en
×

分享给微信好友或者朋友圈

使用微信“扫一扫”功能。
作者简介:

刘佳,女,1996年出生,硕士研究生。主要研究方向为腐蚀与防护、机器视觉和图像识别。E-mail:821775361@qq.com;

王丹,男,1982年出生,博士,副教授。主要研究方向为先进数控系统及装备、机器人技术及集成应用、超精密加工技术及装备、零排放制氢及燃料电池技术。E-mail:wangpick@163.com

通讯作者:

唐鋆磊,男,1983年出生,博士,教授,博士研究生导师。主要研究方向为电化学、腐蚀与防护、表面工程、新能源材料和航空材料。E-mail:tangjunlei@126.com

中图分类号:TG156;TB114

DOI:10.11933/j.issn.1007−9289.20221008001

参考文献 1
陶蕾,宋诗哲,金威贤.基于图像颜色特征的涂层老化失效检测方法[J].化工学报,2009,60(2):415-420.TAO Lei,SONG Shizhe,JIN Weixian.A method for detecting ageing failure of coatings based on image colour features[J].Journal of Chemical Engineering,2009,60(2):415-420.(in Chinese)
参考文献 2
申振腾,陈建平,杜厚栋,等.基于机器视觉的绝缘子RTV涂层质量检测方法研究及应用[J].电磁避雷器,2018(5):233-240.SHEN Zhenteng,CHEN Jianping,DU Houdong,et al.Research and application of machine vision-based insulator RTV coating quality inspection method[J].Electromagnetic Lightning Arrester,2018(5):233-240.(in Chinese)
参考文献 3
YEIH W,HUANG R.Detection of the corrosion damage in reinforced concrete members by ultrasonic testing[J].Cement & Concrete Research,1998,28(7):1071-1083.
参考文献 4
VANAEI H,ESLAMI A.A review on pipeline corrosion in-line inspection(ILI)and corrosion growth rate models[J].International Journal of Pressure Vessels and Piping,2017,149:43-54.
参考文献 5
WICKER M,ALDUSEl B P,JUNG S.Detection of hidden corrosion in metal roofing shingles utilizing infrared thermography[J].Journal of Building Engineering,2018,20:201-207.
参考文献 6
SHAIKHET H,SIVABHARASI N,SASI B,et al.Use of eddy current testing method in detection and evaluation of sensitisation and intergranular corrosion in austenitic stainless steels[J].Corrosion Science,2006,48(6):1462-1482.
参考文献 7
IDRIS S A,JAFAR F A.Image enhancement based on software filter optimization for corrosion inspection[C]//The 5th International Conference on Intelligent Systems,Shanghai,China,Jan 1,2014.Springer Publishing,2014:345-350.
参考文献 8
FERNÁNDEZ-ISLA C,NAVARRO P J,ALCOVER P M.Automated visual inspection of ship hull surfaces using the wavelet transform[J].Mathematical Problems in Engineering,2013,27(5):211-244.
参考文献 9
JAHANSHAHI M R,MASRI S F.Effect of color space,color channels,and sub-image block size on the performance of wavelet-based texture analysis algorithms:an application to corrosion detection on steel structures[J].Computing in Cvil Engineering,2013:685-692.
参考文献 10
SON H,HWANG N,KIM C,et al.Rapid and automated determination of rusted surface areas of a steel bridge for robotic maintenance systems[J].Automation in Construction,2014,42:13-24.
参考文献 11
XIE X.A review of recent advances in surface defect detection using texture analysis techniques[J].ELCVIA:Electronic Letters on Computer Vision and Image Analysis,2008,7(3):1-22.
参考文献 12
ZAIDAN B B,ZAIDAN A A,ALANAZI H O,et al.Towards corrosion detection system[J].International Journal of Computer Science,2010,7(3):33-36.
参考文献 13
ACOSTA M R G,DÍAZ J C V,CASTRO N S.An innovative image-processing model for rust detection using perlin noise to simulate oxide textures[J].Corrosion Science,2014,88:141-151.
参考文献 14
CHEN P H,SHEN H K,LEI C Y,et al.Fourier-transform-based method for automated steel bridge coating defect recognition[J].Procedia Engineering,2011,14:470-476.
参考文献 15
AIJAZI A K,MALATERRE L,TAZIR M L,et al.Detecting and analyzing corrosion spots on the hull of large marine vessels using colored 3d lidar point clouds[J].ISPRS Annals of Photogrammetry Remote Sensing & Spatial Informa,2016,33:153-160.
参考文献 16
PETRICCA L,MOSS T,FIGUEROA G,et al.Corrosion detection using AI:A comparison of standard computer vision techniques and deep learning model[C]//The Sixth International Conference on Computer Science,Amman,Jordan,May 21,2016.IEEE,2016:91-99.
参考文献 17
IGOE D,PARISI A V.Characterization of the corrosion of iron using a smartphone camera[J].Instrumentation Science and Technology,2016,44(2):139-147.
参考文献 18
林福宗.多媒体技术基础[M].4 版.北京:清华大学出版社,2017.LIN Fuzong.Fundamentals of multimedia technology[M].4th ed.Beijing:Tsinghua University Press,2017(in Chinese).
参考文献 19
刘敏,张海兵,林冰,等.实验室加速环境下水性快干环氧厚浆底漆的老化机理及失效过程研究[J].表面技术,2022,51(11):305-317.LIU Min,ZHANG Haibing,LIN Bing,et al.Research on aging mechanism and failure process of waterborne quick drying epoxy paste primer in accelerated environment in laboratory[J].Surface Technology,2022,51(11):305-317.(in Chinese)
参考文献 20
刘佳敏,何宁.基于改进同态滤波的低对比度图像增强 [J].计算机应用与软件,2020,37(3):220-224.LIU Jiamin,HE Ning.Low-contrast image enhancement based on improved homomorphic filtering[J].Computer Application and Software,2020,37(3):220-224.(in Chinese)
参考文献 21
陈春宁,王延杰.在频域中利用同态滤波增强图像对比度[J].微计算机信息,2007(6):264-266.CHEN Chunning,WANG Yanjie.Image contrast enhancement using homomorphicfiltering in the frequency domain,2007(6):264-266.(in Chinese)
参考文献 22
VAPNIK.Statistical learning theory[M].New York:Wiley-Interscience,1998.
参考文献 23
段先云,邓学雄,左启阳.级进模刃孔图形复杂度的研究[J].工程图学学报,2006(5):94-97.DUAN Xianyun,DENG Xuexiong,ZUO Qiyang.The research on the complexity of progressive die edge[J].Journal of Engineering Graphics,2006(5):94-97.(in Chinese)
参考文献 24
杨帆,冯翔,阮羚,等.基于皮尔逊相关系数法的水树枝与超低频介损的相关性研究[J].高压电器,2014,50(6):21-25.YANG Fan,FENG Xiang,RUAN Ling,et al.Study on the correlation between water tree branches and ultra-low frequency dielectric loss based on pearson correlation coefficient method[J].High Voltage Electrical Apparatus,2014,50(6):21-25(in Chinese).
参考文献 25
中华人民共和国国家质量监督检验检疫总局,中国国家标准化管理委员会.GB/T 1766—2008 色漆和清漆涂层老化的评级方法[S].北京:中国标准出版社,2008.General Administration of Quality Supervision,Inspection and Quarantine of the People’ s Republic of China,China National Standardization Administration.GB/T 1766— 2008 Rating Method for Aging of Paints and Varnishes[S].Beijing:China Standards Press,2008.(in Chinese)
目录contents

