Universal Image Enhancer for TEM/SEM Images

Title: Universal Image Enhancer for TEM/SEM Images

Authors:
Riti Paul [1], Abhinav Kumar [2], Baoxin Li [1]

[1]: School of Computing and Augmented Intelligence, Arizona State University, Tempe, AZ, 85287 USA
[2]: Applied Materials Inc., Santa Clara, CA, USA

Abstract:
Transmission Electron Microscopy (TEM) and Scanning Electron Microscopy (SEM) are critical imaging modalities in semiconductor manufacturing, providing high-resolution visualization of material structures for precise metrology and inspection. However, these images are often degraded by complex, acquisition-dependent noise, which impairs the accuracy of downstream tasks such as structural analysis and dimensional measurement. Furthermore, conventional denoising methods, while effective at noise suppression, often oversmooth images, leading to the loss of fine-grained features essential for accurate metrology. This work proposes a robust and efficient universal denoising and feature enhancement framework for TEM and SEM images, designed to reduce noise while preserving and enhancing critical structural details. The framework operates within stringent runtime constraints ( < 1 second per image) to support high-throughput, automated inspection workflows in semiconductor manufacturing environments.

Our motivation stems from two key challenges: (1) the impracticality of acquiring large-scale, clean, annotated datasets in semiconductor settings, and (2) the need for adaptive noise handling across diverse imaging conditions. To overcome these, we adopt a test-time adaptation (TTA) strategy, enabling the model to adapt to each image individually without supervised fine-tuning. By leveraging self-supervised learning, the model refines its representation during inference using only the noisy input image. Specifically, we employ masked image modeling—drawing from masked autoencoder strategies—combined with auxiliary reconstruction tasks to guide the network in learning robust and structure-preserving representations. To enhance adaptation, we introduce a difficulty-aware masking scheme that dynamically emphasizes regions with high noise or structural complexity, ensuring that the model adapts effectively to varying image conditions.

In addition to denoising, we recognize the critical need for unsupervised feature enhancement to counteract the smoothing effect commonly observed post-denoising. To this end, the proposed framework integrates deep learning-based enhancement techniques that simultaneously perform unsupervised contrast enhancement and feature sharpening. We integrate contrastive self-supervised learning objectives that promote the discrimination of subtle local structures’ variations, such as edges and textures. Additionally, perceptual enhancement losses are applied to amplify important structural cues and improve contrast, ensuring preservation of key features vital to metrology.

Our contributions are two-fold:
1. We present a lightweight, self-supervised denoising model that leverages test-time adaptation to generalize across diverse semiconductor images without relying on clean ground truth data.
2. We propose an unsupervised feature enhancement module that integrates contrast amplification and structure sharpening into the denoising pipeline, preserving essential details for accurate metrology.

By effectively addressing both noise reduction and feature preservation, this framework enhances automation, reduces manual intervention, and advances the reliability of semiconductor manufacturing inspection processes.