深度神经网络和高光谱显微图像的二维材料纳米片识别

近年来，二维材料由于其独特的性质而受到了广泛关注。在制备二维层状晶体的各种方法中，机械剥离法获得的薄层二维材料晶体质量高，适用于基础研究及性能演示。然而用机械剥离法从衬底上获得的材料具有一定的随机性，可能包含了少许相对较厚的部分。实现对这些二维薄层材料有效、快速且智能化的表征有利于促进二维材料性能的进一步研究。提出了一种基于深度学习的表征方法，通过搭建的编解码结构的卷积神经网络语义分割算法，可以根据光学显微镜图像进行分割和快速识别二维材料纳米片。卷积神经网络作为深度学习在图像处理领域中的典型算法，能够对光学显微镜图像中的复杂信息进行特征提取。首先采用机械剥离制备MoS₂纳米片样本，通过光学显微镜采集高光谱图像并对样本进行标记，根据样本的厚度范围标记出不同的区域，对标记后的图像进一步处理，包括图像的颜色校准和剪切操作，得到用于网络训练和测试的数据集。针对光学图像中二维纳米薄片存在的低对比度、碎裂等特点，编码时加入残差结构和金字塔池化模型，有助于特征信息的提取；解码时融合编码路径中提取的浅层特征信息，以提高网络分割精度。实验中采用带权重的交叉熵损失函数解决类别数量不平衡问题和采用数据增强扩大数据集。对训练后的网络测试结果表明，模型像素精度为97.38%，平均像素精度为90.38%，均交并比为75.86%。之后通过迁移学习成功地对剥离的单层和双层石墨烯纳米片样本进行了识别，均交并比达到了81.63%，表明该方法具有普适性。通过MoS₂和石墨烯纳米片的识别演示，实现了深度学习在二维材料的光学显微镜图像中的成功应用。该方法有望在更多的二维材料上得到扩展并突破自动动态处理光学显微镜图像的问题，同时为其他纳米材料的高光谱图像处理提供参考。

Identification of Two-Dimensional Material Nanosheets Based on Deep Neural Network and Hyperspectral Microscopy Images

In recent years, two-dimensional materials have received widespread attention due to their unique properties. Among the various methods for preparing two-dimensional layered crystals, the thin-layer two-dimensional material crystals obtained by mechanical exfoliation are of high quality, suitable for basic research and performance demonstration. However, mechanically exfoliated crystals on substrates exhibit a certain degree of randomness, including a few layers and relatively thick flakes. The effective, rapid and intelligent characterization method of these two-dimensional nanostructures is beneficial to further research on the properties of two-dimensional materials. This paper proposes a method based on deep learning, which can segment and quickly identify two-dimensional material nanosheets based on optical microscope images through a convolutional neural network semantic segmentation algorithm built with an encoding-decoding structure. As a typical algorithm for deep learning in the field of image processing, Convolutional neural networks can be applied to the feature extraction in optical microscope images. Firstly, MoS₂ nanosheet samples were prepared by mechanical exfoliation, and high spectroscopic images were acquired by optical microscopy. The nanosheet samples were labeled, and the marked images were further processed, including color calibration and sliding shear operation, to obtain datasets for network training and testing. A semantic segmentation algorithm based on encoding-decoding network structure was designed to rapidly identify nanosheets. Aiming at some flakes in images showing the characteristic of low contrast and fragmentation, residual convolution and pyramid pooling models were added to strengthen the extraction of features during encoding. The shallow feature information extracted from the encoding stage was reused during decoding to improve the network segmentation results. In the experiment, the weighted cross-entropy loss function was used to solve the problem of unbalanced classes, and the dataset was enlarged with data augmentation. Testing on the trained network show that the pixel accuracy was 97.38%, the mean pixel accuracy was 90.38%, and the mean intersection over union was 75.86%. Then, the exfoliated monolayer and bilayer graphenes were identified by transfer learning, and the mean intersection over union reached 81.63%, showing that this technique is universal for the identification of two-dimensional nanosheets. The identification of MoS₂ and graphene nanosheets realizes the application of deep learning in optical microscopy images of two-dimensional materials. This method is expected to apply to more two-dimensional materials and break through the problem of automatic dynamic processing of optical images. Moreover, it provides a reference for hyperspectral images processing of other nanomaterials.