Rapid adaptation of SEM image denoising model using conditional normalization

In semiconductor manufacturing, Scanning Electron Microscope (SEM) images are utilized for quality improvement such as critical dimension (CD) measurements, overlay (OVL) measurement assessments, and defect inspections. SEM images typically suffer from low signal-to-noise ratio (SNR), resulting in poor image quality. To address this, deep learning models have been introduced for image enhancement. A major challenge, however, is that these models require retraining whenever SEM image patterns change, leading to prolonged training times that can disrupt manufacturing workflows. In this study, we propose a method that pre-trains the model on general pattern images and applies fine-tuning for new SEM image pattern types. To enable efficient adaptation to different patterns, we incorporate conditional normalization into our deep learning model. Experimental results demonstrate that our approach reduces training time from hours to minutes while preserving the precision of OVL measurements compared to existing models.