煤岩识别技术是实现开采、掘进智能化和无人化的关键技术之一。为提高煤岩图像识别技术的精度和效率，提出了一种基于塔式池化架构和卷积神经网络技术的煤岩图像分割识别网络模型（Coal-Rock Pyramid Network，CRPN）：① CRPN模型在图像编码部分采用了深度可分离卷积，使用了混合空洞卷积技术和嵌入全局注意力机制的残差卷积模块，在提高计算效率的同时扩大模型的感受野，并降低了全局无关特性对后续特征图的不利影响。CRPN模型在图像解码部分采用了基于空间塔式池化架构的计算框架，弱化了特征图内部不同区域之间关联信息的损失，显著增强了对全局信息的有效表征。② 为保证模型训练的有效性，使用高感光度井下防爆相机对煤矿井下薄煤层采煤工作面原位煤岩图像进行了信息采集，包括完整状态、含有裂隙和阴影、暗光条件、暗光且有支护遮挡等4类典型条件下的煤岩分布图像。通过噪声添加、改变图像的特征和形态等数据预处理方法，建立了含有6 400个有效样本的煤岩高清图像数据库。③ 提出了基于交叉熵损失函数和修正自适应矩估计的训练优化算法，兼顾了模型的训练效率和精度。④ 选择像素准确度和交并比指标对CRPN模型的识别效果进行评估，结果表明CRPN模型的2项指标平均值分别为96.05%和91.54%，优于U-net模型和Segnet模型等现有其他煤岩图像识别方法。CRPN模型单张图像计算时间平均值为0.037 s，高于井下防爆摄像设备25 fps的成像能力，具备现场应用部署条件。将CRPN模型部署在工作面现场获取的动态视频中进行测试，测试结果表明该模型在稳定和抖动条件下均取得了良好的煤岩识别效果，验证了该技术在复杂环境中的可行性、鲁棒性。

Coal androck identification is one of the key technologies to realize intelligent and unmanned mining and heading. In order to further improve the accuracy and efficiency of coal rock image recognition technology,a coal-rock image segmentation model(Coal-Rock Pyramid Network,CRPN) based on spatial pyramid pooling architecture and convolutional neural network is proposed in this paper. ① Depthwise separable convolution is used in the coding part of CRPN. With the support of hybrid dilated convolution and residual convolution embedded with the global attention mechanism,the computational efficiency and the receptive field have been improved,and the adverse effects of global irrelevant features on subsequent feature maps are reduced. The computing framework based on the spatial pooling architecture is applied to the decoding part,weakening the loss of related information between different regions in the feature map,and enhancing the effective representation of global information. ② To ensure the effectiveness of model training,The in-situ coal and rock images from the thin seam mining face are collected by high sensitivity underground explosion-proof SLR camera,including four typical conditions:complete state,crack and shadow,dark light,dark light with wire mesh supporting. The image preprocessing was executed,including noise addition and changing the characteristics and morphology of the image. A high-definition coal-rock image database containing 6 400 valid samples has been established. ③ A training optimization algorithm based on cross-entropy loss function and rectified adaptive moment estimation is proposed,taking into account the efficiency and accuracy. ④ The pixel accuracy and intersection over union are selected as indicators to evaluate the recognition effect of CRPN. The results show that the average values of the two indicators of CRPN are 96.05% and 91.54%,respectively,which are better than other existing coal-rock recognition models such as U-net and Segnet. The average calculation time of a single image of CRPN model is 0. 037 s,which is higher than the frame rate of mining explosion-proof camera equipment(25 fps) and proves its application deployment ability. The CRPN model was deployed in the dynamic video obtained on the working face for testing. The test results show that the model has achieved correct coal-rock recognition results under both stable and jitter conditions,verifying its feasibility and robustness in complex environments. 

1 基于塔式池化架构的煤岩图像分割网络模型架构设计

   1.1 CRPN网络模型整体架构

   1.2 深度可分离卷积构建特征提取网络

   1.3 混合空洞卷积（HDC）扩大特征图感受野

   1.4 嵌入全局注意力机制的残差卷积模块

   1.5 基于空间塔式池化架构的计算框架

2 现场原位煤岩高清图像数据库

   2.1 原始数据采集

   2.2 数据预处理方法

   2.3 煤岩图像数据库的建立

3 模型学习训练

   3.1 训练环境

   3.2 损失函数设计

   3.3 优化算法设计

4 煤岩识别效果及验证

   4.1 主观分析

   4.2 客观分析

   4.3 井下现场动态验证试验

5 结论