    摘要

    图像识别技术广泛应用于涂层领域,图像特征的选择是提升识别率的重要因素,而形状特征在涂层锈点的图像识别中未见报道。基于涂层锈点的颜色和形状特征,结合机器学习对其进行图像识别。通过采集 3 种常见自然光照强度下的 90 张涂层锈点图像,使用同态滤波对图像进行预处理,利用 HSV(色相-饱和度-明度)颜色空间来区分锈点与无锈点区域。然后提取锈点的 8 种形状特征对锈点区域进一步细化,用 Pearson 相关系数对形状特征进行筛选,将颜色特征、单一形状特征、8 种组合形状特征、筛选后的组合形状特征、颜色特征与筛选后组合形状特征的融合特征分别作为参量输入 Linear 核函数、RBF 核函数、Polynomial 核函数和 Sigmoid 核函数 4 种核函数的支持向量机(SVM)对锈点进行识别。研究结果表明:联合 SVM 与颜色、形状特征参量构建的图像识别算法能较准确地识别涂层锈点,其中基于颜色特征与筛选后形状特征的融合特征的准确识别率最高可达 93.33%。形状特征可作为另一种特征信息来提高锈点图像识别的精确度,可为涂层锈点的图像识别技术研究提供参考依据。

    Abstract

    Image recognition technology is used extensively in various industrial fields, and rust spots are a form of commonly occurring damage to coatings in service. Further, the quality inspection of coating surfaces relies primarily on manual work, which is limited by its subjectivity. Therefore, based on machine learning and image processing technology, a recognition method for coating rust spots is proposed that includes a complete set of recognition systems for rust spots from image acquisition, homomorphic filtering preprocessing, feature extraction and screening, and support vector machine(SVM) judgment recognition. Because the coating aging failure is usually detected outdoors, 90 images of coating rust spots under three common natural light intensities, namely sunny days with smooth light, sunny days with backlight, and cloudy days, were collected in this study. Among them, 60 were randomly selected as the test set and the remaining 30 were used as the training set. First, the images were pre-processed using homomorphic filtering to reduce the effect of illumination on the image quality and to facilitate accurate feature extraction. Subsequently, the hue, saturation and value(HSV) color space was used to distinguish the rusted and not-rusted areas which were identified by controlling the value of the H component. Additionally, the eight shape features of the rusted areas were extracted to further refine the rusted areas: rusted area (S), rusted perimeter (L), rusted area (L), rusted area (L), and rusted area (L). These eight shape features were the rust spot area (S), perimeter (L), minimum external rectangle and minimum external ellipse (Sc), roundness (ƒ1), complexity (ƒ2), elongation (ƒ3), compactness (ƒ4), and area concavity ratio (ƒ5). As the number of images in this paper was small and the number of feature parameters was large, it was easy to produce overfitting, thus, the shape features were filtered using the Pearson correlation coefficient, and the area (S), perimeter (L), complexity (ƒ2), and area-concave ratio (ƒ5). Finally, the color features, single shape features, eight combined shape features, screened combined shape features, and fused features of color features and screened combined shape features were input as parameters into the linear kernel function, radials basis function, polynomial kernel function, and sigmoid kernel function, respectively. An SVM with these four kernel functions was used to identify the rusty points. The results showed that the correct recognition rate of rust spots by the SVM based on single-shape features was lower than that based on combined shape features, indicating that the single-shape features cannot accurately describe the rust spots. The correct recognition rate of the four kernel functions based on the four screened shape features was higher than that of the eight unscreened shape features, indicating that the screening of shape features using Pearson's correlation coefficient was effective. The correct recognition rate of the fusion features based on the color features and screened shape features was further improved, and its average correct recognition rate was 89%, among which the SVM recognition of the Linear kernel function was the greatest, and the correct recognition rate reached 93.33%, indicating that the algorithm based on the fusion features could significantly reduce the respective interference factors of the color and shape in the case of the small samples. The SVM can be developed into a classifier that can be used to identify the rust spots on the coating surface. The shape features of the rust spots can provide other feature information to improve the accuracy of image recognition. Further, the machine learning algorithm of the color and shape fusion features is effective and can identify rust spots quickly and accurately compared with the traditional color-based algorithm.

