.2021046

一种融合MeanShift聚类分析和卷积神经网络的Vibe++背景分割方法

刘子豪¹, 贾小军¹, 张素兰¹, 徐志玲², 张俊³

¹ 嘉兴学院数理与信息工程学院，浙江嘉兴 314033
² 中国计量大学质量与安全工程学院，浙江杭州 310018
³ 浙江大学生物系统工程与食品科学学院，浙江杭州 310058

修回日期:2021-02-17 出版日期:2021-03-20 发布日期:2021-03-01
作者简介:刘子豪（1988- ），男，博士，嘉兴学院讲师、硕士生导师，主要研究方向为人工智能与图像处理、基于机器视觉的农产品无损检测。
贾小军（1974- ），男，博士，嘉兴学院副教授、硕士生导师，主要研究方向为人工智能与图像处理。
张素兰（1980- ），女，博士，嘉兴学院讲师，主要研究方向为三值光学计算机、系统结构、嵌入式系统。
徐志玲（1966- ），女，中国计量大学教授，主要研究方向为计量检测成像技术研究与仪器开发、工业机器人检测系统研究与开发。
张俊（1994- ），男，浙江大学生物系统工程与食品科学学院博士生，主要研究方向为基于谱图成像技术的农产品无损检测。
基金资助:
浙江省基础公益研究计划项目(LGG21F030013);浙江省基础公益研究计划项目(LGG20F010010);浙江省基础公益研究计划项目(LGG20F030006);嘉兴市公益计划项目(2020AY10009);嘉兴市公益计划项目(2018AY11008);嘉兴学院科研启动基金(CD70519085)

Vibe++ background segmentation method combining MeanShift clustering analysis and convolutional neural network

Zihao LIU¹, Xiaojun JIA¹, Sulan ZHANG¹, Zhiling XU², Jun ZHANG³

¹ College of Mathematics Physics and Information Engineering, Jiaxing University, Jiaxing 314033, China
² College of Quality ＆Safty Engineering, China Jiliang University, Hangzhou 310018, China
³ College of Biosystems Engineering and Food Science, Zhejiang University, Hangzhou 310058, China

Revised:2021-02-17 Online:2021-03-20 Published:2021-03-01
Supported by:
Zhejiang Public Welfare Technology Research Project(LGG21F030013);Zhejiang Public Welfare Technology Research Project(LGG20F010010);Zhejiang Public Welfare Technology Research Project(LGG20F030006);Jiaxing City Public Welfare Technology Application Research Project of Jiaxing Science and Technology Bureau(2020AY10009);Jiaxing City Public Welfare Technology Application Research Project of Jiaxing Science and Technology Bureau(2018AY11008);Scientific Research Foundation of Jiaxing University(CD70519085)

摘要/Abstract

摘要：

针对传统 Vibe+算法存在噪点和拖影分割错误率较高的问题，提出了一种改进的 Vibe+运动目标分割算法（Vibe++）。首先，通过对视频帧采用传统Vibe+算法处理获取二值图像，基于区域生长算法对结果图中各连通域标记，依据边界面积块差异获取面积筛选阈值，将低于阈值的连通区域视为噪点并删除；然后，引入 5 种不同核函数优化传统 MeanShift 聚类算法，并与卷积神经网络（CNN）进行顺序组合；最后，采用组合模型对已消除噪点图像中的拖影区、非拖影区和拖影边缘区分类，计算拖影区中每个像素点的坐标，定位拖影区并快速删除，获取分割结果。所提算法用于公开数据集的实验结果表明，其可取得 98%以上的分割准确率，具有较好的应用效果和较高的实用价值。

关键词: 背景分割, 聚类分析, 分割准确率, 卷积神经网络

Abstract:

To solve problems of noise points and high segmentation error for image shadow brought by traditional Vibe+ algorithm, a novel background segmentation method (Vibe++) based on the improved Vibe+ was proposed.Firstly, binarization image was acquired by using traditional Vibe+ algorithm from surveillance video.The connected regions were marked based on the region-growing domain marker method.The area threshold was obtained with difference characteristics of boundary area, the connected regions below threshold were treated as disturbing points.Secondly, five different kernel functions were introduced to improve the traditional MeanShift clustering algorithm.After improving, this algorithm was fused effectively with partitioned convolutional neural network.Finally, program of classification of trailing area, non-trailing area and trailing edge area in the resulting image was performed.Position coordinates of the trailing area were calculated and confirmed, and the trailing area was quickly deleted to obtain the final segmentation result.This segmentation accuracy was greatly improved by using the proposed method.The experimental results show that the proposed algorithm can achieve segmentation accuracy of more than 98% and has good application effect and high practical value.

