基于字典学习的电子级玻璃纤维布缺陷分类

Classification of e-glass fiber fabric defects based on dictionary learning

在电子级玻璃纤维布的缺陷分类中,由于每类缺陷特征的多样性以及其丰富的几何结构的存在,用于分类的特征提取方法具有挑战性。提出了一个自动发现特征的框架,即结构不相关性字典学习(dictionary learning with structured incoherence, DLSI)用于提取每类缺陷的特征,并贡献了电子级玻璃纤维布的数据集。首先,利用DLSI学习每类图像的缺陷特征得到一个特定类字典,该字典适合于表示来自该类的电子级玻璃纤维布缺陷,同时很难表示来自其他类的缺陷图像;接着对于待分类图像,利用学习到的特定类字典对其进行重构,得到相应的重构误差;最后根据误差最小准则对待分类图像进行分类。所提出的方法在玻璃纤维布数据集上的平均分类准确率可达96.33%,显示了DLSI模型对玻璃纤维布缺陷分类的适用性。

In the classification of e-glass fiber fabric defects,the feature extraction method for classification is challenging due to the diversity of each type of defect features and its rich geometric structures.In this paper,a framework for automatically discovering features,namely dictionary learning with structured incoherence(DLSI),is proposed to extract features of each type of defect automatically.And contributes to the dataset of e-glass fiber fabric.First,DLSI is used to learn the defect characteristics of each type of images to obtain a specific-class dictionary,which is suitable for representing e-glass fiber fabric defects from the class,while it is difficult to represent defective images from other classes.For the image to be classified,it is reconstructed by using the learned specific-class dictionary to obtain the corresponding reconstruction error.Finally,the classified image is classified according to the error minimum criterion.The average classification accuracy of the proposed method on the e-glass fiber fabric dataset can reach 96.33%,which shows the applicability of the DLSI model to the classification of e-glass fiber fabric defects.