Fringe pattern background and amplitude modulation in DeepDensity: convolutional neural network for local fringe density map estimation

Proposed convolutional neural network based, fast and accurate local fringe density map estimation by DeepDensity was developed to significantly enhance full-field optical measurement techniques, e.g., interferometry, holographic microscopy, fringe projection or moiré techniques. The use of neural networks to determine the final result of the optical measurement may raise legitimate metrological concerns and therefore for the sake of versatility and independence from measurement technique we still recommend the use of fully mathematically sound solutions for both fringe pattern prefiltration and phase retrieval. It is worth to acknowledge that proposed DeepDensity network does not supersede mathematically rigorous phase extraction algorithmic solutions, but it only supports them. For that reason during the neural network learning process it was assumed that the data fed to neural network will be prefiltered so background and amplitude modulation should be successfully minimized. Nevertheless, it is still interesting how sensitive to the prefiltration accuracy is proposed DeepDensity. In this contribution we present a thorough analysis of the DeepDensity numerical capabilities in the case of insufficient fringe pattern (interferogram, hologram, moiregram) background and/or amplitude filtration. The analysis was performed with the use of simulated data and then verified using experimentally recorded fringe pattern.