Key words: background segmentation, clustering analysis, segmentation accuracy, convolutional neural network

中图分类号:

TN391

刘子豪, 贾小军, 张素兰, 徐志玲, 张俊. 一种融合MeanShift聚类分析和卷积神经网络的Vibe++背景分割方法[J]. 电信科学, 2021, 37(3): 133-145.

Zihao LIU, Xiaojun JIA, Sulan ZHANG, Zhiling XU, Jun ZHANG. Vibe++ background segmentation method combining MeanShift clustering analysis and convolutional neural network[J]. Telecommunications Science, 2021, 37(3): 133-145.

图/表 11

图1

图2

图3

图4

图5

图6

表1

所提Vibe++算法性能评价"

分割算法	算法性能评价指标
分割算法	视频帧分割时长/ms	分割准确率	算法时间复杂度
背景差分法^[1]	23.5	83.31%	O(n)
光流法^[2]	86.9	93.12%	O(nlgn)
分水岭算法^[3]	56.9	89.78%	O(lgn)
帧差法^[4]	$12 . 4$	85.72%	O(n)
水平集法^[5]	42.1	87.93%	O(n²)
Vibe^[6]	89.4	92.13%	O(m×n)
Vibe+^[7]	107.6	95.45%	O(nlg(mn))
背景更新恢复法^[8]	34.5	86.05%	O(n³)
Vibe++	132.7	$98 . 58 %$	O(n²lg(mn²))

表1

表2

本文所提方法与先前提出的方法对比"

对比的数据集类型	所提方法相比于之前方法的分割准确率
对比的数据集类型	Previous best	Vibe+	Vibe++
基准类	$95 . 32 %$ ^[22]	92.62%	94.89%
动态背景类	92.12%^[6]	72.91%	$94 . 32 %$
抖动类	95.16%^[6]	80.64%	$98 . 05 %$
间断运动目标类	$96 . 62 %$ ^[23]	75.12%	95.98%
拖影类	94.72%^[6]	83.02%	$97 . 68 %$
热图类	$98 . 29 %$ ^[24]	93.63%	98.16%
所有类（Overall）	97.36%^[6]	87.45%	$98 . 58 %$

