Lightweight Learning Model for Speckle Denoising in Digital Holography

Digital holographic interference (DHI) is a well-known technique in optical metrology for recording and reconstructing 3D deformations. However, the interference patterns captured by DHI are usually affected by speckle noise, which can degrade the quality of the reconstructed deformations. In this work, we propose a lightweight, densely connected deep neural network(DNN) model that can effectively remove speckle noise from DHI images. To aid the DNN model to effectively learn the denoising task, we propose a structural similarity-based loss in addition to the popular \(L_2\) loss. The outcomes of our experiments demonstrate that the proposed model outperforms three well-known conventional methods and existing deep learning methods. In addition, the quantitative results and the analysis plots show 58 \(\%\) and 121 \(\%\) higher PSNR and SSIM scores over the state-of-the-art conventional denoising method. Towards the end, the denoising results presented on a real sample depicts that the proposed model is able to generalize well to real DHI images even with synthetic sample training.