基于改进级联卷积神经网络的织物疵点检测  

Fabric Defect Detection Based on Improved Cascade R-CNN

作　　者：李小庆 张俊杰 杜小勤[1,2,3] 梁晶 袁桦 LI Xiaoqing;ZHANG Junjie;DU Xiaoqin;LIANG Jing;YUAN Hua(Hubei Provincial Engineering Research Center for Intelligent Textile and Fashion,Wuhan 430200;Engineering Research Center of Hubei Province for Clothing Information,Wuhan 430200;School of Computer Science and Artificial Intelligence,Wuhan Textile University,Wuhan 430200;Hubei Garment Art&Culture Research Center,Wuhan 430073;Wuhan Textile and Apparel Digital Engineering Technology Research Center,Wuhan 430073)

机构地区：[1]纺织服装智能化湖北省工程研究中心,武汉430200 [2]湖北省服装信息化工程技术研究中心,武汉430200 [3]武汉纺织大学计算机与人工智能学院,武汉430200 [4]湖北省服饰艺术与文化研究中心,武汉430073 [5]武汉纺织服装数字化工程技术研究中心,武汉430073

出　　处：《计算机与数字工程》2024年第5期1557-1562,1568,共7页Computer & Digital Engineering

基　　金：湖北省普通高校人文社会科学重点研究基地项目(编号:2021HFG007)资助。

摘　　要：为了改进当前织物检测算法样本数量少、织物疵点检测准确率低和定位精准度差的问题,提出一种端到端的改进的织物疵点检测算法。针对公开数据集样本数量少、样本种类不均衡的问题,采用线下与线上结合的数据增广方式,除了基本的数据增广方法,同时引入复制粘贴以及混合的方式对样本进行扩充与增强;针对特征提取算法提取特征不精确的问题,对特征金字塔进行改进,通过加入可变形卷积、递归特征金字塔、可切换的空洞卷积、全局语义信息的方法扩大感受野、增强语义信息。实验结果验证了算法的有效性,该算法对天池雪浪制造数据集9种布匹疵点进行检测,检测是否具有瑕疵的准确率达到97%以上,疵点定位的平均检测精度为56.7%,样本检测效率为2.4 FPS。相对于基础模型定位精准度提升了10%以上,并且检测效果满足工业上的生产需求。The work aims to propose an end-to-end improved algorithm for fabric defect detection in order to solve the prob-lems in the current cloth detection algorithm including few samples,low defect detection accuracy and poor positioning accuracy.Aiming at the problem of lacking samples and imbalance of classes in public data sets,offline and online data augmentation meth-ods are adopted.In addition to basic data augmentation methods,copy-paste and mixup are also introduced to expand and grow sam-ples.Aiming at the poor accuracy features extracted by the feature extraction algorithm,the feature pyramid network is improved by adding deformable convolution,recursive feature pyramid,switchable atrous convolution,global context to enlarge the receptive field and enhance semantic information.The experimental results verify the effectiveness of the algorithm.This algorithm can defect 9 kinds of cloth defects,the accuracy of detecting whether the fabric is defective is above 97%,the average detection accuracy of de-fect location is 56.7%and the efficiency of sample detecting is 2.4 FPS on TIANCHI-XUELANGAI dataset.Compared with the ba-sic model,the positioning accuracy has been improved by more than 10%and the algorithm meeting industrial production needs.

关 键 词：织物疵点检测 级联卷积神经网络 数据增广 递归特征金字塔 可切换空洞卷积 

分 类 号：TS101.9[轻工技术与工程—纺织工程] TP391.4[轻工技术与工程—纺织科学与工程]