基于字典学习对机织物纹理表征的瑕疵检测算法研究

The study of fabric defect detection algorithm based on computer vision is one of important research subjects towards transforming and updating traditional textile industry. At present, the related work mainly focuses on how to extract discriminative feature vectors, and those methods generally confront feature selecting problem as lack of negative samples for training. On the other hand, fabric defect detection is always taken as a multi-class classification problem in terms of classifier design, leading to a poor flexibility in adapting various fabric textures and defect types. Therefore, this project will take the efficient representation of normal fabric texture as the starting point to address fabric defect detection problem. The major research contents are composed of three parts: using learned dictionary to construct the representation model for defect-free fabric texture, and making a profound study upon the representation efficiency with the learned dictionary in order to achieve expected results; studying the mechanism of identifying defects from normal fabric by analysing the representation performance of defective texture with learned dictionary; devising detection algorithms on the basis of well representing normal fabric texture and verifying their effectiveness as well. Finally, this research conducts defect detection operation via a more fundamental way of texture representation by taking advantage of fabric texture features and flexibility of dictionary learning, which could somehow alleviate the feature selection problem and could potentially provide a good reference for other textile applications.

基于计算机视觉的机织物瑕疵检测算法研究是提升传统纺织产业重要的课题之一。当前，相关检测算法的研究工作主要集中在鉴别性特征参数的提取上，由于缺少负样本特征参数的选取往往面临特征选择问题；另一方面，在分类器的设计上研究者又多以多类问题进行处理，导致检测算法适应性不佳。为此，本项目拟以对正常织物纹理进行有效表征为出发点，首先采用字典学习对正常织物纹理进行建模表征，并对学习字典的表征性能进行深入研究，以寻求对织物纹理的最优表征；然后探讨学习字典对瑕疵纹理的表征特性，研究正常纹理与瑕疵纹理区分机理；最后在此基础上进行瑕疵检测算法设计及实验验证。本项目充分结合织物纹理特点及学习字典表征的灵活性，从更为基础的纹理表征层面上进行瑕疵检测，不仅可以巧妙地绕开特征选取问题，同时给其它相关纺织品表观质量表征检测提供借鉴。

基于机器视觉的机织物瑕疵检测算法研究是纺织智能检测的重要课题之一，由于纱线的粘弹和柔软属性强化了正常织物纹理的随机性，加上织物瑕疵尺度和形态变化多样，从而导致检测效果不佳。本项目通过对机织物纹理采用字典学习方法进行表征后再识别瑕疵，即通过建模、最优表征、瑕疵检测机理和检测算法设计与验证四个部分来研究机织物瑕疵检测算法。首先采用字典学习对正常织物纹理进行建模表征，结果表明采用字典学习的方法可以较好表征机织物的纹理，字典个数在10个以内就能较好表征，通过非负约束的求解方法比无约束求解方法的效果略好，但是得到的表达式不唯一；采用DCT过完备字典的重构效果略次于K-SVD学习字典，但是可以得到稳定的表达式。其次研究了机织物密度和织物组织结构对学习字典的表征性能的影响，结果表明，随着机织物密度的增加，重构图像与原图像的相似性逐渐降低；织物组织循环越小，浮长线越短，表面的纹理特征越明显，规则性越好，则重构的误差越小。然后主要设计了两种基于字典学习的机织物瑕疵算法，无监督和半监督织物瑕疵检测算法，通过对字典学习对正常纹理有效表征以及对异常纹理有明显差异而识别瑕疵，其中半监督检测算法对20组样品的检测效果达到平均检出率95.5%，平均误检率为1.9%的水平，明显好于采用傅里叶的方法。最后设计了机织物瑕疵在线检测仪器，通过DSP+FPGA的方式进行实时检测，在30米/分的速度下得到平均检出率为89%，误检率为3.9%的效果。本项目充分结合织物纹理特点及学习字典表征的灵活性与有效性，从更为基础的纹理表征层面上进行有效表征，得到正常纹理表征形式，进而对瑕疵检测，不仅可以巧妙地绕开特征选取问题，同时给其它相关纺织品表观质量表征检测提供借鉴。