Where test data come from and what corrupts real data: synthetic conductivities, noise and modelling errors, and how to avoid overly optimistic results.
Reading order.
- Conductivities for testing and training: Synthetic Conductivity Data.
- Measurement errors: Noise Models for EIT Data.
- Honest simulations: Inverse Crime, and accounting for the modelling error that honest simulations reveal: Approximation Error Approach.
- Degraded images for training denoisers: Spectral Image Corruption.
Related. How much the data can reveal: Decay of Boundary Measurements. Stopping at the noise level: Choosing the Regularization Parameter. Priors trained on such data: Learned Priors.
References
- J. Kaipio, E. Somersalo (2005). Statistical and Computational Inverse Problems. Springer. doi:10.1007/b138659