Contour detection of an object is a fundamental computer vision problem in image processing domain. The goal is to find a concrete boundary for pixel ownership between an OOI (object-of-interest) and its corresponding background. However, contour extraction from low SN SEM images is a very challenging problem as different sources of noise shadow the estimation of underlying structural geometries. As device scaling continues to 3nm node and below, the extraction of accurate CD contour geometries from SEM images especially ADI (after developed inspection) is of utmost importance for a qualitative lithographic process as well as to verify device characterization in aggressive pitches. In this paper, we have applied a U-Net architecture based unsupervised machine learning approach for de-noising CD-SEM images. Unlike other discriminative deep-learning based de-noising approaches, the proposed method does not require any ground-truth as clean/noiseless images or synthetic noiseless images for training. Simultaneously, we have also attempted to demonstrate how de-noising is helping to improve the contour detection accuracy. We have analyzed and validated our result by using a programmable tool (SEMSuiteTM) for contour extraction. We have de-noised SEM images with categorically different geometrical patterns such as L/S (line-space), T2T (tip-to-tip), pillars with different scan types etc. and extracted the contours in both noisy and de-noised images. The comparative analysis demonstrates that de-noised images have higher confidence contour metric than their noisy twins while keeping the same parameter settings for both data input. When the ML algorithm is applied, the contour extraction results would have higher confidence numbers comparing with the ones only applied the conventional Gaussian or Median blur de-noise method. The final goal of this work is to establish a robust de-noising method to reduce the dependency of SEM image acquisition settings and provide more accurate metrology data for OPC calibration.
|