基于多尺度稀疏表示的印花织物瑕疵识别方法研究

In recent years, it has achieved a great progress in raw fabric defect detection and evaluation. However, how to efficiently extract the defect features and how to develop feature analysis and recognition methods which is consistent with the subjective perception for printed fabrics with complex texture, variety of patterns, and variable colors, are still the challenging problems in textile appearance quality inspection. To solve the aforementioned problems well, the proposal focuses on the following three contents. (1)Through the color transformation model, the printed fabric feature method based on multi-scale linear filter is proposed to solve the problem of defect features extraction and defect-related singularity structures analysis. (2)The statistical machine learning and optimization theory is used to explore the low-rank sparse representation of the printed fabrics feature to solve the problem of the defect features expression, and reduce computational cost and boost up computational efficiency. (3) To realize the rapid identification of printed fabric defects, we present the conditional Gaussian mixture model construction scheme which based on adaptive method. This scheme assist worker to analyze the causes of printed fabric defects and will provide a reliable basis for the evaluation of textile appearance quality. The proposal will provide a new theoretical basis for the evaluation of the appearance quality of printed fabrics. The research results have important theoretical significance and engineering application value.

近年来，针对坯布的外观质量检测与评价取得了较大发展。然而，对于纹理结构复杂、花型繁多，颜色多变的印花织物，如何对其瑕疵特征进行有效提取，建立与人类视觉感知相一致的特征分析与识别方法，仍然是纺织品外观质量检测所面临的挑战性问题。为有效解决上述问题，本课题拟开展三个方面的研究：(1)通过颜色模型转化，研究基于多尺度线性滤波器的印花织物瑕疵特征提取方法，解决印花织物瑕疵特征提取及相关奇异结构分析问题；(2)利用统计机器学习和最优化理论，探索印花织物瑕疵特征的低秩稀疏表示，完成瑕疵特征的有效表征，提高瑕疵识别效率；(3)探讨基于自适应的条件高斯混合模型的构建方案，实现印花织物瑕疵的快速识别，分析印花织物瑕疵产生原因，为纺织品外观质量评价提供可靠依据。本课题的研究对印花织物外观质量评价提供新的理论依据，研究成果具有重要的理论意义和工程应用价值。

印花织物在服装制衣、高端家纺制品及家居装饰品中的需求量较大，对其进行瑕疵检测至关重要。本课题基于人眼对于物体感知的多尺度结构特性，深度剖析瑕疵信息与其正常印花织物的组织纹理及颜色的机理和规律，分析印花织物瑕疵产生原因，围绕印花织物颜色模型转换、多尺度瑕疵特征提取及瑕疵特征识别展开研究。提出一系列代表性的方法如：提出了一种基于CS-WNN的CMYK到CIELab的印花织物色彩空间转换方法；提出了一种基于PSO-DBN印花织物色彩空间转换方法；提出了一种基于CenterNet 的印花织物检测方法；提出了一种基于CARL-YOLOF数码印花织物缺陷检测方法；提出了一种基于Criminisi算法非周期印花织物瑕疵检测方法；提出了一种基于MF-SSD的织物疵点检测方法；提出了一种基于改进的快速加权中值滤波与K-means聚类相结合的纹理织物疵点检测方法；提出了一种Mobile-Unet卷积神经网络的织物疵点检测方法，实现端到端的织物缺陷分割。本课题的研究对印花织物外观质量评价提供新的理论依据，有助于提升印花织物产品质量，为我国纺织品外观质量控制提供有效的技术支撑。. 本项目的研究成果包括：培养硕士研究生8名，在读研究生12名。发表与课题相关并标注本基金资助的学术论文共计14篇，其中SCI期刊论文7篇（均已检索），EI期刊论文2篇（均已检索），EI会议论文1篇（均已检索）。获批发明专利4项，实用新型专利1项；登记软件著作权2件。2022年获香港桑麻基金会奖教金；2021年获《纺织学报》优秀论文奖1项。.