Electro-Optical Spatial Resolution via Deep-Learned ESF Correction

The patent replaces traditional mathematical fitting methods (hyperbolic tangent, cubic spline) with a deep learning-based neural network model to correct ESF data. The neural network learns to denoise and correct ESF curves by training on synthetic data generated with known vibration characteristics, enabling effective removal of micro-vibration noise while avoiding the over-fitting and under-fitting problems of conventional fitting methods.

The patent converts the harmful effect of micro-vibrations into a beneficial training opportunity. Instead of eliminating vibrations through equipment shutdown, the system uses synthetic ESF data generated with simulated vibration characteristics to train neural network models. This allows the system to learn to correct vibrational noise patterns, transforming the measurement challenge into an effective noise removal solution without requiring vibration-free conditions during actual measurements.

The patent applies preliminary action by pre-training neural network models on synthetic ESF data generated with simulated vibration characteristics before actual measurements are taken. The training process creates a library of corrected ESF curves that can be quickly applied to real measurements. This preliminary training phase establishes the foundation for effective vibration removal without requiring complex real-time processing during actual measurements.