文献类型：期刊文献

中文题名：基于相关向量机回归的涂泥木线条表面点状小缺陷检测

英文题名：Detection of small point surface defects of gesso primed wood moldings based on relevant vector machine

作者：刘瑞明[1] 黄佳炜[2] 刘勇[2] 孙帅成[2] 王双永[3] 周海宾[3]

第一作者：刘瑞明

机构：[1]江苏海洋大学电子工程学院,江苏连云港222005;[2]江苏海洋大学机械与海洋工程学院,江苏连云港222005;[3]中国林业科学研究院木材工业研究所,北京100091

年份：2019

卷号：4

期号：5

起止页码：115-120

中文期刊名：林业工程学报

外文期刊名：Journal of Forestry Engineering

收录：CSTPCD;;北大核心:【北大核心2017】；CSCD:【CSCD_E2019_2020】；

基金：国家自然科学基金(31770603)

语种：中文

中文关键词：木线条;相关向量机(RVR);核相关系数;表面缺陷检测;信杂比;背景抑制因子

外文关键词：wood molding;relevance vector regression;kernel correlation coefficients;surface defect detection;signal-to-clutter ratio;background suppression factor

分类号：S781.5

摘要：木线条是一种用途广泛的建筑材料,主要用于装饰、装修和家具制造等行业。涂泥木线条具有美观、强度高、耐潮湿等优点,深受用户的青睐。在生产过程中,为了提高产品质量,需要对涂泥木线条的表面缺陷进行检测,基于数字图像处理技术的缺陷检测方法已成为主要技术手段。在图像中,面积较大的缺陷相对容易检测,但小的点状缺陷由于包含像素少、缺乏纹理特征等特点,检测非常困难。使用传统的、基于滤波的方法检测小缺陷,主要利用了小缺陷的高频特性,但高频图像噪声会造成误检,导致检测效果并不理想。本研究提出了基于相关向量回归结合后处理的方法对小缺陷进行检测,相关向量回归与支持向量回归相比,具有超参数少、表达更稀疏、核函数不需要满足梅西定理等优点。该方法通过以下步骤实现小缺陷检测:首先利用相关向量回归算法对图像进行处理,将回归值作为像素灰度值构建回归图像,然后求原始图像与回归图像的差图像,求取差图像的核相关系数后,再经过取反、二值化和局部平均等后处理,最终得到信杂比较大的检测图像。通过检测指标的比较,与基于Top?hat滤波的检测算法相比,本研究提出的方法具有更好的检测效果。
Molding is a kind of widely used building material,which is utilized primarily for decorating and furniture manufacturing.The gesso primed wood molding is favored by users because it has beautiful appearance,high strength and moisture resistance.In the molding production,surface defects of gesso primed wood moldings must be checked and eliminated in order to improve the quality of products.The defect detection methods based on digital image processing technology have been widely applied in the wood industry.It is relatively easy to detect those defects with a large area in images.However,detecting tiny point defects is very difficult due to fewer pixels and lack of texture features,which is a current technical challenge.The traditional defect detection methods based on filtering technology mainly utilize the high?frequency characteristics of tiny defects.Unfortunately,image noise has also high?frequency characteristics,which may cause false detection and the detection performance is unsatisfactory.In this paper,a method based on relevance vector regression combined with some post?processing was applied to detect tiny point defects on the gesso primed wood molding.Compared with support vector regression,the relevance vector regression does not need to set many hyper?parameters.Its representation was sparser than the support vector regression,which means less computation required by the relevance vector regression.Another advantage was that its kernel function was not necessary to satisfy the Mercer??s condition.This method executed the detection of tiny point defects with the following steps.Firstly,the image was processed by the relevance vector regression algorithm,and the regression value of every pixel was obtained.The regression value was used as the pixel gray value to construct a regression image.The pixels of tiny defects were smoothed in the regression image.The background pixels were unchanged.Then,the difference between the original image and the regression image was calculated.In the different images,the different values of tiny defects were larger than those of background.The kernel correlation coefficient of the difference image was computed.The post?processing methods,such as inversion,binarization and local averaging,were executed in order to improve the detection performance further.Finally,the detection image with high signal?to?clutter ratio and large back?ground suppression factor was constructed.By comparing the detection indicators,the proposed method had better detection performance than that of the Top?hat filter?based detection algorithm.