基于深度学习的机织印花布疵点实时检测方法研究  

Research on real-time detection method of woven printed fabric defects based on deep learning

作　　者：郭同建 徐洋[1] 陈慧敏[1] 余智祺 孙以泽[1] GUO Tongjian;XU Yang;CHEN Huimin;YU Zhiqi;SUN Yize(College of Mechanical Engineering,Donghua University,Shanghai 201620,China)

机构地区：[1]东华大学机械工程学院,上海201620

出　　处：《东华大学学报（自然科学版）》2023年第1期33-38,共6页Journal of Donghua University(Natural Science)

基　　金：国家重点研发计划资助项目(2018YFB1308800)。

摘　　要：为满足纺织业内机织印花布瑕疵检测的实时性需求,基于利用回归思想进行检测的单阶段算法模型YOLOv_(3)(you only look once version 3),提出一种改进的机织印花布疵点实时检测方法。通过优化骨干网络结构,引入可变形卷积,提高印花背景下模型的瑕疵特征提取能力;设计新的损失函数,提高瑕疵分类和定位的精准度;引入几何中位数剪枝算法,去除深层网络冗余参数,进一步提高系统检测速度。试验结果表明,改进算法的模型在测试集上准确率可达92.02%,检测精度显著提高,每张图片检测平均耗时22.61 ms,满足工厂的实时性要求。In order to meet the demand for real-time defect detection of woven printed fabric within the textile industry, an improved real-time detection method for woven printed fabric defects is proposed based on YOLOv_(3), a single-stage algorithm model for detection using regression ideas. The method optimizes the structure of backbone and introduces deformable convolution to improve the defect feature extraction capability of the model in the printing context;designs a new loss function to improve the accuracy of defect classification and localization;introduces a pruning algorithm based on geometric median to remove the redundant parameters of the deep network and further improves the detection speed of the system. The experimental results show that the model with the improved algorithm can reach 92.02% accuracy on the test set. The detection accuracy is significantly improved, and the average time spent per image detection is 22.61 ms, which meets the real-time requirements of the factory.

关 键 词：机织印花布 疵点检测 深度学习 YOLOv3 轻量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TS101[自动化与计算机技术—计算机科学与技术]