文献类型：期刊文献

中文题名：基于改进Pix2PixGAN的织物疵点检测算法

英文题名：Fabric defect detection algorithm based on improved Pix2PixGAN

作者：郜仲元[1,2];余灵婕[1,2];章玉铭[3];支超[1,2];陈梦琦[1,2]

机构：[1]西安工程大学纺织科学与工程学院,陕西西安710048;[2]西安工程大学省部共建智能纺织材料与制品国家重点实验室(培育),陕西西安710048;[3]绍兴文理学院元培学院纺织服装与艺术设计分院,浙江绍兴312000

年份：2022

卷号：50

期号：3

起止页码：9

中文期刊名：毛纺科技

外文期刊名：Wool Textile Journal

收录：Scopus、北大核心、北大核心2020

基金：国家自然科学基金项目(51903199);陕西省自然科学基础研究计划资助项目(2019JQ-182);陕西省教育厅科研计划资助项目(18JS039);2020年陕西高校“青年杰出人才支持计划”;西安工程大学研究生创新基金项目(chx2021004)。

语种：中文

中文关键词：疵点检测;生成对抗网络;Pix2PixGAN;数据增强

外文关键词：defect detection;generative adversarial network;Pix2PixGAN;data enhancement

中文摘要：针对现阶段深度学习应用于纺织疵点检测时,由于疵点织物数据量不足而导致检测模型准确率低、疵点识别种类少的问题,提出一种基于改进Pix2PixGAN网络的训练数据增强方法。无疵点织物图像比较容易获取,利用多层深度Pix2PixGAN网络,可在无疵点织物图像上自动生成疵点,从而实现疵点图像数据的增强。首先对数据集进行预处理,得到语义分割图;然后加深U-net网络,利用双重Pix2PixGAN网络加强疵点与纹理的融合;最后将新生成的疵点图像数据加入原训练集完成数据增强。分别以数据增强前后的疵点织物样本作为训练集,采用Faster R-CNN目标检测模型进行对比实验。实验结果表明,数据增强方法可有效提高织物疵点检测的效果。对于线状、破洞和污渍3种疵点,与原训练数据集相比,数据增强后的检测模型平均精度分别从73%、75%、62%提升到84%、79%、65%。

外文摘要：When deep learning is applied to intelligent textile defect detection,the insufficient training data may result in low accuracy and poor adaptability of varying defect types of the trained defect model.To address the above problem,a training data enhancement method based on improved Pix2PixGAN network was proposed.Compared with the defected fabric images,the defective-free fabric images were much easier to obtain.According to the above reasons,multi-layer deep Pix2PixGAN network was established to generate defects automatically on the defect-free fabric images,thus enhancing the training data of fabric defect images.Firstly,the image preprocessing was applied to the defect-free fabric image to obtain semantic segmentation graph.Then,a cascade neural network based on double U-net structure was designed based on Pix2PixGAN network to strengthen the fusion of defect and texture.Finally,the semantic segmentation images were inputted to the improved Pix2PixGAN network to generate new fabric defect images,merging the new image with the original defect data to realize the data enhancement.Comparison experiments were performed using the Faster-RCNN detection model on the training data before and after data enhancement,respectively.The experimental results show that the data enhancement method can effectively improve the accuracy of fabric defect detection.With regard to the defects of linear defects,holes and stains,the Average Precision of the detection model are increased from 73%,75%and 62%to 84%,79%and 65%.