  • 0 前言

  • 涂料涂装作为一种有效的材料保护技术,已广泛应用于各个行业。涂层在服役过程中受到物理或化学因素的破坏,会导致涂层的外观会发现一些变化,如起泡、锈点、变色、龟裂等[1]。其中,锈点是影响涂层质量和使用寿命的主要因素之一。目前,涂层表面的质量检测主要依靠于人工[2],而人工对涂层表面的锈点检测又受限于其主观性,很难客观地量化锈点的特征。因此,研究快速准确识别涂层锈点的技术手段,有利于及时科学防治,从而提高涂层的使用寿命。

  • 表面腐蚀可以通过传统的无损检测(NDT)技术进行检测和评估,如超声检测[3]、磁粉探伤[4]、红外[5]和涡流检测[6]等。这些方法检测周期较长且专业性强。近年来,数字图像处理技术广泛应用于生物医学、航空航天技术和地球科学等各个领域。在腐蚀领域,腐蚀损伤的形式和特点可以通过腐蚀图像来表示,以评估腐蚀类型和分析腐蚀程度,并成为研究腐蚀规律的重要基础。通过建立合适的分割准则和识别模型就能够实现对腐蚀区域的定量描述,对腐蚀的检测问题可以转化为计算机视觉问题。传统的一般方法是对腐蚀图像采用小波变换[7-9],提取各子图像中的能量、熵值[10],并以此来检测图像中的腐蚀区域;有学者利用锈蚀区域和未锈蚀区域的颜色差异[710],对图像进行分割;也有学者认为,腐蚀的加剧,金属表面粗糙度[911]也随之发生变化,反映在腐蚀图像中,腐蚀边缘处像素的灰度值与其他处的灰度值不同[12-13],并以此区分腐蚀和非腐蚀部位;更进一步,有学者提出用无损评估(NDE)[914]分析腐蚀表面的纹理变化[11-12],再结合自组织特征映射网络(SOM)[13]对腐蚀损伤分类,或者训练支持向量机(SVM)[1014]对腐蚀进行分类和检测,而分类的依据基于腐蚀区域的颜色特性[15-17]

  • 锈点是涂层服役过程中较为常见的损伤,在图像识别领域,部分学者采用基于颜色模型的方法[710],在各种图像条件下从背景中分割腐蚀区域,通常依赖于颜色通道包含的信息。然而,基于颜色的算法容易受到图像采集的光照条件影响,且单独的颜色不足以准确区分涂层表面的不同损伤。

  • 本文提出颜色和形状特征同时应用,并结合支持向量机的锈点图像识别方法,减少了仅使用颜色特征识别时光照和其他损伤带来的干扰信息,有助于提高锈点的识别率,且使用形状特征对锈点进行判断和识别,目前没有相关报道。本文首先提取锈点 HSV 颜色模型中的 H 通道的信息作为颜色特征的输入,HSV 颜色模型是人类视觉系统的色彩感知方式之一,包括色相(H)、饱和度(S)、明度 (V)[18],相对于 RGB 颜色模型更能准确地描述锈点的色相特点;然后提取锈点的 8 种形状特征,考虑到本研究的样本数量不够充足,容易产生过拟合的风险,使用 Pearson 相关系数的排名筛选策略,将得到的相关性较高的 4 种形状特征作为 SVM 的输入,为了证实这一策略的可靠性,也将单一的形状特征、8 种组合形状特征作为模型的输入,对比其识别率,证明筛选策略是有效的。此外,也对比了颜色和形状特征同时应用的 4 种不同核函数的 SVM 的识别率,探索更精确的机器学习方法。

  • 1 图像采集及预处理

  • 1.1 图像采集

  • 涂层样品取自实验室加速老化试验环境,以 Q235 型碳钢为基材,涂料为双组份水性快干环氧厚浆漆(6DV,中远佐敦,青岛)[19]。图像采集装置如图1 所示:工业相机(MV-CA060-10GC,海康威视,杭州)有效像素 2000 万;变焦镜头(H281050-6 MP,威科迈,厦门)焦距 10~50 mm;条形光源 (KM-BRD30030,威科迈,厦门)最大亮度 140 000 lx。涂层老化失效检测通常在户外现场检测,因此选取 3 种常见天气的光照强度值进行采集,用条形光源调节亮度值,其光照强度范围见表1。本文共使用 30 个不同的具有锈点的涂层样品,并对每个样品进行编号,拍摄照片时,保持相机和光源与涂层样品垂直,入射光线与拍照角度平行,拍摄照片共 90 张,不同光照条件下各 30 张,图像大小为 1 250 ×2 400 像素,格式为 JPEG,将所有图像随机分为两组,一组包含 60 张作为训练集,一组包含 30 张作为测试集。采集的部分图像如图2 所示,对所有图像进行分类标注,将晴天顺光、晴天逆光、阴天分别用 A、B、C 表示,如晴天顺光下的样品 1 标注为 A1。

  • 图1 图像采集装置

  • Fig.1 Image capture device

  • 表1 三种常见天气条件下的光照强度

  • Table1 Light intensity in three common weather conditions

  • 图2 采集的部分图像

  • Fig.2 Part of the image captured

  • 1.2 图像预处理

  • 对采集的涂层锈点图像作预处理,不仅能使锈点与涂层背景对比更清晰,还更能准确获取涂层锈点的颜色与形状特征,从而更有效地对锈点进行识别和分类。本文采集的图像是不同光照强度下的涂层锈点图像,光线对成像的影响较大,光照分布不均匀会影响涂层表面的锈点识别和定位。因此,对原始图像使用同态滤波进行预处理。从图3 可以看出,在晴天逆光和阴天的光照条件下,采用同态滤波后,整幅图像光照分布更加均匀,且锈点边缘得到了一定程度的锐化。同态滤波的原理是通过利用图像的照度分量和反射分量使照明更均匀来加强阴影区域的细节。其数学模型为[20]

