鲁棒表征下的织物瑕疵异常检测算法研究

Automated fabric defect inspection is one of important research subjects for textile manufacturing in information intelligence upgrading. At present, the performance of related work mainly relies on extracting discriminative textural features, hardly guaranteeing their optimisation in identifying defects, while the traditional time-frequency methods e.g. Gabor filter also confronts the challenge of parameter optimization. Thus, this project starts with the dependence between defect and normal fabric texture, identifying defects based on normal texture being well modelled. Essentially, representing fabric texture with a learned dictionary is an optimisation problem, and how to find the learned dictionary that is able to robustly model normal texture and well discriminate defects is our primary work; then the defect enhancement and discrimination mechanism involving visual saliency are to be studied; then studying the novelty detection algorithm and segmentation algorithm and devising detection algorithms base on them; finally, conducting static and dynamic experiments to evaluate the performance of devised algorithms. This research addresses that well representing normal fabric texture is the key necessary for the effective defect detection with the mathematical optimization and visual attention mechanics considered, managing to recognise defects in a more nature way as our visual perception does.

机织物表面瑕疵的自动检测研究是提升我国纺织加工信息智能化的重要课题之一，当前瑕疵异常检测算法多依赖于鉴别性特征的提取，很难保证最优性，而传统的时频分析方法如Gabor滤波器又面临着参数优化问题。本项目以瑕疵和正常纹理的依存关系为出发点，在有效表征织物纹理基础上进行瑕疵鉴别。所采用的学习字典表征方法其实质为一个最优化问题，如何构建能在鲁棒描述正常纹理下对瑕疵有最优的鉴别性的字典学习模型是本项目首要研究内容；然后在此基础上对结合视觉显著性的瑕疵的增强和区分机理进行研究；接着对基于单分类算法和分割算法进行研究并设计相应检测算法；最后对所设计的检测算法在静态和动态环境测试和实验分析。本研究强调对正常织物纹理的表征是有效瑕疵检测的关键，注重数学优化和视觉注意机制的引入，力求让瑕疵的辨别更符合人的视觉感知。

纺织品质量控制及智能检测是提升我国纺织加工信息智能化的重要课题之一，其中织物瑕疵检测算法的研究是该课题的核心。织物瑕疵检测其实质为一个单分类或异常检测问题，而当前瑕疵检测算法多依赖于鉴别性特征的提取及参数优化。本项目另辟蹊径以瑕疵和正常纹理的依存关系为出发点，在有效表征织物纹理基础上进行瑕疵异常鉴别。首先对字典系数施加正则化约束下的织物纹理表征模型及其鲁棒性及鉴别性进行了研究；其次在表征基础上对正常纹理与瑕疵间的区分机理进行了研究；最后对基于FCM单分类器、疵点分割算法进行研究并设计相应瑕疵检测算法。主要取得的理论成果有：（1）构建了施加正则化约束下的字典表征模型，有效地提高了表征织物纹理的鲁棒性；（2）提出一维投影下局部纹理增强方法及鉴别性特征；（3）在FCM聚类分析基础上，构建了适应于织物瑕疵异常检测器。. 本项目最大的创新点在于以对机织物纹理的表征为入口点，很好地绕开特征提取方法所面临的特征选取问题。对织物表征模型的优化设计为建立统一的以表征为思路的机织物瑕疵检测算法框架提供基础，同时也为自动验布机算法设计、机织物的参数识别及织物仿真等应用有重要借鉴作用。