IC Image Segmentation Using Eigenvalue Clustering

The influence of real defects on yield has been becoming a major obstruction in integrated circuits (IC) manufacturing. Since yield loss is connected closely with the outlines of real defects, it is a key step to extract these outlines for the yield model studying. In order to obtain real defects, a test chip that includes three structures (comb structure, shake structure, and nest structure) is designed and implemented in 0.5 mum process, and more than 358 simple defect images, each of which only includes one real defect, are collected. Furthermore, for the purpose of extracting outlines of real defects, a novel segmentation method based on the color model and clustering algorithm is presented, in which the piecewise linear discriminant functions and clustering criterion are created. The experimental results show that the proposed method can separate the outlines from IC defect images satisfactorily, no matter their backgrounds are uniform or no uniform, which provides a basis for the later studying of the defect outline and yield models.