Deep learning enhanced quantitative phase imaging in digital holographic microscopy with attention empowered deep...

Using a pixelated polarization camera in digital holographic microscopy enables snapshot quantitative phase imaging, but at the cost of decreased resolution. In addition, the phase images also suffered from coherent noise. To address these two issues, we propose a deep image prior (DIP) based on an attention-enhanced double Unet (ADUnet+) to realize phase image interpolation and denoising simultaneously. Different from the supervised learning method that requires large training datasets, the DIP method adopts an unsupervised learning strategy and accomplishes these two joint tasks from a single noisy phase map by exploiting the implicit prior of ADUnet+, where the attention mechanism is used to effectively extract the features underlying the phase maps by rescaling weights of different feature maps in the network. Different learning strategies of the ADUnet+ are compared, and it is demonstrated that a sequential two-step mapping procedure yields better performance than a one-step procedure. Comparative results demonstrate that our method outperforms other baseline methods on both simulations and experiments, validating its high universal capability for various samples.