Gabor小波和FT方法应用于疵点检测的若干理论问题研究

发布时间：2018-03-26 14:32

本文选题：疵点检测　切入点：Gabor滤波簇　出处：《武汉纺织大学》2017年硕士论文

【摘要】：提高疵点辨识精度和效率对提升纺织品质量具有重要意义。针对疵点图像光照不均和对比度低的问题,开展基于Gabor小波簇的疵点图像增强方法研究。首先,利用由3个尺度和5个方向的15个Gabor滤波簇对疵点图片进行不同方向和尺度的滤波,减少图像不均和对比度低对特征提取精度的影响;然后,将滤波图像划分成面积相等互不重合的邻域,并从邻域中提取高维特征向量。接下来,针对Gabor特征向量维数高和冗余信息大的问题,使用等距映射方法对Gabor特征进行非线性降维,剔除高维特征中冗余信息,强化分类器拟合能力,达到强化Gabor特征灵敏度的目的。其次,针对等距映射算法在Gabor特征降维过程中遇到的结构参数选择困难的问题,应用Ncut准则作为适度函数建立结构参数优化模型,使用离散离子群算法进行参数优化,提出基于粒子群和Ncut准则的等距映射参数优化方法;针对等距映射算法新增样本低维特征提取困难的问题,利用样本在高维空间和低维空间几何结构相同的假设建立新样本低维嵌入模型,提出新增样本低维特征提取方法。最后,将低维特征输入概率神经网络分类器中进行疵点辨识,突破疵点图像光照不均和对比度低等对疵点检测精确的制约。实验研究中,利用2组不同纹理的疵点图片数据进行实验研究,结果表明:基于Gabor滤波簇和等距映射算法的疵点检测准确率达97%左右。但是,同时也存在滤波器数量多、运算量大的问题。为提高疵点检测效率,利用频域协调算法抗噪能力强和计算量小优点,替代Gabor滤波器簇用于疵点图像增强,达到提高检测效率的目的。针对频域协调算法在疵点检测中遇到的疵点辨识精度受高斯滤波器模板尺寸影响大的问题,利用Ncut准则作为适度函数,建立高斯滤波器模板尺寸优化模型,使用离散离子群算法进行参数优化;针对Lab颜色空间对单一颜色纺织品疵点显著效果不明显的问题,利用HSV颜色空间代替Lab颜色空间,强化显著效果;针对色调特征、饱和度特征和亮度特征取值范围不同且变化不一致导致显著值不能很好地体现各个分量作用的问题,展开了色调特征、饱和度特征和亮度特征的归一化研究,建立显著值归一化模型。最后,采用灰度共生矩阵进行特征提取,将提取的特征向量输入概率神经网络进行疵点辨识。通过改进频域协调显著方法和Gabor滤波簇方法的对比实验研究发现:基于改进频域协调显著算法的疵点检测方法能够在保证疵点检测精度的前提下,运算速度比Gabor小波方法提高70%。
[Abstract]:Improving the accuracy and efficiency of defect identification is of great significance to improve the quality of textiles. Aiming at the problem of uneven illumination and low contrast of defect image, the defect image enhancement method based on Gabor wavelet cluster is studied. Using 15 Gabor filter clusters with three scales and five directions to filter defect images in different directions and scales to reduce the influence of uneven image and low contrast on the accuracy of feature extraction. The filtered image is divided into two neighborhoods whose area is equal to each other, and high dimensional feature vectors are extracted from the neighborhood. Then, for the problems of high dimension of Gabor eigenvector and large redundant information, The method of equidistant mapping is used to reduce the nonlinear dimension of Gabor features, eliminate redundant information from high dimensional features, and enhance the classifier fitting ability to enhance the sensitivity of Gabor features. Aiming at the difficulty of selecting structural parameters in the process of Gabor feature dimensionality reduction using the isometric mapping algorithm, the structural parameter optimization model is established by using the Ncut criterion as an appropriate function, and the discrete ion swarm algorithm is used to optimize the structure parameters. A parameter optimization method for equidistant mapping based on particle swarm optimization and Ncut criterion is proposed, and it is difficult to extract low-dimensional feature of new samples in offset mapping algorithm. Based on the assumption that the geometric structure of samples is the same in high-dimensional space and low-dimensional space, a new low-dimensional embedding model of samples is established, and a new low-dimensional feature extraction method is proposed. Finally, the low-dimensional feature is input into the probabilistic neural network classifier for defect identification. In the experimental research, two groups of defect image data of different textures are used to carry out experimental research. The results show that the defect detection accuracy is about 97% based on Gabor filter cluster and equidistant mapping algorithm. Using the advantages of strong anti-noise ability and small computational complexity of frequency domain coordination algorithm, instead of Gabor filter cluster, it can be used for defect image enhancement. Aiming at the problem that the defect identification accuracy of frequency domain coordination algorithm is greatly affected by the size of Gao Si filter template, the Ncut criterion is used as a moderate function. The template size optimization model of Gao Si filter is established, and the discrete ion swarm algorithm is used to optimize the parameters. Aiming at the problem that the Lab color space has no obvious effect on single color textile defects, the HSV color space is used to replace the Lab color space. Aiming at the problem that the significant value of each component is not well reflected due to the difference in the range of the values of hue feature, saturation feature and luminance feature, the color feature is developed. The normalization of saturation feature and luminance feature is studied, and the normalized model of significant value is established. Finally, the gray level co-occurrence matrix is used to extract the feature. The feature vector input probabilistic neural network is used for defect identification. Through comparing the improved frequency domain coordination saliency method and Gabor filter cluster method, it is found that the defect detection method based on improved frequency domain coordination saliency algorithm is based on improved frequency domain coordination saliency algorithm. The method can guarantee the precision of defect detection, The operation speed is 70% higher than that of Gabor wavelet method.
【学位授予单位】：武汉纺织大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】