基于非下采样Contourlet域高斯混合模型的布匹瑕疵识别算法

针对经典瑕疵检测算法不能很好地描述瑕疵特征和不易区分相似类瑕疵类别的问题,提出了一种非下采样Contourlet域高斯混合模型的布匹瑕疵识别算法。首先利用非下采样Cont-ourlet变换(NSCT)得到图像的多方向、多尺度和平移不变表示,并通过代价函数挑选一个最优子带;由于子带瑕疵和非瑕疵区域系数差别较小,很难直接选取阈值,采用标准差法可以有效避免这个问题,获得比较准确的检测结果;然后计算瑕疵区域的统计特征得到特征向量;接着引入最小误分率函数,联合估计样本的高斯混合模型参数;最后采用贝叶斯分类器进行分类。在9类瑕疵上的实验结果表明,本文算法与几种经典算法相比得到更高的分类正确率。

Fabric defect recognition algorithm based on Gaussian mixture model in nonsubsampled Contourlet domain

Classical detection algorithms can not describe the defect characteristics in a better way,and it is difficult to classify the defects of similar classes.To solve these problems,an algorithm based on Gaussian Mixture Model(GMM) in nonsubsampled Contourlet domain for fabric defect recognition is proposed.First,the Nonsubsampled Contourlet Transform(NSCT) is used to obtain the representations in multi-directions,multi-scales and translation invariants,and an optimal sub-band is selected by the cost function.The difference of coefficients between defect and non-defect regions of the sub-band is smaller;therefore,the threshold is difficult to choose.The standard deviation method can effectively avoid this problem to get accurate results.After that the feature vectors are calculated by the statistical characteristics of the defect region.Then,the minimum misclassification function is used for joint estimation of the GMM parameters.Finally,the classification is completed by Bayesian classifier.Experiments on defects of nine categories show that the proposed algorithm can achieve better classification accuracy compared with the traditional algorithms.