Image Reconstruction Using Pre-Trained Autoencoder on Multimode Fiber Imaging System

Multimode fiber (MMF) based endoscopy could reach high resolution and is fine enough for vivo imaging. However, the received images are speckles due to the mode crosstalk and sensitivity to environment of MMF, which makes image reconstruction the main challenge. We propose to use pre-trained autoencoder for image reconstruction from speckles to original images, which shows high performance and fast convergence speed. The network architecture includes two parts, i.e ., encoder and decoder. In the first step, we pre-train the network to initialize the parameters of the decoder. In the second step, the network can learn the mapping relation between speckle patterns and original images. We conduct experiment of transmitting over one-meter MMF with 50- $\mu \text{m}$ -core to verify this method. Structural similarity (SSIM), peak signal-to-noise ratio (PSNR), and mean square error (MSE) are measured to evaluate the performance. Compared with U-net, the SSIM increases by 11% with pre-trained autoencoder. Moreover, the training of pre-trained autoencoder is both fast and steady.