基于低秩稀疏矩阵分解的织物疵点检测算法研究

发布时间：2018-05-07 08:57

本文选题：织物图像 + 疵点检测　；参考：《中原工学院》2017年硕士论文

【摘要】：织物疵点检测是纺织品质量控制系统中一个核心环节,直接影响着系统的性能。从纹理复杂的织物图像中检测形态多样的疵点具有重要的应用价值。该问题的解决也有利于对其它工业产品表面缺陷检测提供新的解决思路。现有织物疵点检测多采用传统模式识别的方法,如统计分析、频谱分析等。近年来,受压缩感知和稀疏表示理论的推动,低秩稀疏矩阵分解模型在图像处理和模式识别中也获得广泛的应用,并且在目标检测中达到很好地检测效果。低秩稀疏矩阵分解模型与人类视觉系统的低秩稀疏性相吻合,通过将图像矩阵分解为低秩阵和稀疏阵,实现目标与背景的有效分离,特别地,对于织物图像,视觉上具有高度冗余性,相对于自然图像中的目标检测,织物疵点检测能够更好地符合了低秩稀疏矩阵分解模型。另外,织物图像的特征提取也是疵点检测的关键步骤。对图像提取好的特征,并构建合适的低秩稀疏矩阵分解模型,并利用有效的优化求解方法和对分解得到的疵点分布图采用有效的阈值分割算法,才能准确和有效的定位出疵点的位置和区域。为此,本论文提出应用方向梯度直方图和低秩分解、基于Gabor滤波器和低秩分解、基于GHOG和低秩矩阵恢复以及基于生物视觉特征提取及低秩表示的织物疵点检测算法。所做的工作以及研究成果如下:1)提出基于Gabor滤波器和低秩分解的织物疵点检测算法。首先,对织物图像提取Gabor滤波器特征,再对生成的特征图进行均匀分块,并将所有的图像块特征组合成特征矩阵。对于生成的特征矩阵,构建合适的低秩分解模型,通过快速近端梯度方法优化求解,从而生成低秩矩阵和稀疏矩阵,最后采用最优阈值分割算法对由稀疏阵生成的疵点分布图进行分割,从而定位出疵点的区域和位置。2)提出了应用方向梯度直方图(HOG)和低秩分解的织物疵点检测算法。首先,将织物图像划分为大小相同的图像块,提取每个图像块的HOG特征,并将图像块特征组成特征矩阵。针对特征矩阵,构建有效的低秩分解模型,通过增广拉格朗日方法优化求解,生成低秩阵和稀疏阵;最后采用最优阈值分割算法对由稀疏阵生成的疵点分布图进行分割,从而定位出疵点区域。3)提出了基于GHOG及低秩分解的模式织物疵点检测算法。对于前两种检测算法只能检测纹理较为简单的织物疵点图像,本论文提出了一种基于GHOG和低秩恢复的模式织物疵点检测算法。首先,对图像进行Gabor滤波,从而生成相应的Gabor特征图,然后将对应的方向上的Gabor特征图进行均匀分块,并提取HOG特征,从而生成最后的GHOG特征,并将所有图像块的特征向量进行级联生成特征矩阵。对特征矩阵,构建低秩分解模型,并利用方向交替方法进行优化求解,产生低秩矩阵和稀疏矩阵,并对由稀疏矩阵产生的疵点分布图采用最优阈值分割算法进行分割,从而定位出疵点的位置。4)提出了基于生物视觉特征提取及低秩表示的织物疵点检测算法。生物视觉对客观世界的表征是完备的,能支持各种复杂的高级视觉任务。本文引入一种借鉴人类视觉感知和视网膜表征机理的特征表示方法。在该特征表示的基础上,利用KSVD在测试图像上训练出正常织物图像块字典。基于学习出的字典,建立特征矩阵的低秩表示模型,并利用ADMM方法进行求解,从而提高算法检测效果及自适应性。
[Abstract]:Fabric defect detection is a key link in the quality control system of textiles, which directly affects the performance of the system. It is of great application value to detect the variety of defects from the texture of a complex texture. The solution of this problem is also helpful to provide a new solution for the surface defect detection of other industrial products. Point detection mostly uses traditional pattern recognition methods, such as statistical analysis, spectrum analysis and so on. In recent years, the low rank sparse matrix decomposition model has also been widely used in image processing and pattern recognition, and has been widely used in image processing and pattern recognition. The model is consistent with the low rank sparsity of human visual system. By decomposing the image matrix into low rank array and sparse array, the target and the background are separated effectively. In particular, the fabric image is highly redundant. The fabric defect detection can better meet the low rank sparsity compared with the target detection in the natural image. In addition, the feature extraction of the image of the fabric is also the key step of the defect detection. The features extracted from the image and the suitable low rank sparse matrix decomposition model are constructed, and the effective optimization method and the effective threshold segmentation algorithm are used to determine the defect distribution. In this paper, we propose the application of directional gradient histogram and low rank decomposition, based on Gabor filter and low rank decomposition, GHOG and low rank matrix restoration, and fabric defect detection algorithm based on biological visual feature extraction and low rank representation. The work and research results are as follows: 1) proposed based on Gab Or filter and low rank decomposition algorithm for fabric defects detection. First, the feature of the fabric image is extracted from the Gabor filter, and then the generated feature graph is partitioned evenly, and all the image block features are combined into the feature matrix. In the end, the low rank matrix and the sparse matrix are generated. Finally, the optimal threshold segmentation algorithm is used to segment the defect distribution map generated by the sparse array, and the defect location and location.2 are located. The fabric defect detection algorithm is proposed by using the direction gradient histogram (HOG) and the low rank decomposition. First, the fabric image is divided into the size phase. In the same image block, the HOG features of each image block are extracted and the feature matrix of the image block is formed. An effective low rank decomposition model is constructed for the feature matrix. The low rank array and sparse array are generated by the augmented Lagrange method to generate the low rank array and the sparse array. Finally, the optimal threshold segmentation algorithm is used to divide the defect distribution map generated by the sparse array. The defect detection algorithm based on GHOG and low rank decomposition is proposed. The first two detection algorithms can only detect the fabric defect image with simple texture. In this paper, a pattern detection algorithm based on GHOG and low rank recovery is proposed in this paper. First, the image is filtered by Gabor, The corresponding Gabor feature graph is generated, then the Gabor feature map of the corresponding direction is partitioned evenly, and the features of the HOG are extracted, thus the final GHOG features are generated, and the feature vectors of all the image blocks are cascaded to generate the feature matrix. The low rank decomposition model is constructed for the feature matrix, and the direction alternation method is used to optimize the feature matrix. A low rank matrix and a sparse matrix are generated, and the defect distribution map produced by the sparse matrix is segmented with the optimal threshold segmentation algorithm, and the location of the defect location.4 is located. A fabric defect detection algorithm based on the feature extraction of biological vision and low rank representation is proposed. The representation of the raw object vision to the objective world is complete and can be supported. In this paper, we introduce a feature representation method that draws on the mechanism of human visual perception and retina representation. On the basis of this feature, we use KSVD to train normal fabric image block dictionary on the test image. Based on the learning dictionary, the low rank representation model of the feature matrix is built, and ADMM is used. Method is used to improve the detection effect and adaptability of the algorithm.

【学位授予单位】：中原工学院
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TS101.97;TP391.41

【参考文献】