  • f(x,y)=fi(x,y)fr(x,y)
    (1)
  • 式中,fi 表示涂层随空间位置不同时的光照强度分量,fr代表涂层反射到人眼的反射分量。

  • 图3 不同光照强度下的同态滤波图

  • Fig.3 Homomorphic filtering under different illumination strips: (a1) - (c1) original images, (a2) - (c2) after homomorphic filtering

  • 本文采用巴特沃斯滤波函数[21]为:

  • H(u,v)=rH-rL1+cD0nD(u,v)2+rL
    (2)
  • 式中,rH代表高频增益,rL代表低频增益,D0为截至频率,c 为常数。

  • 进行傅里叶和指数变换,得到增强的图像 gxy

  • g(x,y)=exphi(u,v)exphr(u,v)
    (3)
  • 1.3 SVM 分类器

  • SVM 支持向量机(Support vector machines)是由 VAPNIK[22]提出的一种基于统计学理论的机器学习算法,多用于小样本数据的学习、分类和预测,其超平面表达式为:

  • ωφ(x)+b=0
    (4)
  • 式中,ω 表示超平面法向量,ϕ 为映射函数,b 为分类阈值。由于不同的核函数代表不同的 SVM 性能,影响对涂层锈点的识别率。本文分别采用线性核函数(Linear)、多项式核函数(Polynomical)、径向基核函数(RBF)、Sigmoid 核函数建立的 SVM 对融合特征的涂层锈点进行分类识别。

  • 2 涂层锈点的特征提取

  • 2.1 锈点的颜色特征

  • HSV 颜色模型更侧重于色彩表示,此模型受光照影响较小,且与人眼对颜色的主观认识相对比较符合,因此选用 HSV 空间模型。由采集的原始图像观察可知,涂层表面的锈点及腐蚀产物多表现为红棕色、黄色、暗褐色,这些颜色多介于品红与红色、红色与黄色之间。图4 显示了 H 通道控制不同颜色时的取值范围,对应选择 H 的取值范围为 0~0.167、 0.883~1。

  • 图4 H 通道控制不同颜色的值范围

  • Fig.4 H channel controls the value range of different colors

  • 2.2 锈点的形状特征

  • 2.2.1 形状特征提取

  • 在样本图像中,锈点的形状大多呈圆形、椭圆型及其他不规则形状,需要将人眼观察到的形状特征从图像中提取出来,并建立一组特征向量能够准确的描述锈点。

  • 如图5 所示,选取锈点轮廓的周长、面积、长、宽、外切圆直径、外切矩形为基本几何特征,使用 Matlab 数学软件对几何特征进行计算并提取所有的形状特征参数,以下为本文所需的形状特征。

  • (1)锈点面积(S):涂层表面锈点区域的像素之和。

  • (2)周长(L):涂层锈点区域的外轮廓所有像素的中心距离之和。

  • (3)最小外接矩形和最小外接椭圆(Sc):能够包含锈点的最小面积的矩形和椭圆。

  • (4)圆形度(f1):锈点单位面积周长的大小,判断锈点接近于圆形的程度。计算公式:

  • f1=4πSL2
    (5)
  • (5)复杂度(f2):以锈点的周长、面积为尺度描述复杂度,圆的复杂度定为 0 [23]。计算公式:

  • f2=1-4πSL2
    (6)
  • (6)伸长度(f3):锈点外接矩形的宽与长的比值。计算公式:

  • f3=min{a,b}max{a,b}
    (7)
  • 式中 ab 表示外接矩形的宽和长。

  • (7)紧凑度(f4):锈点的最小内切圆直径 d 与长轴 b 之比。计算公式:

  • f4=db
    (8)
  • (8)面积凹凸比(f5):锈点面积与最小外接矩形和最小外接椭圆面积的比值定义为凹凸性。计算公式:

  • f5=SSc
    (9)
  • 图5 形状特征示意图

  • Fig.5 Sketch map of shape feature

  • 2.2.2 形状特征的筛选

  • 在提取大量涂层锈点图像形状特征参数后,由于图像样本数据之间存在相关性,且本文图像数量相对较少,特征数量较多,易产生过拟合的风险,在后续识别中会导致识别精度降低。因此,本文通过计算形状特征的皮尔逊相关系数 p 进行特征筛选,去除与目标值相关度较低的特征量,缩短训练时间,提高识别精度。皮尔逊相关系数法是可以描述变量之间关系密切程度的一种统计学方法[24],其计算公式如下:

  • p=i=1n Xi-X-Yi-Y-i=1n Xi-X2i=1n Yi-Y-2
    (10)
  • 式中,n 为样本数量,Xi 为样本的特征值,X-为平均特征值,Y-Yi 的平均值,Yi 为涂层锈点程度 [25],根据 GB / T1766—2008,涂层锈点等级评定体系基于表面的锈点数量来确定,见表2。p 值介于−1 和 1 之间,通常以绝对值表示,其值域等级解释见表3。

