SEM Deep Learning Multiclass Noise Level Classification With Data Augmentation
Scanning electron microscopy (SEM) plays an important role in providing high-resolution imaging in various fields, including industrial chip manufacturing, materials science, and nanoscale biology. However, high-resolution imaging often compromises image quality due to noise, such as Gaussian noise,...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10993397/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Scanning electron microscopy (SEM) plays an important role in providing high-resolution imaging in various fields, including industrial chip manufacturing, materials science, and nanoscale biology. However, high-resolution imaging often compromises image quality due to noise, such as Gaussian noise, which is caused by electronic noise and signal amplification. Overcoming noise in SEM images traditionally involves the use of various filters, such as the Wiener filter and Gaussian filter, but these methods are often limited in efficiency when handling complex noise patterns, and they can be time-consuming. In recent years, the rapid evolution of deep learning technology has been applied to various machine vision tasks such as classification, detection, denoising, and super-resolution. In this study, a deep learning network, the Gaussian Noise Level Classification Network (GNLCN), is proposed to Classify noise levels in SEM images from 0.0000 to 0.0100, incremented by 0.001. A novel Hybrid Spatial State-Space Model (SSM) block is introduced which uniquely integrates a state-space model with Residual Squeeze-and-Excitation (SE) and other feature enhancement mechanisms to improve spatial feature extraction. The network uses a dual-head classifier, consisting of both a standard classification head and an ordinal classification head. GNLCN is compared with other state-of-the-art (SOTA) models, including ConvNeXT, ResNeXt, EfficientNetV2, Vision Transformer, and VGG-19. Three different SEM datasets are used, namely the NFFA-EUROPE dataset, EPFL CVLab EM dataset, and Biofilms SEM dataset. The proposed network achieved the highest evaluation in accuracy (92.76%), precision (92.77%), recall (92.76%), and F1 score (92.74%), as well as the shortest training time in the NFFA-EUROPE dataset. It also demonstrated consistently strong results on the EPFL CVLab EM dataset and Biofilms SEM dataset, surpassing most competing methods, and thus sets a new benchmark in SEM image analysis. This model significantly enhances noise level classification in SEM images, can be widely applied in various machine vision tasks, and holds the potential to revolutionise SEM imaging techniques. |
|---|---|
| ISSN: | 2169-3536 |