极限学习机在纺织品图像处理中的应用

发布时间：2018-02-26 15:44

本文关键词： 色差分类疵点检测极限学习机差分进化算法多尺度字典学习　出处：《浙江理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：在现代纺织织造行业,纺织品成品的质量直接关系到其销量,只有提高纺织品的质量才能提高行业竞争力。通常来说,影响纺织品质量的两个关键因素是色差和疵点。在传统的纺织品织造行业,色差和疵点的检测大多都是通过经验丰富的技术人员的肉眼来完成。但是由于现场环境的恶劣和技术人员的视觉疲劳,检测效率往往非常低下,强烈依赖于技术人员的主观意识。因此基于机器视觉的纺织品图像检测应运而生,不但解决了传统的检测方法依赖人工完成,而且大大提高了检测的正确率。本文的研究工作主要围绕基于机器学习的纺织品色差检测和疵点检测两个方面,致力于构建具有良好稳定性的色差分类模型和疵点检测模型,并用于解决纺织品在印染和织造过程中的色差和疵点问题。论文的主要工作和研究成果概括如下:(1)对纺织品的色差,疵点的基本概念和研究现状进行了简要的介绍,分析和比较了各种色差分类和疵点检测算法的特点及优缺点。针对纺织品的色差分类,着重研究了具有良好分类性能的极限学习机算法,并且对极限学习机的理论基础和存在的问题进行了分析和探讨。针对纺织品的疵点检测,重点研究了基于字典学习的疵点检测算法,并且结合稀疏编码理论对字典学习方法的工作原理和隐含的问题进行了深入的分析和讨论,为本文后续的工作提供坚实的理论基础。(2)为了建立纺织品色差分类模型,本文提出了一种基于动态参数选择的差分进化算法优化正则化极限学习机方法。首先,针对传统的极限学习机随机生成输入权重和隐层偏置的问题,利用差分进化算法良好的全局搜索能力迭代优化求取极限学习机的输入权重和隐层偏置;同时,由于传统极限学习机在计算输出权重时只考虑到经验风险,故引入了代表结构风险的正则化参数,防止生成病态解矩阵;最后,采用动态参数选择方法选取差分进化算法最优的参数组合模型,构建最优化的色差分类模型。实验结果表明,相比于差分进化算法优化原始极限学习机的色差分类方法,本文提出方法取得较高的分类精度,并且具有较好的稳定性和泛化能力。(3)纺织品是由经纬纱线按一定的规则编织而成,故其表面呈现高相关性,疵点区域可视为局部异常纹理。考虑到织物表面的纹理结构和疵点的尺寸,本文提出一种基于多尺度字典学习的疵点检测方法,所提出的多尺度字典学习方法能够更加清晰的描述纺织品图像细节。由于原始字典学习算法计算量大,故在学习字典阶段,引入改进的字典学习算法KSVD,加快字典更新速度。并提出了自适应差分进化算法优化正则化极限学习机的纺织品疵点检测模型。在模型训练阶段,引入自适应变异算子以解决差分进化算法寻优过程中的参数设置问题。实验结果表明,与传统的应用Gabor滤波器方法,形态学操作方法和局部二值模式相比,本文提出的方法能精确的定位疵点区域,实现较高的疵点检出率。并且,本文提出的方法针对纯红色布和纯粉色布的疵点也能实现非常好的检测。
[Abstract]:In modern textile weaving industry, textile product quality is directly related to its sales, only to improve the quality of textiles to improve the competitiveness of the industry. Generally speaking, the two key factors affecting the quality of textiles is color and defects. In the traditional textile industry, color difference and defect detection is almost completed by technical personnel experience rich eye. But because of the bad environment of visual fatigue and technical personnel, the detection efficiency is very low, is strongly dependent on the technical personnel's subjective consciousness. Therefore textile machine vision detection based on image came into being, not only solved the traditional detection method relies on manual work, but also greatly improve the detection accuracy. This thesis is mainly focus on the two aspects of machine learning and textile defect detection based on color difference detection, is committed to building a The color classification model and the defect detection model of good stability, and is used to solve the textile printing and dyeing and weaving process in color and defects. The main work and research results are summarized as follows: (1) the color of textiles, briefly introduces the basic concepts and research status of defects, analyzes and compares the characteristics of various color classification and defect detection algorithm and its advantages and disadvantages. According to the color classification of textiles, focuses on the ultimate has good classification performance of machine learning algorithms, and discusses the theoretical basis and extreme learning machine and problems. In view of the textile defect detection, focusing on the defect detection algorithm based on dictionary learning the working principle, and combined with the theory of sparse encoding dictionary learning method and hidden problems are analyzed and discussed deeply, for the following Provide a solid theoretical basis. (2) in order to establish the textile color classification model, this paper proposes an algorithm of optimal regularization parameter selection method of dynamic extreme learning machine based on the difference. First, the traditional extreme learning machine randomly generated input weights and hidden layer bias problem, using the differential global evolutionary algorithm good search limit from learning the input weights and hidden layer offset machine for iterative optimization ability; at the same time, because of the traditional extreme learning machine in the calculation of output weights when considering only the empirical risk, so the introduction of the regularization parameter represents the structural risk, prevent the formation of ill posed matrix; finally, selection methods difference parameter combination model optimal evolution algorithm for dynamic parameters using optimization model, color classification. The experimental results show that compared to the original color difference optimization evolutionary algorithm for extreme learning machine The classification method proposed in this paper has high classification accuracy method, and has good stability and generalization ability. (3) is made of textile warp and weft yarn woven according to certain rules, so the surface showed a high correlation, the defect region can be regarded as abnormal local texture. Considering the texture and structure of fabric surface size in this paper, a method of detecting defects based on multiscale dictionary learning, multiscale dictionary learning method proposed can describe the textile image details more clearly. Because the original dictionary learning algorithm. So in the dictionary learning stage, the introduction of improved dictionary learning algorithm KSVD, accelerate the dictionary update speed. And put forward the points evolutionary algorithm regularized extreme learning machine textile defect detection model. During the training phase adaptive model, introducing the adaptive mutation operator to solve differential evolution Algorithms for parameters optimization in the process of setting problem. Experimental results show that the application of Gabor filter with the traditional method, morphological operation method and local two value model, the proposed method can achieve accurate defect location, high defect detection rate. Moreover, this paper proposed a method for pure red and pure pink cloth cloth defects can achieve very good detection.

【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【相似文献】