直线导轨面缺陷检测与分类方法

Defect Detection and Classification Method of Linear Guide Surface

为解决复杂背景下直线导轨面缺陷识别难的问题,提出了一种基于灰度共生矩阵(Gray level co occurrence matrix,GLCM)和非负矩阵分解(Non negative matrix factor,NMF)的纹理背景抑制来实现缺陷特征增强的方法。首先,利用GLCM多特征统计量重构导轨面背景纹理图,实现一定程度上的纹理背景抑制;接着,将纹理图均分成若干子图像块,随机抽取一定的子图像块进行NMF训练;然后,将NMF分解出的基图像同纹理图中相同大小图像块遍历求其欧式距离,并将距离平均后赋值给纹理图中相应图像块的中心像素点,以进一步实现纹理背景抑制和缺陷特征增强。最后,基于K means聚类和支持向量机对缺陷进行分类识别。实验结果中对测试集中的划痕、裂纹和撞伤缺陷识别准确率分别为89.06%,88.46%和95.12%,表明该方法能抑制纹理背景和增强缺陷特征,有效分离出缺陷并识别其类型。

In order to solve the difficulty of segmenting the linear guide surface defects from image with a complex background,a method based on gray level co-occurrence matrix(GLCM)and non-negative matrix factorization(NMF)to suppress the texture background to realize defect feature enhancement was proposed.Firstly,the GLCM multi-feature statistics was used to reconstruct the background texture map of the linear guide surface to achieve a certain degree of texture background suppression.Then,the texture was divided into several sub-image blocks,and a certain sub-image block was randomly selected for NMF dimension reduction.Next,the basic matrix decomposed by NMF was traversed by the same size image block in the texture map to find its Euclidean distance,and the averaged distance was assigned to the center pixel of the corresponding image block in the texture image to realize texture background suppression and features enhancement.Finally,the defects were classified based on K-means clustering and support vector machine.In the experiment,the recognition accuracy of scratches,cracks and crash defects in the test set are 89.06%,88.46% and 95.12%,which shows that the proposed method can suppress the texture background effectively and enhance the defect features of the linear guide surface image,and it can separate the defects and identify their types accurately.