工件表面缺陷图像检测中的自适应聚类

目的 针对工件表面形态复杂、干涉交叠缺陷难以实现自动分离、分类图像检测的情况，提出一种工件表面交叠缺陷自适应图像聚类方法。方法 首先提取工件表面缺陷二值图像，采用混合概率主成分分析器估计缺陷位置各像素点局部切空间信息，并改进局部切空间之间的相似性矩阵；然后通过改进局部密度峰值自适应方法，基于相似性矩阵确定聚类中心点和数目；最后通过谱多流形聚类，将各分析器所包含的像素点分配至不同缺陷流形结构中，实现多个交叠缺陷分离、检测。结果 首先通过比较计算与实际测量的长度、宽度来验证该方法对相互交叠结构缺陷良好、准确的分离效果，平均相对误差分别为0.957%和0.650%。其次为了体现该方法对于分离工件表面相互交叠缺陷的有效性及优越性，使用k-means聚类、谱聚类与该方法进行对比实验，证明了该方法良好的聚类效果。最后对所设计方法的稳定性进行测试，统计检测结果的平均ME值均在6%以下，正确聚类数目率高达99%~100%。结论 该方法能够较为准确地自动识别工件表面图像中存在相互干涉的不同缺陷，并进行分离。

Adaptive Clustering Method of Image Detection for Work-piece Surface Defect

The work aims to propose an adaptive image clustering method for overlapping defects on work-piece surface, so as to solve the problem that the complex and mutually interfered defects on work-piece surface are difficult to be separated automatically and classified and identified by images. Firstly, the binary image of the workpiece surface defects was extracted. The principal component analyzer of mixed probabilities was used to estimate the local tangent space of each pixel on the defects and improve the similarity matrix between the local tangent space of each defect location. Then, the clustering center point and quantity were determined through the improved density peaks adaptive method based on the similarity matrix. Finally, the pixels included in each analyzer were assigned to the different defect manifolds through SMMC to realize the separation and detection of workpiece defects. Firstly, the good and accurate separation effect of this method on the interfered structure defects was verified by comparing the calculated and actual length and width. The average relative errors were 0.957% and 0.650%. Secondly, in order to reflect the effectiveness and superiority, this method was compared with k-means clustering and spectral clustering, which proved a good clustering effect. Finally, the stability of the method was tested. The average ME value was below 6% for the statistical test results, and the correct cluster number rate was as high as 99%~100%. The experimental results show that this method can automatically separate different defects that interfere with each other in the surface image of the workpiece more accurately.