表2

表3

图7

图8

参考文献 27

[1]	DIM M , BASHIR A , BASIT A ,et al. Abandoned object detection using frame differencing and background subtraction[J]. International Journal of Applied Mathematics and Computer Science, 2020,11(7): 676-681.
[2]	ZHOU S T , ZHANG Q , GAO R ,et al. High-accuracy atmospheric turbulence compensation based on a Wirtinger flow algorithm in an orbital angular momentum-free space optical communication system[J]. Optical Communication, 2020,477(2): 345-353.
[3]	DERIVAUX S , FORESTIER G , WEMMERT C ,et al. Supervised image segmentation using watershed transform,fuzzy classification and evolutionary computation[J]. Pattern Recogntion Letters, 2010,31(15): 2364-2374.
[4]	ALGETHAMI N , REDFERN S . A robust tracking-by-detection algorithm using adaptive accumulated frame differencing and corner Features[J]. Journal of Imaging, 2020,6(4): 1-10.
[5]	LYRAS K G , HANSON B , FAIRWEATHER M ,et al. A coupled level set and volume of fluid method with a re-initialisation step suitable for unstructured meshes[J]. Journal of Computation Physics, 2020,407(12): 124-139.
[6]	ALVAREZ-MEZA A M , MOLINA-GIRALDO S , CASTELLANOS-DOMINGUEZ G . Background modeling using object-based selective updating and correntropy adaptation[J]. Image and Vision Computing, 2016,45(13): 22-36.
[7]	BARNICH O , VAN D M . ViBe:a universal background subtraction algorithm for video sequences[J]. IEEE Transaction of Image Process, 2011,20(6): 1709-1724.
[8]	GULESIR G , BERGMANS L , AKSIT M ,et al. Vibes:a visual language for specifying behavioral requirements of algorithms[J]. Journal of Visual Languages ＆ Computing, 2013,24(5): 350-364.
[9]	ZHOU X Q , LIU X F , JIANG A M ,et al. Improving video segmentation by fusing depth cues and the visual background extractor (ViBe) algorithm[J]. Sensors, 2017,17(5): 56-67.
[10]	彭丰平, 鲍苏苏, 曾碧卿 . 基于自适应区域生长算法的肝脏分割[J]. 计算机工程与应用, 2010,46(33): 198-200.
	PENG F P , BAO S S , ZENG B Q . Segmentation of liver based on adaptive region growing[J]. Computer Engineering and Applications, 2010,46(33): 198-200.
[11]	钮圣唬, 王盛, 杨晶晶 ,等. 完全基于边缘信息的快速图像分割算法[J]. 计算机辅助设计与图形学学报, 2012,24(11): 1410-1419.
	NIU S X , WANG S , YANG J J ,et al. A fast segmentation algorithm fully based on edge information[J]. Journal of Computer-Aided Design ＆ Computer Graphics, 2012,24(11): 1410-1419.
[12]	CAO X F , QIU B Z , XU G D . BorderShift:toward optimal MeanShift vector for cluster boundary detection in high-dimensional data[J]. Pattern Analysis and Applications, 2019,22(3): 1015-1027.
[13]	MITTAL G , KORUS P , MEMON N . FiFTy:large-scale file fragment type identification using convolutional nural networks[J]. IEEE Transactions on Information Forensics and Security, 2021,16(3): 28-41.
[14]	KEINOSUKE F , LARRY D H . The estimation of the gradient of a density function with application in pattern recognition[J]. IEEE Transations on Information Theory, 1975,21(1): 32-40.
[15]	CHENG Y Z . Mean shift,mode seeking,and clustering[J]. IEEE Transations on Pattern and Intelligence, 1995,17(8): 790-799.
[16]	姜枫, 顾庆, 郝慧珍 ,等. 基于内容的图像分割方法综述[J]. 软件学报, 2017,28(1): 160-183.
	JIANG F , GU Q , HAO H Z ,et al. Survey on content-based image segmentation methods[J]. Journal of Software, 2017,28(1): 160-183.
[17]	LIU Z , CHENG F , HONG HM . Identification of impurities in fresh shrimp using improved majority scheme-based classifier[J]. Food Analytical Methods, 2016,16(3): 10-20.
[18]	任洪蛾, 白杰云 . Lab颜色空间中基于动态聚类的颜色分级[J]. 计算机工程, 2013,39(6): 231-235.
	REN H E , BAI J Y . Color grading based on dynamic clustering in lab color space[J]. Computer Engineering, 2013,39(6): 231-235.
[19]	彭宁嵩, 杨杰, 刘志 ,等. MeanShift 跟踪算法中核函数窗宽的自动选取[J]. 软件学报, 2005,16(9): 1542-1550.
	PENG N S , YANG J , LIU Z ,et al. Automatic selection of kernel-bandwidth for Mean-Shift object tracking[J]. Journal of Software, 2005,16(9): 1542-1550.
[20]	邓柳, 汪子杰 . 基于深度卷积神经网络的车型识别研究[J]. 计算机应用研究, 2016,33(3): 930-932.
	DENG L , WANG Z J . Deep convolution neural networks for vehicle classification[J]. Application Research of Computers, 2016,33(3): 930-932.
[21]	DROOGENBROECK M V , PAQUOT O . Background subtraction:experiments and improvements for ViBe[C]// Proceedings:In Change Detection Workshop(CDW).[S.l.:s.n.], 2012.
[22]	HOFMANN M , TIEFENBACHER P,RIGOLL,G . Background segmentation with feedback:The pixel-based adaptive segmenter[C]// Proceedings of Computer Vision and Pattern Recognition Workshops(CVPR). Piscataway:IEEE Press, 2012: 38-43.
[23]	GOYETTE N , JODOIN P , PORIKLI F ,et al. Changedetection net:A new change detection benchmark dataset[C]// Proceedings of Computer Vision and Pattern Recognition Workshops(CVPR). Piscataway:IEEE Press, 2012.
[24]	CHEN Z , ELLIS T . A self-adaptive gaussian mixture model[J]. Computer Vision and Image Understanding, 2014(122): 35-46.
[25]	何利, 李伯全 . 复杂背景中奶粉罐内配件数量检测研究[J]. 机电工程, 2014,31(8): 1098-1102.
	HE L , LI B Q . Detection of the number of accessories in the milk powder cans in complex background[J]. Journal of Mechanical ＆ Electrical Engineering, 2014,31(8): 1098-1102.
[26]	周文静, 查志华, 吴杰 . 改进圆形 Hough 变换的田间红提葡萄果穗成熟度判别[J]. 农业工程学报, 2020,36(9): 205-213.
	ZHOU W J , ZHA Z H , WU J . Maturity discrimination of “Red Globe” grape cluster in grapery by improved circle hough transform[C]// Proceedings of Transactions of the Chinese Society of Agriculture Engineering(Transactions of the ASABE).[S.l.:s.n.], 2020,36(9): 205-213.
[27]	LIU Z H , CHENG F , ZHANG W . Recognition-based image segmentation of touching pairs of cooked shrimp (Penaeus Orientalis) using improved pruning algorithm for quality measurement[J]. Journal of Food Engineering, 2017,195(6): 166-181.