  • 表2 锈点数量等级

  • Table2 Grade of the number of rust spots

  • 表3 皮尔逊相关系数值域等级

  • Table3 Range grade of Pearson correlation coefficient

  • 2.3 基于融合特征的涂层锈点识别方法

  • 通过以上分析,本文根据涂层锈点的颜色和形状特征,提出的锈点识别流程图如图6 所示。

  • 图6 算法流程图

  • Fig.6 Algorithm flow chart

  • (1)采集 3 种不同光照强度下的原始涂层锈点图像,采用同态滤波的方法进行预处理,得到光线更均匀、锈点更清晰的图像。

  • (2)根据锈点的色相特征,选择 H 分量取值范围,提取锈点的颜色特征。

  • (3)根据锈点的形状特点,计算其 8 种形状特征值,为提高后续计算精度,对 8 种特征值用皮尔逊相关系数进行筛选,得到 4 种相关性较高的特征值。

  • (4)将颜色特征和筛选后的形状特征输入 SVM,再用 SVM 识别测试集中的图像,判断和识别图中的锈点。

  • 3 结果与讨论

  • 3.1 涂层锈点的特征分析

  • 3.1.1 颜色特征

  • 图7 显示了 H 通道取值范围为 0~0.167、 0.883~1 下的识别效果,可以看出,使用颜色模型可以识别出大部分的锈点区域,但是涂层表面容易被腐蚀产物遮盖,且涂层表面的锈迹与锈点颜色接近,容易被误识别为锈点。

  • 3.1.2 形状特征

  • 表4 列示了 3 种光照条件下的部分样品的 8 种形状特征参数,可以看出,相同光照条件下的不同样品上锈点的形状特征存在差异,这是因为不同样品的锈点生长周期不同,表现出的形状特征也不同;不同光照条件下的相同样品的锈点特征参数值也存在差异,但差异较小,这是因为在图像预处理时,使用了同态滤波,减少了光照对图像的影响。

  • 图7 通过 H 分量提取的锈点部分

  • Fig.7 Rust spot area extracted by H component

  • 表4 部分图像的 8 种形状特征参数值

  • Table4 Value of eight shape feature parameters of partial images

  • 表5 列出了 8 种形状特征的 p 值,结合表2 可知,面积 S、复杂度 ƒ2 与涂层锈点程度强相关,周长 L、面积凹凸比 ƒ5 与涂层锈点程度中等程度相关,圆形度 ƒ1、紧凑度 ƒ4 与涂层锈点程度弱相关,伸长度 ƒ3 与涂层锈点程度极弱相关。

  • 表5 8 种形状特征的相关性系数

  • Table5 Pearson correlation coefficient of each shape feature

  • 图8 为这 4 种形状特征参数的频率分布图,如图8a 所示,锈点面积大小 S 分布在 0~0.4 mm2;如图8b 所示,复杂度 ƒ2 大小范围在 5~10;如图8c 所示,周长 L 分布在 0~5 mm; 如图8d 所示,面积凹凸比 ƒ5 在 0~3。本文选取强相关和中等程度相关的面积 S、复杂度 ƒ2、周长 L、面积凹凸比 ƒ5 作为输入的形状特征。

  • 图8 4 种形状特征参数的频率分布直方图

  • Fig.8 Frequency distribution histogram of four shape feature parameters

  • 3.2 基于单一形状特征 SVM 的识别结果

  • 在 SVM 模型中,为验证各特征的分类可行性,以识别率作为评价指标,识别率等于所有分类正确的样本除以总样本。基于单一的形状特征识别锈点,是将采集的图像中所有形状特征的值分布范围通过 matlab 提取出来,如锈点面积大小分布在 0~0.4 mm2,将该范围作为参数输入 SVM 分类器中,作为对锈点识别的依据。表6 列示了单一形状特征作为参数输入的 SVM 识别率,可以看出,最小外接矩形 Sc、圆形度 ƒ1、伸长度 ƒ3、紧凑度 ƒ4作为单一特征时的不同核函数 SVM 对涂层锈点的识别率较低,均在 41%以下;锈点面积 S、周长 L、复杂度 ƒ2、面积凹凸比 ƒ5 作为单一特征时的识别率,虽然略高于圆形度 ƒ1、伸长度 ƒ3、紧凑度 ƒ4、最小外接矩形 Sc 作为特征参数时的识别率,但综合来看,单一形状特征作为参数输入的 SVM 对涂层锈点的识别率不高,均低于 65%。原因是涂层表面除锈点损伤外,还存在其他浮尘颗粒和微小起泡,而这些浮尘和起泡具有和锈点相似的形状特征;再者,涂层表面的锈点生长阶段不同,表现的形貌有所不同,如锈点在早期时,其表面形貌呈圆形或椭圆形,但随着腐蚀周期延长,锈点的表面形貌会出现无定式的变化,这是因为表面锈点密度变大,锈点增多,锈点之间会出现合并的现象。所以,单一的形状特征无法准确描述在多种损伤混合的涂层表面和不同生长阶段的锈点。

  • 表6 基于单一形状特征的 SVM 识别率(%)

  • Table6 Recognition rate of SVM based on single shape feature (%)

  • 3.3 基于组合形状特征 SVM 的识别结果

  • 表7 列示了组合形状特征的 SVM 识别率,本文所有 8 种形状特征的不同核函数的 SVM 识别率在 68%以上,均低于筛选后的 4 种形状特征组合的识别率,这是因为本文图像样本数量相对较少,特征数量较多,产生了过拟合。筛选后的形状特征组合的识别率均低于 85%,其原因是,在采集图像之前未对涂层表面做任何清洁处理,存在一些浮尘或颗粒,会被误识为锈点,所以需要结合锈点的颜色特征。

  • 表7 基于组合形状特征的 SVM 识别率(%)

  • Table7 Recognition rate of SVM based on combined shape feature (%)

  • 3.4 基于 HSV 与形状特征的融合特征 SVM 的识别结果

  • 图9 显示了融合 HSV 与筛选后的 4 种形状特征在不同核函数下的识别率,由图可知,基于 HSV 颜色特征的 SVM 识别率最高为 85.71 %,基于筛选后的 4 种形状特征的 SVM 识别率最高为 87.52 %,而基于融合特征的 SVM 识别率最高为 93.33 %。因此,须要结合锈点的形状特征,消除通过颜色特征被误识别的锈点区域。对于涂层的锈点识别,应该选择 HSV 和形状的融合特征。

