Computational study demonstrates blind SNR estimation in scanning electron microscopy images, highlighting dataset-dependent model performance.
Key Points
To develop a blind signal-to-noise ratio estimation framework for scanning electron microscopy images that operates without clean reference images.
Constructed HSCF-KANet by fusing lightweight convolutional neural network features with multi-scale local variance and power spectral density descriptors using a Kolmogorov-Arnold Network regression head.
Evaluated the framework on three scanning electron microscopy datasets (EPFL CVLab EM, NFFA-EUROPE, and Biofilm) across synthetic Gaussian noise variances ranging from 0.001 to 0.010.
HSCF-KANet achieved a mean absolute error of 0.011 dB on EPFL CVLab EM, 0.049 dB on NFFA-EUROPE, and 0.314 dB on Biofilm.
Ablation experiments showed that combining convolutional neural networks with Kolmogorov-Arnold Networks yielded the best root mean squared error on NFFA-EUROPE, whereas pairing them with Random Forest was optimal on Biofilm.
Cross-dataset evaluations indicated that incorporating frequency-domain features reduced generalization performance across different image sets.