面向织物疵点检测神经网络模型的研究进展  

Research Progress of Neural Network Model for Fabric Defects Detection

作　　者：刁宇涵 祝双武 赵妍 DIAO Yuhan;ZHU Shuangwu;ZHAO Yan(School of Textile Science and Engineering,Xi'an Polytechnic University,Xi'an 710048,China)

机构地区：[1]西安工程大学纺织科学与工程学院,西安710048

出　　处：《纺织科技进展》2025年第3期21-29,共9页Progress in Textile Science & Technology

基　　金：中国纺织工业联合会科技指导性项目(2019057);陕西省教育厅科研计划项目(18JS042)。

摘　　要：疵点严重影响了织物外观质量,织物疵点自动检测技术对提高检测效率、降低人工成本、提高纺织企业生产智能化水平都具有重要的意义;因基于深度学习的神经网络具有强大的特征提取能力,近些年越来越多的研究人员将其用于织物疵点自动检测过程中,提出了很多用于织物疵点检测的神经网络模型。为了提高织物疵点的检测性能和效率,对基于CNN(Convolutional Neural Networks)、生成模型和DETR(Detection Transformer)等当前主流网络模型的检测原理进行概述;分析以这几种网络为主干的多个神经网络模型,讨论其优缺点以及目前它们在织物疵点检测上的应用状况和面临的挑战;展望DETR相关算法的研究趋势。Fabric defects significantly compromise the aesthetic quality of textiles.Automatic defect detection technology has been recognized as crucial for enhancing inspection efficiency,reducing labor costs,and advancing intelligent production capabilities in tex-tile enterprises.Deep learning neural networks have been increasingly employed in automated defect detection due to their robust fea-ture extraction capacity,with numerous specialized models developed in recent years.To optimize detection performance and opera-tional efficiency,the operational principles of mainstream network architectures including convolutional neural networks(CNNs),generative models,and detection transformers(DETR)were systematically examined.Multiple neural network frameworks derived from these architectures were critically analyzed,with their respective strengths and limitations comparatively evaluated.Current im-plementation status and persistent challenges in industrial deployment were discussed.Potential research directions for DETR-based algorithms in this domain were identified.

关 键 词：深度学习 织物疵点检测 卷积神经网络(CNN) 生成模型 DETR 

分 类 号：TS107[轻工技术与工程—纺织工程]