  • 对比 4 种核函数的 SVM 对涂层锈点的识别率,基于融合特征的 Linear 核函数的 SVM 对涂层锈点的识别率最高,为 93.33 %;基于融合特征的 Sigmoid 和 RBF 核函数的 SVM 次之,分别为 89.30%、 88.00%;基于融合特征的 Polynomial 核函数的 SVM的识别率最低,为 86.00%,表明 Linear 核函数的融合特征 SVM 对涂层锈点的识别分类性能较好。图10 为基于融合特征的不同核函数的 SVM 部分识别效果图,识别到的锈点用红色圆圈表示,从图中可以看出相同图片使用不同核函数识别的效果不一样,针对同一数据集,核函数在一定程度上会影响识别的精度。

  • 图9 不同核函数的 SVM 的锈点识别率

  • Fig.9 Rust point recognition rate of pitting of SVM with different kernel function types

  • 3.5 结果分析

  • 锈点是涂层在服役过程中常见的损伤类型,在服役环境和时间的变化下,锈点的生长情况也会发生变化。在锈点萌发的初期,锈点的形貌表现为较规则的面积较小的圆形或椭圆形,颜色多为黄色或浅棕色;随着腐蚀周期的加长,涂层表面出现更为密集和凹凸不平的锈点,且相邻的锈点存在粘连的现象,其面积和轮廓周长逐渐增大,形状呈不规则状,且锈点的颜色也随之加深,多为深棕色或暗褐色。锈点的萌发与生长伴随着颜色和形状的变化,在颜色方面的变化较为明显,容易被观察到。因此,颜色特征能够大致识别和判断锈点,但采集图像环境的光照强度会影响视觉对颜色的感知。虽然同态滤波图像预处理降低了这种影响,但是无法达到理想的效果(图3)。涂层服役失效过程中,表面不止锈点会释放颜色信息,金属腐蚀产物迁移痕迹、环境污染物等印迹很容易被误识别为锈点颜色,导致基于 HSV 颜色特征的识别算法精确度受到限制。锈点的形状特征可提供另一种特征信息来提高图像识别的精确度,在本研究中,提取所有锈点的面积、复杂度、周长、面积凹凸比这 4 种形状特征,在锈点总面积、周长、最小外接矩形、圆形度、复杂度、伸长度、紧凑度、面积凹凸比等特征信息中,识别效果最好,但同样会受到锈迹、表面颗粒物等的干扰,其识别率最高只能达到 81.86%。颜色和形状特征同时使用时,平均识别率为 89%,其中 Linear 核函数工作效果最好,识别率为 93.33%,在小样本情况下,可以显著减少颜色和形状的各自干扰因素,比使用 HSV 颜色模型的识别率提升 7.62%,比使用形状特征的识别率提升 5.81%。

  • 图10 基于融合特征的不同核函数的 SVM 识别结果

  • Fig.10 Recognition results of SVM with different kernel functions based on fused features

  • 为进一步提升锈点的识别率,在未来工作中可做以下改进:一方面,增加训练数据集的大小,可将本文方法作为未来深度学习的训练标签,利用深度学习强大的学习能力训练大样本的模型;另一方面,提高训练数据的图像质量,如更高的对比度、分辨率等。

  • 4 结论

  • (1)使用不同核函数的支持向量机对基于 HSV、筛选后的形状特征、融合特征的锈点进行识别,结果表明支持向量机可发展为用于涂层锈点识别的分类器。

  • (2)引入形状特征对锈点进行描述,识别率有明显提升,减少了因颜色特征而导致的误识别。

  • (3)基于颜色和形状融合特征的锈点识别,对实现现场条件下涂层锈点的自动化检测可提供参考依据。

  • 参考文献

    • [1] 陶蕾,宋诗哲,金威贤.基于图像颜色特征的涂层老化失效检测方法[J].化工学报,2009,60(2):415-420.TAO Lei,SONG Shizhe,JIN Weixian.A method for detecting ageing failure of coatings based on image colour features[J].Journal of Chemical Engineering,2009,60(2):415-420.(in Chinese)

    • [2] 申振腾,陈建平,杜厚栋,等.基于机器视觉的绝缘子RTV涂层质量检测方法研究及应用[J].电磁避雷器,2018(5):233-240.SHEN Zhenteng,CHEN Jianping,DU Houdong,et al.Research and application of machine vision-based insulator RTV coating quality inspection method[J].Electromagnetic Lightning Arrester,2018(5):233-240.(in Chinese)

    • [3] YEIH W,HUANG R.Detection of the corrosion damage in reinforced concrete members by ultrasonic testing[J].Cement & Concrete Research,1998,28(7):1071-1083.

    • [4] VANAEI H,ESLAMI A.A review on pipeline corrosion in-line inspection(ILI)and corrosion growth rate models[J].International Journal of Pressure Vessels and Piping,2017,149:43-54.

    • [5] WICKER M,ALDUSEl B P,JUNG S.Detection of hidden corrosion in metal roofing shingles utilizing infrared thermography[J].Journal of Building Engineering,2018,20:201-207.

    • [6] SHAIKHET H,SIVABHARASI N,SASI B,et al.Use of eddy current testing method in detection and evaluation of sensitisation and intergranular corrosion in austenitic stainless steels[J].Corrosion Science,2006,48(6):1462-1482.

    • [7] IDRIS S A,JAFAR F A.Image enhancement based on software filter optimization for corrosion inspection[C]//The 5th International Conference on Intelligent Systems,Shanghai,China,Jan 1,2014.Springer Publishing,2014:345-350.

