Center of Shanxi Engineering Research for Coal Mine Intelligent Equipment, Taiyuan University of Technology
College of Mechanical and Vehicle Engineering, Taiyuan University of Technology
Post-doctoral Workstation, Shanxi Coking Coal Group Co. Ltd.

煤岩识别技术是实现煤矿工作面智能无人开采的关键技术之一。为进一步提高基于机器视觉实现煤岩界面图像识别的精度和效率，提出一种基于改进DeeplabV3+和迁移学习的煤岩界面图像识别网络模型：首先，使用轻量化MobilenetV2模块作为骨干特征提取网络，减少网络模型参数，提高语义分割效率；然后，在编码器和解码器中引入卷积注意力机制模块（CBAM），提高模型特征提取能力，并实现不同层级特征信息有效融合，提升模型分割精度；其次，采用迁移学习训练方法，克服样本分布差异性，增强模型泛化性，以适应于不同应用场景下的煤岩识别任务。应用自制煤岩分割数据集和综采面煤岩分割数据集验证模型性能，与FCN、SegNet、U-net、DeeplabV3+网络模型作对比试验，并选择准确度、平均交并比、推理时间等指标对模型识别效果进行评估。消融试验结果表明，改进DeepLabV3+网络模型在自制煤岩分割数据集上准确度和平均交并比分别为94.67%和93.48%，测试用时42.58 ms/张，采用推理加速框架TensorRT优化后推理时间可达6.14 ms/张，与其他模型相比，改进DeepLabV3+对煤岩边界细节特征提取能力更强，分割精度和处理效率更高。最后通过构建综采工作面含有煤岩层的煤岩图像分割数据集对改进DeepLabV3+模型采用迁移学习方法进行训练测试，实现了煤矿井下工作面的煤岩界面图像识别，验证了该方法在实际煤岩图像识别任务的可行性和有效性。

Coal-rock identification is one of the key technologies to realize intelligent and unmanned mining. To further improve the accuracy and efficiency of coal-rock interface image recognition based on machine vision, a coal-rock interface image recognition network model based on the improved DeeplabV3+ and transfer learning is proposed. Firstly, the lightweight MobilenetV2 module is used as the backbone feature extraction network to reduce the net-work model parameters and improve the semantic segmentation efficiency; Secondly, the Convolutional Block Attention Module (CBAM) is introduced into the encoder and decoder to improve the model feature extraction ability, effectively fuse feature information at different levels, and improve the model segmentation accuracy; Thirdly, the transfer learning training method is adopted to overcome the difference of sample distribution and enhance the generalization of the model, so as to adapt to the coal-rock recognition tasks in different application scenarios. The performance of the model is verified by using the self-made coal rock segmentation data set and the coal rock segmentation dataset of the fully-mechanized mining face. The model is compared with FCN, SegNet, U-net, DeeplabV3+network models, and the accuracy, average intersection ratio, and inference time indexes are selected to evaluate the model recognition effect. The ablation experiment results show that the accuracy and the mean intersection over union of the improved DeeplabV3+ network model in the self-made coal-rock segmentation dataset are 94.67% and 93.48%, respectively, and the test time is 42.58 ms·sheet-1. In addition, the model inference time can reach 6.14 ms·sheet-1 after optimization with the inference acceleration framework TensorRT. Compared with other models, the improved DeeplabV3+ shows stronger ability to extract detailed features of coal-rock boundaries, higher segmentation accuracy and processing efficiency. Finally, the dataset of coal-rock image segmentation with coal-rock layers in the fully mechanized mining face is constructed. The improved DeepLabV3+model is trained and tested on the dataset by using the transfer learning method, which realizes the coal-rock interface image recognition of the underground fully mechanized mining face, and verifies the feasibility and stability of this method in the actual coal rock image recognition task.

国家重点研发计划资助项目（2020YFB1314000）；山西省科技计划揭榜招标资助项目（20201101008）；山西省应用基础研究计划资助项目（202103021223123）

闫志蕊，王宏伟，耿毅德. 基于改进DeeplabV3+和迁移学习的煤岩界面图像识别方法[J]. 煤炭科学技术，2023，51（S1）：429−439.

YAN Zhirui，WANG Hongwei，GENG Yide. Coal-rock interface image recognition method based on improved DeeplabV3+ and transfer learning[J]. Coal Science and Technology，2023，51（S1）：429−439.