序号	CNN结构变更	分割准确率	分割时间/ms
1	删除FC1	98.51%	118.91
2	删除FC2	78.32%	103.54
3	删除Pooling1	97.45%	124.76
4	删除Pooling2	97.28%	128.54
3	只保留Conv1	68.25%	73.75
4	只保留Conv2	69.36%	69.82
5	只保留Conv1+Conv2	83.48%	83.24
6	删除Conv2+Pooling2	85.31%	112.42
7	删除Conv1+Pooling1	87.37%	108.31
8	删除Conv5	90.23%	122.13
9	删除Conv6	91.46%	124.54
	删除Conv5+Conv6	89.79	101.85
10	调整Conv1,2 卷积核数量32,64	95.14%	98.73
11	调整Conv1,2 卷积核数量64,128	96.58%	114.54
12 (基准)	调整Conv1,2 卷积核数量128,256	98.58%	132.77
13	调整Conv1,2 卷积核数量256,512	97.32%	156.48

一种融合MeanShift聚类分析和卷积神经网络的Vibe++背景分割方法

Vibe++ background segmentation method combining MeanShift clustering analysis and convolutional neural network

在线阅读

PDF下载

可视化

摘要/Abstract

引用本文

使用本文

图/表 11

参考文献 27

相关文章 10

Metrics

推荐阅读 0

[1]	陈靓, 钱亚冠, 何志强, 关晓惠, 王滨, 王星. 深度卷积神经网络的柔性剪枝策略[J]. 电信科学, 2022, 38(1): 83-94.
[2]	曹靖城, 张继东, 史国杰. 一种使用边缘增强技术提高相似图片检索召回率的方法[J]. 电信科学, 2021, 37(1): 76-84.
[3]	贾川民, 赵政辉, 王苫社, 马思伟. 基于神经网络的图像视频编码[J]. 电信科学, 2019, 35(5): 32-42.
[4]	江俊, 黄骅, 任条娟, 张登辉. 基于峰值密度聚类的电信业投诉热点话题检测方法[J]. 电信科学, 2019, 35(5): 97-103.
[5]	苏赋,吕沁,罗仁泽. 基于深度学习的图像分类研究综述[J]. 电信科学, 2019, 35(11): 58-74.
[6]	李强,康子路. 基于深度时空特征卷积—池化的视频人群计数方法[J]. 电信科学, 2018, 34(6): 72-79.
[7]	徐宏伟,严迪群,阳帆,王让定,金超,向立. 基于卷积神经网络的电子变调语音检测算法[J]. 电信科学, 2018, 34(2): 46-57.
[8]	王炳鑫,侯岩,方红旺,陈雨泽,刘建. 面向“削峰填谷”的电力客户用电行为分析[J]. 电信科学, 2017, 33(5): 164-170.
[9]	邢惠钧,昌硕. 基于移动小车的行人监控系统[J]. 电信科学, 2017, 33(2): 120-127.
[10]	翟健宏,李伟,葛瑞海,杨茹. 基于聚类与贝叶斯分类器的网络节点分组算法及评价模型[J]. 电信科学, 2013, 29(2): 51-57.