In semiconductor inspection and metrology on scanning electron microscopy (SEM) images, image noise affects the results of inspection and metrology. Image accumulation is effective for denoising but slows image grabbing duration. To get low noise images in high throughput inspection and metrology, we developed a novel denoising algorithm that converts a lowaccumulated image into a clean image like a high-accumulated image. Noise2Noise is one of the image denoising technologies by deep learning for natural images. In this method, clean images are not required for training because the Noise2Noise model is generated with pairs of original images and noisy images created by artificially adding Gaussian noise to the original images. It is more practical than other deep learning methods because collecting clean images is usually difficult. On the other hand, Noise2Noise doesn’t perform enough in SEM images because the noise on the SEM image is not Gaussian noise. To solve this problem, in this study, we analyzed SEM noise characteristics by changing SEM conditions to create the artificial SEM noise. Furthermore, we developed the novel denoising algorithm which is based on Noise2Noise but is specialized to train the artificial SEM noise. We confirmed the improvement of the roughness precision of the proposed method compared to the deep denoise model trained using simple artificial noise. We discuss the impact on throughput advantage of inspection and metrology by applying the proposed method in NGR3500.
|