Deep learning-based multimode fiber imaging of multiple objects with different spatial coherence and different polarization

It is well known that when a single object illuminated by a spatially coherent laser passes through a multimode fiber (MMF), a speckle pattern is obtained, meaning that the information of the object is lost. In recent years, deep learning has been successfully employed to reconstruct images with high-fidelity from the speckle. In this paper, we consider a more general situation in which two objects are located in two planes and illuminated by a laser with different spatial coherence and orthogonal polarization, and then passed through a MMF. In distal end of the MMF, a speckle is obtained, and the information of two objects are lost. A neural network named as TDUNet (two-decoder-UNet) has been built to reconstruct the images. It is shown that the trained TDUNet could reconstruct images with high quality from a single speckle. It is also found that even when the spatial coherence of the illuminating laser is low, the neural network can reconstruct the images of two objects in two different planes with high fidelity. The experimental results demonstrate that the performance of reconstruction imaging of the object (digital) in the first plane is slightly better than that in the second plane.