Neural Denoising of Electron Microscope Images Without Clean Ground Truth

The patent replaces conventional mechanical image processing techniques (like Gaussian blur or median filtering) with a deep learning-based neural network approach. The neural network learns to distinguish between noise and genuine edge information through training on synthetic images with known ground truth, allowing it to denoise while preserving sharp edges and fine structural details that would otherwise be lost in traditional filtering methods.

The patent changes the fundamental parameter of how denoising is achieved by moving from deterministic mathematical operations to data-driven learning. The neural network models the noise characteristics and learns optimal denoising transformations through training, enabling adaptive noise removal that preserves edge information better than fixed-parameter filtering methods.

The patent creates synthetic copies of the actual EM images by adding controlled noise patterns to clean versions. These synthetic noisy images serve as training data, allowing the neural network to learn denoising without requiring access to the original clean ground-truth images. The synthetic images replicate the noise characteristics of real EM data while providing the clean targets needed for supervised learning.

The patent uses synthetic images as disposable training data that can be generated on demand from the actual EM images. Instead of requiring rare and valuable clean ground-truth images, the system creates numerous synthetic training pairs by processing the same real images through controlled noise addition, providing abundant training data without compromising the scarcity of real clean data.