Physical priors accelerate neural network training for imaging through multimode fiber

In multimode fiber (MMF) imaging, neural networks typically require large datasets and extensive training times, which present substantial challenges for both data acquisition and practical deployment. To address this limitation, we propose a pre-training approach that utilizes data generated from a physical model. By incorporating physically based training data, the required training time for actual fiber data is reduced to approximately one-third. Additionally, this method demonstrates superior performance on the test set, with a 0.04 increase in the average structural similarity index measure (SSIM) and a 31.25% reduction in the average mean squared error (MSE) compared to the baseline model. This approach not only accelerates the training process but also improves the model's accuracy and generalization capabilities, offering a more efficient and effective solution for MMF imaging.