基于骨架局部曲率分水岭算法的颗粒特征识别分割方法

Particle Feature Recognition and Segmentation Method Based on Skeleton Local Curvature Watershed Algorithm

摘要: 针对晶圆生产过程软损伤吸杂工艺中，判断晶圆背部经软损伤处理后的吸杂能力，对背部类圆形杂质颗粒进行密度检测时，存在大量粘连颗粒无法准确计数的问题。本文基于Halcon视觉平台提出轮廓骨架局部曲率分析分水岭分割方法对粘连目标进行分割。首先将经过图像预处理的晶圆表面颗粒图像通过距离变换和分水岭算法进行初步分割。对分割后的区域进行凸性筛选，选择可能粘连的区域进行骨架的局部曲率分析，进一步分割粘连目标。解决在无法获取颗粒全部轮廓的情况下，大规模粘连分割问题，实现颗粒计数。实验结果表明，所提出的优化方法对于复杂情况下的粘连颗粒物能很好地适应，综合准确率达到96.7%，比传统的分水岭算法提高了8.5个百分点，验证了本文所提方法的有效性。

Abstract: For the soft damage gettering process in the wafer production process, to determine the gettering ability of the back of the wafer after the soft damage treatment, when the density of the back round impurity particles is detected, there is a problem that a large number of adhesion particles cannot be accurately counted. Based on the Halcon vision platform, this paper proposes a watershed segmentation method to analyze the local curvature of the contour skeleton to segment the adhesion target. First, the preprocessed wafer surface particle image is preliminarily segmented through distance transformation and watershed algorithm. Convexity screening is performed on the segmented regions, and the regions that may be adhered are selected for local curvature analysis of the skeleton, and the adhered targets are further segmented. Solve the problem of large-scale adhesion and segmentation when the full outline of the particles cannot be obtained, and realize the particle counting. The experimental results show that the proposed optimization method can adapt well to the adhesion of particles in complex situations, and the comprehensive accuracy rate reaches 96.7%, which is 8.5 percentage points higher than the traditional watershed algorithm, which verifies the effectiveness of the method proposed in this paper.