Robust and fast auto-focusing using convolutional neural networks for off-axis digital holography

In this paper, we propose a robust and accurate estimation method for the distance required for digital holography (DH) reconstruction using convolutional neural networks (CNN) in off-axis DH (off-axis DH). This method applies adaptive spectral pooling to reflect distance-related optical characteristics and minimize information loss during the training phase. Simulations and experiments have confirmed that the proposed method is more robust and accurate than search-based or CNN-based distance estimation methods.