Deep Learning Network for Speckle De-Noising in Severe Conditions

Abstract

Digital holography is well adapted to measure any modifications related to any objects. The method refers to digital holographic interferometry where the phase change between two states of the object is of interest. However, the phase images are corrupted by the speckle decorrelation noise. In this paper, we address the question of de-noising in holographic interferometry when phase data are polluted with speckle noise. We present a new database of phase fringe images for the evaluation of de-noising algorithms in digital holography. In this database, the simulated phase maps present characteristics such as the size of the speckle grains and the noise level of the fringes, which can be controlled by the generation process. Deep neural network architectures are trained with sets of phase maps having differentiated parameters according to the features. The performances of the new models are evaluated with a set of test fringe patterns whose characteristics are representative of severe conditions in terms of input SNR and speckle grain size. For this, four metrics are considered, which are the PSNR, the phase error, the perceived quality index and the peak-to-valley ratio. Results demonstrate that the models trained with phase maps with a diversity of noise characteristics lead to improving their efficiency, their robustness and their generality on phase maps with severe noise.