基于改进级联R-CNN的面料疵点检测方法

来源:用户上传      作者:许胜宝 郑�f默 袁德成

　　摘 要：由于布匹疵点种类分布不均，部分疵点具有极端的宽高比，而且小目标较多，导致检测难度大，因此提出一种改进级联R-CNN的布匹疵点检测方法。针对小目标问题，在R-CNN部分采用在线难例挖掘，加强对小目标的训练;针对布匹疵点极端的长宽比，在特征提取网络中采用了可变形卷积v2来代替传统的正方形卷积，并结合布匹特征重新设计边界框比例。最后采用完全交并比损失作为边界框回归损失，获取更精确的目标边界框。结果表明：对比改进前的模型，改进后的模型预测边界框更加精确，对小目标的疵点检测效果更好，在准确率上提升了3.57%，平均精确度均值提升了6.45%，可以更好地满足面料疵点的检测需求。
　　关键词：级联R-CNN;面料疵点;检测;可变形卷积v2;在线难例挖掘;完全交并比损失
　　中图分类号：TS101.8;TP391
　　文献标志码：A
　　文章编号：1009-265X（2022）02-0048-09
　　收稿日期：20210607 网络出版日期：20210803
　　基金项目：国家重点研发计划“智能机器人专项”项目（2018YFB1308803）
　　作者简介：许胜宝（1993-），男，辽宁丹东人，硕士研究生，主要从事计算机视觉方面的研究。
　　通信作者：郑f默，E-mail：zhengliaomo@sict.ac.cn
　　A method for fabric defect detection based on improved cascade R-CNN
　　XU Shengbao1a， ZHENG Liaomo2，3， YUAN Decheng1b
　　（1a.College of Computer Science and Technology; 1b. College of Information Engineering， Shenyang
　　University of Chemical Technology， Shenyang 110142， China; 2.Shenyang Institute of Computing
　　Technology， Chinese Academy of Sciences， Shenyang 110168， China; 3.Shenyang CASNC
　　Technology Co.， Ltd.， Shenyang 110168， China）
　　Abstract： To solve the difficult detection problem due to the uneven distribution of different fabric defects， extreme aspect ratios existing in some defects， and a large number of small targets， a method for fabric defect detection based on improved cascade R-CNN was proposed. The difficult examples were mined online in R-CNN part to strengthen small target training. To address the issue of the extreme aspect ratio of fabric defects， the traditional square volume in the feature extraction network was replaced by deformable convolution v2. The scale of the bounding box was redesigned according to the characteristics of the fabric. Finally， the complete intersection over union loss was adopted as the bounding box regression loss， and a more accurate target bounding box was obtained. The experimental results indicated that the improved model was more accurate in predicting the bounding box than that before improvement， and it achieved a better effect on small target detection. The accuracy was improved by 3.57%， and the average accuracy was improved by 6.45%. Therefore， it can better meet the requirements of fabric defect detection.
　　Key words： cascade R-CNN; fabric defect; detection; deformable convolution v2; online difficult example mining; complete intersection over union loss
　　布匹缺陷z测任务的难点可能有以下几个方面：小目标问题，缺陷具有极端的宽高比，样本不均衡。在MS COCO数据集[1]中，面积小于32×32像素的物体被认为是小目标。小目标具有分辨率低，图像模糊，携带的信息少的特点，导致其特征表达能力弱，也就是在提取特征过程中，能提取到的特征非常少，不利于其检测;布匹疵点由于生产工艺的原因常常具有极端的宽高比，例如断经、断纬等，给其边界框的预测增添了难度;样本不均衡是指部分疵点拥有大量的训练样本，而另一部分疵点则只有少数的样本，让分类器学习起来很困难。

转载注明来源:https://www.xzbu.com/1/view-15427025.htm