基于约束条件的非负矩阵分解算法及其在纤维自动识别中的应用研究

Non-negative matrix factorization (NMF) has been widely employed in computer vision and pattern recognition fields since the learned bases can be interpreted as a natural parts-based representation of the input space, which is consistent with the psychological intuition of combining parts to form a whole. It has been shown that learning performance could be significantly enhanced if the geometrical structure was exploited and the local invariance was considered. In addition, the labels are incorporated into the dictionary learning stage and obtained a discriminative dictionary. It gives us a possible that we can incorporate into the NMF model both intrinsic geometrical structure and discriminative information which have been essentially ignored in prior works. Specifically, both the graph Laplacian and supervised label information are jointly utilized to learn the projection matrix in the new model. As one of two significant steps in object recognition, Feature Coding is more important than Dictionary Learning. A dictionary can be produced by the k-means clustering algorithm, and in the coding stage, each descriptor was represented by a nearest basis vector, in which there is only one nonzero entry in each coefficient vector (i.e. sparse matrix). And then each descriptor can be encoded by applying soft inner product coding scheme. We propose to combine the sparse representation of object (fiber) with spatial information of image, which are also important, however, existing NMF methods do not take these two aspects into consideration simultaneously. The spatial information of fiber (the local structure of it) and sparse representation of image can be used together in the spatial pyramid matching (SPM) model for fiber recognition, which has not been applied in the fiber recognition to our knowledge. As far as we know, different types of fibers, such as cotton, flax, and synthetic fibers, are often blended together in textile manufacturing to improve performances of end-use products. The proportion of each component plays an important role in determining the ultimate quality on the end product, and thus needs to be monitored quantitatively for quality inspection. An automatic, fast, accurate fiber component analysis method instead of traditional way is not only the urgent needed by China's inspection and quarantine departments as well as enterprises, but also is one of the worldwide technical difficulties in the commodity inspection field. We propose a research project that applies the constrained NMF technology to automatic identifications of blended fibers. The accuracy, objectivity and testing efficiency will be improved by our inspection work, and a greater social and economic benefit can be brought to our country's textile exports, as well as the monitoring unit to supervise the importation of new fiber, textiles and clothing.

图像表示是模式识别研究中关键问题之一。非负矩阵分解是一种有效的图像表示方法，近年来已被广泛应用到计算机视觉、信号处理、模式识别和图像处理等领域。但是，非负矩阵分解在处理高维数据时的效率瓶颈以及无法同时考虑样本类别信息和固有几何结构信息的缺陷制约了非负矩阵分解的应用范围和应用研究的发展。本项目针对非负矩阵分解存在的局限性，通过系统地研究基于约束条件的新特性非负矩阵分解模型在维数约减、特征提取和分类识别等关键技术中的应用，以解决非负矩阵分解存在的数据冗余性较大、无监督训练过程不利于后续识别率提高等缺陷，拓展和推动非负矩阵分解的应用，具有重要的理论研究意义和实用价值。同时，将研究成果用于解决混纺纤维的纤维自动分类问题，对纺织品截面纤维进行准确的图像表示，并开发完成基于显微图像的纤维混纺织品自动检测系统。其研究成果将为解决纺织品检验领域纤维自动识别与分析这一世界性难题带来创新性的突破。

不同的纺织品纤维横截面在显微镜下形状各异，呈现出一定的几何形状，不规则的截面形状使纤维具有不同的特征，进而导致了纤维产品有不同的特性。因此，纤维的纺比是纤维产品质量的重要检测指标。..图像表示是模式识别中研究的热点问题之一。其中非负矩阵分解（NMF）的是种非常有效的图像表示方法。..目前非负矩阵分解降维需借助计算量较大的迭代操作来实现，因此，其处理高维数据时常遇到效率瓶颈。另外，NMF无法同时考虑样本类别信息和固有几何结构信息，这些缺陷都制约了非负矩阵分解的应用范围和应用研究的发展。..针对传统NMF模型的不足，本项目通过系统研究基于约束条件的新特性非负矩阵分解模型在维数约减、特征提取和分类识别等关键技术中的应用，以解决非负矩阵分解存在的数据冗余性较大、无监督训练过程不利于后续识别率提高等缺陷，拓展和推动非负矩阵分解的应用，并在理论研究的基础上研究并开发了混纺织品纤维自动识别原型系统。..1、.提出了一种新的成对约束图正则非负局部坐标分解(PCGNLCF) 方法，其中成对约束的非负局部坐标分解(PCGNLCF)用于图像表示。实验结果优于现有最好算法。..2、.基于Self-Organizing Map算法提出了一种在线自组织散列方法SOH。该方法采用类SOM的算法，将训练数据映射到超立方体的每个顶点，并赋以二进制编码，以保持仿射不变性。实验结果显示：SOH方法相比于当前各种先进的散列方法具有更好的检索性能，同时具有更低的计算复杂度和存储空间。..3、.结合SIFT特征的非负性，提出了一种基于空间金字塔匹配的稀疏性约束非负矩阵分解算法。将这一研究结果应用于混纺纤维图像做了实验比较。实验结果表明，该算法的识别精度高于PCA、LDA和NMF等几种经典算法。..4、.我们将稀疏且正交的约束加于自编码器应用于纤维目标的特征提取 ，使得它成为有高度判别力的特征描述子。实验结果显示：我们提出的方法相比于当前各种特征描述方法具有更好的精确性和效率。..这一项目的研究，是对于混纺织品纤维自动识别方面的研究的突破，也能够对显微图像中目标特征提取与识别有着重要的借鉴作用。