一种用于光刻SEM图像轮廓提取的Canny优化算法

Optimization of Canny Algorithm for Contour Extraction of Lithography SEM Image

根据摩尔定律,半导体芯片单位面积上的晶体管数量每18个月翻一番,当前量产先工艺节点已经演进到3nm,特征尺寸不断缩小,图案越发复杂,芯片制造过程中出现缺陷的概率升高,对检测技术及其关联设备提出了极高的要求。光刻作为半导体制造中的核心模块,其工艺质量直接决定了后续工艺准确度,由于图案尺寸小于光源波长,衍射效应显著,导致曝光图案的高频信息丢失,光刻扫描电子显微镜(scanning electron microscope,SEM)图像作为曝光图案的直接表示,基于SEM图像的测量在半导体制造中对于图案化工艺质量的评估和监控显得至关重要。SEM图像不仅能够提供图像的特征宽度和特征-特征间距尺寸测量,同时还提供了其它有关图案质量的丰富信息,然而通过视觉检测在SEM图案质量检测方面留下了相当大的模糊空间。为了缩小模糊范围并获得更多关于图案质量的统计定量信息,通常会提取SEM图像轮廓来进一步计量分析。基于轮廓信息,我们能够估计出曝光图案任何位置的尺度信息,例如侧壁角度,临界尺寸等,这些尺度信息可用于光刻热点检测和光刻OPC模型验证等。基于局部信息的经典轮廓提取算法在处理存在噪声的低对比度光刻SEM图像方面尚有不足,如传统的prewit、log、sobel等算子,利用模板匹配无法获得令人满意的轮廓结果;轮廓提取领域的canny算法因其适用性广和效果较好,被广泛应用,但其梯度图像及阈值先验设置往往具有较大主观性,且没有充分利用待处理场景信息,对于较为复杂场景效果仍有待提升。本文通过结合通过结合光刻SEM图像的灰度、拓扑和轮廓长度先验信息,对canny算法进行优化,利用先验信息获得合适的阈值,完成对伪边缘(轮廓)的有效过滤,最终得到了与光刻SEM图像高度吻合的轮廓结果。

According to the Moore’s Law,the number of transistors in unit area doubles every 18 months.The current mass-production process node has evolved to 3nm,the feature size is shrinking and the pattern is more complex,then the probability of defects in the chip manufacturing process is increasing.Extremely high requirements for defect detection technology and related equipment have been put forward.As a core module in semiconductor manufacturing,lithography directly determines the quality of subsequent processes.Since the pattern size is smaller than the wavelength of the light source,the diffraction effect is significant,resulting in the loss of high-frequency information of the exposure pattern.The lithography SEM image is used as a direct representation of the exposure pattern,and SEM image-based measurements are crucial for the evaluation and monitoring of process quality.SEM images not only provide feature width and feature-to-feature spacing dimension,but also provide a wealth of information about pattern quality.However,visual inspection leaves considerable room for ambiguity,in order to narrow the blurring range and obtain more statistical quantitative information,SEM image contours are usually extracted for further metrological analysis.Based on the contour information,we can estimate the scale at any position of the exposure pattern,such as sidewall angle,critical dimension,etc.,these scales can be used for lithography hot spot detection and lithography OPC model verification.The classical contour extraction algorithm based on local information is often insufficient in processing low-contrast lithography SEM images with noise,such as traditional prewit,log,sobel and other operators,can not obtain satisfactory contour results.The Canny algorithm is widely used because of its efficiency and good effect,but its gradient image and threshold settings are often subjective,and do not make full use of scene information to be processed.In our work,the canny algorithm is optimized by combining the grayscale,topology and contour length prior information of the lithography SEM image,and the appropriate threshold is obtained by using the prior information to complete the effective filtering of false contours,and finally the obtained results are in good agreement with the original SEM images.