College of Energy Engineering, Xi'an University of Science and Technology
Key Laboratory of Western Mine Exploitation and Hazard Prevention, Ministry of Education, Xi'an University of Science and Technology
Sichuan Coal Group Xuyong No.1 Coal Mine Co., Ltd.

当前应用于煤岩图像识别的深度卷积神经网络模型存在体积庞大、计算过程冗杂等问题，难以满足实时检测要求，且对低照度、高粉尘等复杂环境适应性差。针对上述问题，提出了一种融合钻孔地质信息的煤岩图像识别方法。首先，通过改进的谱残差显著性检测（ISRSD）算法增强煤岩图像质量，有效减弱复杂环境对煤岩图像特征造成的不利影响；然后，使用加入注意力机制的VGG（AVGG）深度卷积神经网络模型——在VGG的基础上进行剪枝、加入卷积注意力模块（CBAM）和引入自适应学习率调整策略，高效提取煤岩图像特征；最后，利用贝叶斯模型融合煤岩图像特征和由钻孔地质柱状图获取的钻孔地质信息，提升煤岩分类的准确性和鲁棒性。实验结果表明，经ISRSD算法增强后的图像目标更突出，色彩失真程度更低，且边缘、纹理等图像特征保留相对完整； AVGG模型的准确率与VGG模型相当，但平均推理时间、参数量及模型大小分别仅为VGG模型的15.61%，33.44%及33.40%；与仅使用AVGG模型识别煤岩图像相比，利用贝叶斯模型融合钻孔地质信息后，准确率提高了1.85%，达97.31%。

The current deep convolutional neural network models applied to coal-rock image recognition have problems such as large volume and cumbersome calculation process. It is difficult to meet real-time detection requirements, and it has poor adaptability to complex environments such as low lighting and high dust. In order to solve the above problems, a coal-rock image recognition method integrating drilling geological information is proposed. Firstly, the improved spectral residual saliency detection (ISRSD) algorithm is used to enhance the quality of coal-rock images, effectively reducing the adverse effects of complex environments on the features of coal-rock images. Secondly, the method uses the attentional VGG (AVGG) deep convolutional neural network model. The AVGG performs pruning based on VGG, adds convolutional block attention module (CBAM), and introduces adaptive learning rate adjustment strategy to efficiently extract coal-rock image features. Finally, the Bayesian model is used to integrate the features of coal-rock images with the geological information obtained from the borehole geological column chart, in order to improve the accuracy and robustness of coal-rock classification. The experimental results show that the image enhanced by the ISRSD algorithm has more prominent targets, lower color distortion, and relatively complete preservation of image features such as edges and textures. The accuracy of the AVGG model is comparable to that of the VGG model, but the average inference time, parameter count, and model size are only 15.61%, 33.44%, and 33.40% of the VGG model, respectively. Compared with using only the AVGG model to recognize coal-rock images, using the Bayesian model to fuse drilling geological information improves accuracy by 1.85%, reaching 97.31%.

陕西高校青年创新团队项目（陕教函〔2022〕943号）。

李季，马潇锋，吴洁琪，等. 融合钻孔地质信息的煤岩图像识别方法[J]. 工矿自动化，2024，50（8）：38-43, 68.

LI Ji, MA Xiaofeng, WU Jieqi, et al. Coal-rock image recognition method integrating drilling geological information[J]. Journal of Mine Automation，2024，50（8）：38-43, 68.