Multimode fiber image reconstruction based on parallel neural network with small training set under wide temperature variations

Abstract

Much artificial neural networks (ANN) related research has focus on isothermal imaging, but many applications require variable temperature imaging capabilities. Therefore, a parallel neural network model is proposed for high quality reconstruction of unknown temperature specklegrams, which can be used to recover images of large-scale thermal variations using only a small training set. The results show that the proposed system achieves averaged SSIM value of 0.8035 on the 800-training set and 0.7702 on the 400-training set, even 0.6350 on 160-training set in the range of 16–56 °C, and has good immunity to specklegram cropping and laser power. Moreover, due to the fast training of the model and easy deployment of the system based on small training set, the temperature range can be easily extended by adding a few more models, which considerably enhances the role of multimode fiber (MMF) in encrypted image transmission and clinical long-term monitoring.