为提高煤岩镜质组显微组分的识别准确率和识别模型性能，减少识别模型训练中的人工干预，以轻量级网络模型ShuffleNet V2作为主干网络，提出一种嵌入改进注意力机制的煤岩镜质组显微亚组分识别轻量级深度学习网络模型。首先针对镜质组样本数据量较少的问题，采用随机裁剪、旋转、镜像及加噪等方法对初始样本数据进行增强，提高训练模型泛化能力；然后以大型数据集ImageNet上完成预训练的ShuffleNet V2模型为基础，采用迁移学习的方法对模型进行微调，即在本文增强的镜质组显微图像训练样本集上将预训练模型中靠近输入的若干层冻结，对靠近输出的网络层权值进行微调，生成深层特征提取层；最后在ShuffleNet V2模型的输出部分嵌入通道注意力机制ECA，并根据煤岩镜质组不同显微亚组分图像具有明显纹理差异的特点对注意力机制进行改进，构建端到端的轻量级网络识别模型，实现对煤岩镜质组7类显微亚组分的自动识别。实验结果表明：采用本文方法对煤岩镜质组显微亚组分进行识别，其平均准确率可达97.85%，较之ShuffleNet V2原模型可提升5.71%；与经典的神经网络相比，本文模型具有较高的识别准确率、较少的网络参数和计算量及较快的收敛速度；与其他轻量级网络及嵌入其他注意力机制相比，该网络在保持较少参数量的同时，识别的准确率有明显的提高。

In order to improve the recognition accuracy of coal vitrinite macerals and performance of recognition model, reduce the manual intervention in the training of recognition model, a lightweight deep learning network model for coal vitrinite submaceral recognition, which is embedded with an improved attention mechanism, is proposed. Firstly, aiming at the problem of small amount of vitrinite sample data, a data enhancement method of random clipping, rotating, mirroring and noise adding is employed to enhance the original sample data and improve the generalization ability of the recognition model. Then, a backbone network is constructed based on the ShuffleNet V2 model pre-trained on a large data set ImageNet. A transfer learning method, which freezes some layers near the input and adjust the weight parameters of layers near the output of the pre-trained model according the enhanced training sample set of vitrinite microscopic image, is employed to generate some deep feature extraction layers. Finally, a channel attention mechanism ECA, which is improved according to the characteristics of distinct texture feature of vitrinite microscopic image, is embedded before the output of the ShuffleNet V2 model, and a new lightweight end-to-end recognition network model is constructed to implement the automatic recognition of seven submacerals of coal vitrinite. The experimental results show that the average recognition accuracy of submacerals with this method can reach 97.85%, gone beyond 5.71% of the original ShuffleNet V2 model. Compared with the classical neural network, the proposed model has higher accuracy, lesser number of parameters and calculation, and rapid convergence velocity. Compared with other lightweight networks and other attention embedding mechanisms, the proposed network not only maintains a less amount of parameters, but also significantly improves the recognition accuracy.