    • [8] FERNÁNDEZ-ISLA C,NAVARRO P J,ALCOVER P M.Automated visual inspection of ship hull surfaces using the wavelet transform[J].Mathematical Problems in Engineering,2013,27(5):211-244.

    • [9] JAHANSHAHI M R,MASRI S F.Effect of color space,color channels,and sub-image block size on the performance of wavelet-based texture analysis algorithms:an application to corrosion detection on steel structures[J].Computing in Cvil Engineering,2013:685-692.

    • [10] SON H,HWANG N,KIM C,et al.Rapid and automated determination of rusted surface areas of a steel bridge for robotic maintenance systems[J].Automation in Construction,2014,42:13-24.

    • [11] XIE X.A review of recent advances in surface defect detection using texture analysis techniques[J].ELCVIA:Electronic Letters on Computer Vision and Image Analysis,2008,7(3):1-22.

    • [12] ZAIDAN B B,ZAIDAN A A,ALANAZI H O,et al.Towards corrosion detection system[J].International Journal of Computer Science,2010,7(3):33-36.

    • [13] ACOSTA M R G,DÍAZ J C V,CASTRO N S.An innovative image-processing model for rust detection using perlin noise to simulate oxide textures[J].Corrosion Science,2014,88:141-151.

    • [14] CHEN P H,SHEN H K,LEI C Y,et al.Fourier-transform-based method for automated steel bridge coating defect recognition[J].Procedia Engineering,2011,14:470-476.

    • [15] AIJAZI A K,MALATERRE L,TAZIR M L,et al.Detecting and analyzing corrosion spots on the hull of large marine vessels using colored 3d lidar point clouds[J].ISPRS Annals of Photogrammetry Remote Sensing & Spatial Informa,2016,33:153-160.

    • [16] PETRICCA L,MOSS T,FIGUEROA G,et al.Corrosion detection using AI:A comparison of standard computer vision techniques and deep learning model[C]//The Sixth International Conference on Computer Science,Amman,Jordan,May 21,2016.IEEE,2016:91-99.

    • [17] IGOE D,PARISI A V.Characterization of the corrosion of iron using a smartphone camera[J].Instrumentation Science and Technology,2016,44(2):139-147.

    • [18] 林福宗.多媒体技术基础[M].4 版.北京:清华大学出版社,2017.LIN Fuzong.Fundamentals of multimedia technology[M].4th ed.Beijing:Tsinghua University Press,2017(in Chinese).

    • [19] 刘敏,张海兵,林冰,等.实验室加速环境下水性快干环氧厚浆底漆的老化机理及失效过程研究[J].表面技术,2022,51(11):305-317.LIU Min,ZHANG Haibing,LIN Bing,et al.Research on aging mechanism and failure process of waterborne quick drying epoxy paste primer in accelerated environment in laboratory[J].Surface Technology,2022,51(11):305-317.(in Chinese)

    • [20] 刘佳敏,何宁.基于改进同态滤波的低对比度图像增强 [J].计算机应用与软件,2020,37(3):220-224.LIU Jiamin,HE Ning.Low-contrast image enhancement based on improved homomorphic filtering[J].Computer Application and Software,2020,37(3):220-224.(in Chinese)

    • [21] 陈春宁,王延杰.在频域中利用同态滤波增强图像对比度[J].微计算机信息,2007(6):264-266.CHEN Chunning,WANG Yanjie.Image contrast enhancement using homomorphicfiltering in the frequency domain,2007(6):264-266.(in Chinese)

    • [22] VAPNIK.Statistical learning theory[M].New York:Wiley-Interscience,1998.

    • [23] 段先云,邓学雄,左启阳.级进模刃孔图形复杂度的研究[J].工程图学学报,2006(5):94-97.DUAN Xianyun,DENG Xuexiong,ZUO Qiyang.The research on the complexity of progressive die edge[J].Journal of Engineering Graphics,2006(5):94-97.(in Chinese)

    • [24] 杨帆,冯翔,阮羚,等.基于皮尔逊相关系数法的水树枝与超低频介损的相关性研究[J].高压电器,2014,50(6):21-25.YANG Fan,FENG Xiang,RUAN Ling,et al.Study on the correlation between water tree branches and ultra-low frequency dielectric loss based on pearson correlation coefficient method[J].High Voltage Electrical Apparatus,2014,50(6):21-25(in Chinese).

    • [25] 中华人民共和国国家质量监督检验检疫总局,中国国家标准化管理委员会.GB/T 1766—2008 色漆和清漆涂层老化的评级方法[S].北京:中国标准出版社,2008.General Administration of Quality Supervision,Inspection and Quarantine of the People’ s Republic of China,China National Standardization Administration.GB/T 1766— 2008 Rating Method for Aging of Paints and Varnishes[S].Beijing:China Standards Press,2008.(in Chinese)

  • 参考文献

    • [1] 陶蕾,宋诗哲,金威贤.基于图像颜色特征的涂层老化失效检测方法[J].化工学报,2009,60(2):415-420.TAO Lei,SONG Shizhe,JIN Weixian.A method for detecting ageing failure of coatings based on image colour features[J].Journal of Chemical Engineering,2009,60(2):415-420.(in Chinese)

    • [2] 申振腾,陈建平,杜厚栋,等.基于机器视觉的绝缘子RTV涂层质量检测方法研究及应用[J].电磁避雷器,2018(5):233-240.SHEN Zhenteng,CHEN Jianping,DU Houdong,et al.Research and application of machine vision-based insulator RTV coating quality inspection method[J].Electromagnetic Lightning Arrester,2018(5):233-240.(in Chinese)

    • [3] YEIH W,HUANG R.Detection of the corrosion damage in reinforced concrete members by ultrasonic testing[J].Cement & Concrete Research,1998,28(7):1071-1083.

