A novel adaptive noise model selection framework for blind denoising of Scanning Electron Microscopy images

Scanning Electron Microscopy (SEM) is crucial for imaging micro-nano particles in material science, wherein faster dwell times increase noise, while slower dwell times increase the risk of sample damage. The dwell time is defined as the time required for an electron beam to remain focused per pixel of the surface. Rapid dwell times and high-quality maintenance are essential in certain instances; however, this introduces noise, necessitating denoising. There is uncertainty in the noise information from SEM images, which makes it more difficult to apply traditional or well-known supervised denoising techniques. Additionally, self-or unsupervised methods require processing time and computation, which makes us choose supervised and blind methods, but still has problems with prior noise information. To address this issue, an Adaptive Blind Denoising framework that identifies and adapts to local noise patterns without prior knowledge is proposed. The framework comprises three modules: a Dataset Synthesizer that utilizes the power set theory to generate noise combinations, a denoising network that combines Shifted Windows Transformer, Convolutional Neural Network in U-shaped network (UNet), and a Noise Model Ranker (NMRanker) that adapts the optimal noise patterns. It was validated on SEM images using metrics such as the peak signal-to-noise ratio (PSNR), Structural Similarity Index Measure (SSIM), keypoint detection (necessary for further three-dimensional reconstruction), and visual assessment. The framework demonstrated improved denoising by accurately adapting to noise patterns from noisy input, with Speckle and Poisson noise combinations exhibiting optimal performance. This approach enhances SEM imaging denoising and can be validated in other fields.