Speckle de-noising via deep learning in digital holographic interferometry

Abstract

Speckle de-noising can improve digital holographic interferometric phase measurements but may affect experimental accuracy. A deep learning-based speckle de-noising algorithm is developed referring to the U-Net and the DenseNet architectures using a conditional generative adversarial network established by the generator and the discriminator network. The loss functions that guide generator training consist of a mixture of a static spatial distance norm metric designed by considering the peak signal-to-noise ratio parameter, and a dynamic metric generated from the discriminator that grows with the generator in training. Datasets obtained from speckle simulations 4-f system are shown to provide improved noise feature extraction. Therefore, the proposed method offers better performance than other de-noising algorithms For processing experimental strain data from digital holography.