    • [4] VANAEI H,ESLAMI A.A review on pipeline corrosion in-line inspection(ILI)and corrosion growth rate models[J].International Journal of Pressure Vessels and Piping,2017,149:43-54.

    • [5] WICKER M,ALDUSEl B P,JUNG S.Detection of hidden corrosion in metal roofing shingles utilizing infrared thermography[J].Journal of Building Engineering,2018,20:201-207.

    • [6] SHAIKHET H,SIVABHARASI N,SASI B,et al.Use of eddy current testing method in detection and evaluation of sensitisation and intergranular corrosion in austenitic stainless steels[J].Corrosion Science,2006,48(6):1462-1482.

    • [7] IDRIS S A,JAFAR F A.Image enhancement based on software filter optimization for corrosion inspection[C]//The 5th International Conference on Intelligent Systems,Shanghai,China,Jan 1,2014.Springer Publishing,2014:345-350.

    • [8] FERNÁNDEZ-ISLA C,NAVARRO P J,ALCOVER P M.Automated visual inspection of ship hull surfaces using the wavelet transform[J].Mathematical Problems in Engineering,2013,27(5):211-244.

    • [9] JAHANSHAHI M R,MASRI S F.Effect of color space,color channels,and sub-image block size on the performance of wavelet-based texture analysis algorithms:an application to corrosion detection on steel structures[J].Computing in Cvil Engineering,2013:685-692.

    • [10] SON H,HWANG N,KIM C,et al.Rapid and automated determination of rusted surface areas of a steel bridge for robotic maintenance systems[J].Automation in Construction,2014,42:13-24.

    • [11] XIE X.A review of recent advances in surface defect detection using texture analysis techniques[J].ELCVIA:Electronic Letters on Computer Vision and Image Analysis,2008,7(3):1-22.

    • [12] ZAIDAN B B,ZAIDAN A A,ALANAZI H O,et al.Towards corrosion detection system[J].International Journal of Computer Science,2010,7(3):33-36.

    • [13] ACOSTA M R G,DÍAZ J C V,CASTRO N S.An innovative image-processing model for rust detection using perlin noise to simulate oxide textures[J].Corrosion Science,2014,88:141-151.

    • [14] CHEN P H,SHEN H K,LEI C Y,et al.Fourier-transform-based method for automated steel bridge coating defect recognition[J].Procedia Engineering,2011,14:470-476.

    • [15] AIJAZI A K,MALATERRE L,TAZIR M L,et al.Detecting and analyzing corrosion spots on the hull of large marine vessels using colored 3d lidar point clouds[J].ISPRS Annals of Photogrammetry Remote Sensing & Spatial Informa,2016,33:153-160.

    • [16] PETRICCA L,MOSS T,FIGUEROA G,et al.Corrosion detection using AI:A comparison of standard computer vision techniques and deep learning model[C]//The Sixth International Conference on Computer Science,Amman,Jordan,May 21,2016.IEEE,2016:91-99.

    • [17] IGOE D,PARISI A V.Characterization of the corrosion of iron using a smartphone camera[J].Instrumentation Science and Technology,2016,44(2):139-147.

    • [18] 林福宗.多媒体技术基础[M].4 版.北京:清华大学出版社,2017.LIN Fuzong.Fundamentals of multimedia technology[M].4th ed.Beijing:Tsinghua University Press,2017(in Chinese).

    • [19] 刘敏,张海兵,林冰,等.实验室加速环境下水性快干环氧厚浆底漆的老化机理及失效过程研究[J].表面技术,2022,51(11):305-317.LIU Min,ZHANG Haibing,LIN Bing,et al.Research on aging mechanism and failure process of waterborne quick drying epoxy paste primer in accelerated environment in laboratory[J].Surface Technology,2022,51(11):305-317.(in Chinese)

    • [20] 刘佳敏,何宁.基于改进同态滤波的低对比度图像增强 [J].计算机应用与软件,2020,37(3):220-224.LIU Jiamin,HE Ning.Low-contrast image enhancement based on improved homomorphic filtering[J].Computer Application and Software,2020,37(3):220-224.(in Chinese)

    • [21] 陈春宁,王延杰.在频域中利用同态滤波增强图像对比度[J].微计算机信息,2007(6):264-266.CHEN Chunning,WANG Yanjie.Image contrast enhancement using homomorphicfiltering in the frequency domain,2007(6):264-266.(in Chinese)

    • [22] VAPNIK.Statistical learning theory[M].New York:Wiley-Interscience,1998.

    • [23] 段先云,邓学雄,左启阳.级进模刃孔图形复杂度的研究[J].工程图学学报,2006(5):94-97.DUAN Xianyun,DENG Xuexiong,ZUO Qiyang.The research on the complexity of progressive die edge[J].Journal of Engineering Graphics,2006(5):94-97.(in Chinese)

    • [24] 杨帆,冯翔,阮羚,等.基于皮尔逊相关系数法的水树枝与超低频介损的相关性研究[J].高压电器,2014,50(6):21-25.YANG Fan,FENG Xiang,RUAN Ling,et al.Study on the correlation between water tree branches and ultra-low frequency dielectric loss based on pearson correlation coefficient method[J].High Voltage Electrical Apparatus,2014,50(6):21-25(in Chinese).

    • [25] 中华人民共和国国家质量监督检验检疫总局,中国国家标准化管理委员会.GB/T 1766—2008 色漆和清漆涂层老化的评级方法[S].北京:中国标准出版社,2008.General Administration of Quality Supervision,Inspection and Quarantine of the People’ s Republic of China,China National Standardization Administration.GB/T 1766— 2008 Rating Method for Aging of Paints and Varnishes[S].Beijing:China Standards Press,2008.(in Chinese)

  • 手机扫一扫看