SPIE Journal Paper | 5 July 2024
Wei Xiao, Fazhan Zhao, Hongtu Ma, Kun Zhao, Qing Li
KEYWORDS: Scanning electron microscopy, Denoising, Performance modeling, Image processing, Image denoising, Education and training, Electron microscopes, Image quality, Data modeling, Diffusion
BackgroundScanning electron microscope (SEM) images of integrated circuits (ICs), which can provide details and information about the internal structure of ICs, play a critical role in hardware trust and assurance tasks. However, noise from various factors, such as equipment performance, environmental interference, etc., has a serious impact on the imaging quality of SEM images, which further leads to an imperfect performance on image feature extraction and analysis.AimTo solve the issues mentioned above, an SEM image denoising model based on the denoising diffusion probabilistic models (DDPM) is applied.ApproachThe applied model consists of a forward Gaussian noise addition process and a reverse denoising process. In the forward Gaussian noise addition process, a Markov chain of diffusion steps is used to gradually add random noise to images. In the reverse denoising process, we utilize an improved Unet architecture, which integrates the polarized self-attention (PSA) module into the residual blocks of the Unet structure to gradually transform the standard Gaussian noise to target ground truth (GT) distribution conditioned on the noisy images. Also, in the improved Unet architecture, we add an attention mechanism not only at 2x downsampled resolution but also at 4x and 8x downsampled resolutions to further enhance model performance. Then, we incorporate the complete form of the standard deviation into the reverse process, enabling more precise pixel-level probability predictions. Furthermore, a dataset consisting of noisy images and GT images pairs is proposed.ResultsThrough training and evaluation on the dataset, the proposed model exhibits good performance on the test set, achieving a peak signal-to-noise ratio (PSNR) of over 35.82 and a structural similarity (SSIM) of over 0.96.ConclusionsThe proposed model exhibits good performance in SEM image denoising and quality improvement, which provides insight for the denoising method research of SEM images of ICs, further contributing to a positive effect on hardware trust and assurance tasks.
