神经网络让多模光纤成像抗温变，50℃温差仍能清晰成像
Neural Network-Based Multimode Fiber Imaging and Characterization Under Thermal Perturbations
- 用单隐藏层密集神经网络提升成像对温度变化的鲁棒性
- 温度变化50℃时仍可准确重建自然图像，斑点图样剧烈变化也不影响
- 通过模型参数反推传输矩阵，揭示成像精度与温敏性的关系
基于机器学习的多模光纤（MMF）成像在医疗内窥等场景具有广阔前景，但其模态传输特性易受环境扰动影响。本文实验表明，采用神经网络（NN）的MMF成像方案对热扰动具有显著鲁棒性：即使在相对于训练条件温度变化高达50℃的情况下，自然图像仍可成功重建，尽管热致斑点图案发生显著变化。研究发现，仅含一个隐藏层的密集神经网络性能优于适用于标准计算机视觉任务的卷积神经网络。此外，我们证明了神经网络参数可用于重构近似传输矩阵，并揭示图像重建精度直接关联于MMF传输特性的温度依赖性。
原文摘要 · Abstract (English)
Multimode fiber (MMF) imaging aided by machine learning holds promise for numerous applications, including medical endoscopy. A key challenge for this technology is the sensitivity of modal transmission characteristics to environmental perturbations. Here, we show experimentally that an MMF imaging scheme based on a neural network (NN) can achieve results that are significantly robust to thermal perturbations. For example, natural images are successfully reconstructed as the MMF's temperature is varied by up to 50$^{\circ}$C relative to the training scenario, despite substantial variations in the speckle patterns caused by thermal changes. A dense NN with a single hidden layer is found to outperform a convolutional NN suitable for standard computer vision tasks. In addition, we demonstrate that NN parameters can be used to understand the MMF properties by reconstructing the approximate transmission matrices, and we show that the image reconstruction accuracy is directly related to the temperature dependence of the MMF's transmission characteristics.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。