Transfer learning and generalization of a neural-network-based multimode fiber position and imaging sensor under thermal perturbations

Abstract

Fiber specklegrams created through multimode fibers have been utilized in a wide variety of imaging and sensing applications. In a practical environment it is useful to have a system that, after an initial calibration, produces reconstructed images and tracks the current state of relevant physical quantities such as the position or shape of the fiber. Neural networks provide the ability to extract and generalize the information contained in the specklegrams in order to produce accurate predictions even when experimental conditions are altered from their calibration state. Using a single-layer fully-connected network, we demonstrate the capability to perform robust image reconstruction under discrete position and temperature perturbations, as well as to classify the fiber position for different reconstructed images and thermal perturbations. On a test set with the same thermal state and bend configuration as for the training data, the network achieves near-perfect reconstruction and classification accuracy. Further, our neural network can generalize its position classification and image reconstruction capabilities across a 30 °C temperature perturbation with a high degree of accuracy, providing measurements that are robust to these variations, as well as calibrations that are stable and effective for several days. Moreover, using a low-temperature training set to initialize network weights, transfer learning is demonstrated by re-calibrating on a small set of data and shown to significantly improve robustness and performance of the imaging system on a high-temperature perturbation test set.