Citation

Lew, Kai Liang (2026) Noise analysis and denoising in Scanning Electron Microscopy images using deep learning. PhD thesis, Multimedia University.

Full text not available from this repository.

Abstract

Accurate estimation and management of additive white Gaussian noise (AWGN) are fundamental to scanning electron microscopy (SEM) workflows because noise degrades defect detection, metrology, and morphological analysis. This thesis develops and evaluates statistical and deep learning methods for three main tasks: classifying Gaussian noise variance, estimating single-image signal-to-noise ratio (SNR) without manual calibration, and performing blind denoising when the noise level is unknown. Experiments use three publicly available datasets, including EPFL CVLab, NFFA EUROPE, and Biofilm. Images from each dataset are corrupted with Gaussian noise with variances from 0.001 to 0.010, incrementing by 0.001, before being split into training, validation, and test sets. Models are evaluated using classification accuracy for noise-level classification, mean absolute error (MAE) in dB for SNR estimation, and PSNR/SSIM for denoising. Two deep learning models are developed for noise level classification, namely modified MobileNetV3 variants and the proposed Gaussian Noise Level Classification Network (GNLCN). The modified MobileNetV3 variants are combined with other modules. The GNLCN integrates a hybrid spatial state-space module, multi-scale feature fusion, and a dual-head classifier to reduce confusion between adjacent variance levels. GNLCN achieves 98.98% accuracy on EPFL, 94.66% on NFFA, and 95.27% on Biofilm. The state-space modules and ordinal head allow the model to distinguish closely spaced variance levels. For single-image SNR estimation, three methods are proposed, namely, quadratic, sigmoid ACF estimator (QSE), CNN-based Calibration Map Network (CalibNet) and Hybrid Statistical and CNN Feature Kolmogorov-Arnold Network (HSCF-KANet). QSE is the statistical method that combines the quadratic, sigmoid and ACF functions to estimate SNR. The QSE runs fast and is interpretable, but it becomes unstable at low SNRs. CalibNet estimates dB-scale SNR from a single image using a regressor and a learned calibration map. HSCF-KANet fuses spatial CNN features and statistical features via a Kolmogorov-Arnold layer, reaching MAEs of 0.010 dB in EPFL CVLab, 0.208 dB in NFFA-EUROPE and 0.371 dB in Biofilm. For the blind denoising, the proposed HyperDenoiseNet performs blind denoising using a U Net-based architecture with self conditioned FiLM modulation and squeeze and excitation with spatial gating, achieving PSNR of 31.34 dB and SSIM of 0.905 in the EPFL CVLab dataset, PSNR of 30.91 dB and SSIM of 0.808 in the NFFA-EUROPE dataset, and PSNR of 36.18 dB and SSIM of 0.883 in the Biofilm dataset. Overall, the developed methods provide practical, efficient tools for assessing and restoring SEM noise.

Item Type:	Thesis (PhD)
Additional Information:	Call No.: Q325.73 .L49 2026
Uncontrolled Keywords:	Deep learning (Machine learning)
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	22 Jul 2026 08:51
Last Modified:	22 Jul 2026 08:51
URII:	http://shdl.mmu.edu.my/id/eprint/16379

Downloads

Edit (login required)