【摘要】 针对传统散射成像方式范围小、系统稳定性要求高、设备复杂等问题，提出了一种基于深度学习的散射成像实验设计方案。该方案采用监督学习的方式，利用神经网络从实验散斑数据中学习光线在散射介质中的传播规律，实现对散斑图像的重建。具体实验过程：首先构建散射系统，并在其基础上获取不同种类的散斑图案用于网络训练；然后使用Unet 3+网络模型对散射过程进行拟合，通过多尺度跳跃连接的方式实现对图案中低维特征的高效利用；最后利用训练好的网络实现对不同种类的散斑图案的高质量重建。实验结果表明，本方案无需昂贵的硬件设备，结构简单，易于操作，训练好的Unet3+网络相比于广泛应用的Unet模型具有更高的重建质量，且针对新类型的散斑数据体现出一定的泛化能力。更多还原

【Abstract】 This study takes the scattering imaging problem as the research background and proposes a experiment schedule based on deep-learning. In this work, we adopt the method of supervised learning and utilize neural network to learn the propagation of light in the scattering medium from the speckle dataset in order to reconstruct object information. Specifically, an experimental scattering optical system is built, then different kinds of speckle patterns are obtained to train the network. Besides, the Unet 3+ network is taken to fit the relationship between groundtruth and speckle patterns, and the low-dimensional features are effectively utilized in the way of multi-scale skip connections. The experimental results show that our scheme is facility to operate without extra hardware support, and the well-trained Unet 3+ model can achieve much higher reconstruction quality compared with Unet network, and has certain generalization ability for new types of speckle input.更多还原