Novel real-time algorithms and system architecture for defect detection on patterned textiles

全球纺织业生产1700亿米的织物，创造了6400亿英镑的市场价值。25%的纺织品是图案织物。缺陷检测是一个主要的行业挑战。未能在缺陷容忍范围内提供纺织品可能导致整批召回，从而导致代价高昂的客户索赔和下游生产延迟，影响零售商的库存管理。不良的缺陷管理也是工业废物的一个主要来源。传统的缺陷检测方法依赖于人工检测，这是无效的（<65%的检出率）。虽然机器视觉系统为自动缺陷检测（95%）和管理提供了复杂的平台；这些系统仅限于普通纺织品。虽然模式匹配和神经网络方法以前已经尝试过图案纺织品；由于在可变形的基材（纺织品）上进行模式匹配的极端复杂性，以及为每种模式类型训练神经网络所需的时间，所有这些都未能提供实用的解决方案。因此，图案纺织品的制造商仅限于人工检查。该项目以市场领先的平纹纺织品视觉系统为基础，将开发新的递归模板匹配算法，用于解决复杂的图案变形，实现有效的图案减法，从而揭示潜在的缺陷。视觉系统将提供：i)相机和照明系统，以实现高速（100米/分钟）的最佳图像捕获；Ii)利用统计分析实现纺织新产品系统配置自动化的自我培训软件；Iii)先进的图像分析，用于检测95%的缺陷；Iv)能够学习和自动化客户决策规则的缺陷分类系统；V)记录和检索完整卷图图像的系统，以供后续审查和质量控制；vi)生成具有完整缺陷数据的纺织卷图，用于优化纺织品切割计划（提高产量），改善下游加工和质量保证。目前市场上还没有为图案纺织品提供这些功能的视觉系统。项目重点的主要领域在于开发新的递归模板匹配算法。这些将集成到现有的视觉系统平台中，并通过领先的纺织品和服装制造商巴宝莉的工厂试验进行验证。该项目的一个重要里程碑将是在领先的行业展览（ITMA - Barcelona）上展示原型系统。图案纺织品视觉系统的潜在市场约为10亿英镑。该合作伙伴关系的目标是在5年内实现约848万英镑的业务增长（累计销售额约1690万英镑），从而创造42个新工作岗位，并产生30倍的投资回报率。

The global textile industry produces \>170 billion metres of fabric generating a market value of \>£640 Billion. \>25% of textiles are patterned fabrics.Defect detection represents a major industry challenge. Failure to provide textiles within defect tolerance limits can lead to whole batch recalls, resulting in costly customer claims and downstream production delays impacting on retailer stock management. Poor defect management is also a major source of industrial waste.Traditional methods for defect detection rely on human inspection which is ineffective (<65% detection). Whilst machine vision systems offer sophisticated platforms for automated defect detection (\>95%) and management; these systems are restricted to plain textiles.Whilst pattern matching and neural network approaches have previously been tried for patterned textiles; all have failed to provide a practical solution due to the extreme complexities associated with pattern matching on deformable substrates (textiles) and the time required to train a neural network for each pattern type. Manufacturers of patterned textiles are therefore limited to manual inspection.Building on a market leading vision system for plain textiles, the project will develop novel recursive template matching algorithms for the resolution of complex pattern deformations, enabling efficient pattern subtraction and thereby revealing underlying defects.The vision system will offer: i) camera and lighting system for optimum image capture at high speed (\>100m/min); ii) self-training software utilising statistical analysis to automate system configuration for new textile products; iii) advanced image analysis for detection of \>95% defects; iv) defect classification system able to learn and automate client decision rules; v) system for recording and retrieval of complete roll map images for subsequent review and quality control; and vi) generation of textile roll maps with complete defect data, for optimised textile cut plan (increased yield), improved downstream processing, and quality assurance. No vision system on the market offers these features for patterned textiles.The main areas of project focus lie in development of the novel recursive template matching algorithms. These will be integrated into the existing vision system platform and validated through factory trials with leading textile and clothing manufacturer Burberry. An important milestone in the project will be demonstration of a prototype system at the leading sector exhibition (ITMA - Barcelona).The potential addressable market for patterned textile vision systems is ~£1 billion. The partnership targets ~£8.48 million business growth within a 5 year period (~£16.9 million cumulative sales) thereby creating \>42 new jobs and generating a \>30-fold ROI.