Image transmission through a multimode fiber based on transfer learning

The multimode fiber is a kind of scattering medium, in which the light travels along different optical modes with different phase speeds. High-quality optical communications and medical endoscopic imaging can be carried out through a multimode fiber. However, one has to face a speckle pattern formed at the exit due to the distortion of the incident wave caused by multiple mode superposition and mode coupling. The convolutional neural network model U-net can be utilized to fit the input and output data to achieve the reconstruction of the input images from the speckle patterns at the output. Note that enough data is needed for training the network. However, it is often encountered that the amount of data is insufficient in practice. In order to solve this problem, a speckle image reconstruction method based on transfer learning and convolution learning model is proposed in this work. This model is able to realize the reconstruction of speckle images, and greatly reduce the demand for the quantity of training data. The experimental results confirm that our transfer learning model can reconstruct the speckle images effectively.

Automated summary

In this work, a speckle image reconstruction method based on transfer learning and convolution learning model is proposed, which can effectively reconstruct speckle images and greatly reduce the demand for the quantity of training data.