煤矸石分选作为煤矿开采过程中保证煤炭充分利用的一个必要环节，煤矸石自动化分选是建设智慧矿山不可或缺的技术。随着图像处理技术以及深度学习技术的不断发展，基于深度学习的目标检测算法开始应用在煤矸石分选场景并取得了显著成果。本文提出了一种基于YOLOv5的无锚框深度学习煤矸石检测算法，并针对煤矸石分选场景设计了嵌入式平台部署方案。

本文采用多种改进策略有效提高煤矸石检测算法性能。针对YOLOv5目标检测算法中正负样本分配不均衡的问题，采用无锚框策略完成目标框回归任务，同时更新了样本分配策略以及联合损失函数。针对煤矸石受环境影响难以被准确检测的问题，在YOLOv5网络结构中引入CA注意力机制，增强目标在复杂背景中的显著度，提高了特征的表达能力。YOLOv5目标检测算法在模型检测头部分采用共享卷积层完成分类和回归任务，两种任务在空间维度上的不一致性限制了检测性能，将共享卷积层解耦合生成两个卷积分支结构分别在不同的空间维度上完成分类和回归任务避免了不一致性问题。改进的煤矸石检测模型在三个不同地区煤矸石数据集上精度AP_50-95分别达到了74.9%、70.4%、82.6%，相对于原始YOLOv5算法提升了3.0%、4.7%、3.6%。实验结果表明，基于YOLOv5的无锚框深度学习煤矸石检测算法能够有效地提高煤矸石检测性能。

针对嵌入式平台算力较低以及功耗限制的问题，本文对无锚框深度学习煤矸石检测模型的网络结构进行了轻量化改进。通过重新设计主干网络实现模型轻量化，并利用TensorRT优化模型结构，降低模型参数的数值精度，提升嵌入式平台资源利用率，同时搭建了煤矸分选系统。在三个不同地区煤矸石数据集上，优化后模型的检测精度AP_50-95达到了72.2%、68.9%、79.2%，嵌入式平台上的检测速率达到52FPS，煤矸石分选实验中准确率均在95.0%以上。实验结果表明，轻量化煤矸石检测模型能够有效地检测煤矸石目标，检测性能优于当前优秀的轻量化目标检测模型，满足煤矸石分选应用场景需求。

Coal gangue separation is a necessary link to ensure the efficient utilization of coal in the process of coal mining, and the realization of automatic separation of coal gangue is an indispensable technology for the construction of smart mines. With the development of image processing technology and deep learning technology, object detection algorithms based on deep learning have been applied in coal gangue separation scenarios and have achieved remarkable results. The paper proposes an anchor-free deep learning coal gangue detection algorithm based on YOLOv5 and designs an embedded platform deployment scheme for coal gangue separation scenarios.

The paper adopts a variety of improvement strategies to effectively improve the performance of the coal gangue detection algorithm. Aiming at the problem of an unbalanced distribution of positive and negative samples in the YOLOv5 object detection algorithm, the anchor-free strategy is used to complete the box regression task, and the sample allocation strategy and joint loss function are updated at the same time. Aiming at the problem that coal gangue is difficult to be accurately detected due to the influence of the environment, the CA attention mechanism is introduced into the YOLOv5 network structure to enhance the saliency of the object in the complex background and improve the expression ability of the feature. The YOLOv5 object detection algorithm uses a shared convolutional layer in the model detection head to complete the classification and regression tasks. The inconsistency of the two tasks in the spatial dimension limits the detection performance. The shared convolutional layer is decoupled to generate two convolutional branch structures respectively. Completing the classification and regression tasks in different spatial dimensions avoids inconsistency issues. The improved coal gangue detection model achieves AP_50-95 of 74.9%, 70.4%, and 82.6% on coal gangue datasets in three different regions, which are 3.0%, 4.7%, and 3.6% higher than the original YOLOv5 algorithm. The experimental results show that the anchor-free deep learning coal gangue detection algorithm based on YOLOv5 can effectively improve the coal gangue detection performance.

Given the problems of low computing power and power consumption limitations of embedded platforms, the paper makes a lightweight improvement on the network structure of the anchor-free deep learning coal gangue detection model. The model is lightweight by redesigning the backbone network, and TensorRT is used to optimize the model structure, reduce the numerical accuracy of model parameters, and improve the resource utilization of the embedded platform. At the same time, a coal gangue separation system is built. On the coal gangue datasets in three different regions, the detection accuracy AP_50-95 of the optimized model reaches 72.2%, 68.9%, and 79.2%, and the detection rate on the embedded platform reaches 52FPS. The accuracy rates in the coal gangue separation experiments were all above 95.0%. The experimental results show that the lightweight coal gangue detection model can effectively detect coal gangue and the detection performance is better than the current excellent lightweight object detection models, which can meet the needs of coal gangue separation application scenarios.

[1] 国家统计局. 中华人民共和国2020年国民经济和社会发展统计公报[EB/OL]. (2021-02-28)[2021-5-27].

[2] 国家统计局. 中国统计年鉴[J]. 北京: 中国统计出版社, 2021.

[3] 罗敏, 罗国金, 伍霞, 等. 某煤矸石页岩烧结砖项目职业病危害因素与关键控制点分析[J]. 职业卫生与应急救援, 2017, 35(04):366-368.

[4] 雷建红. 煤矸石的污染危害与综合利用分析[J]. 能源与节能, 2017(04):90-91+147.

[5] 郭秀军. 煤矸石分选技术研究与应用[J]. 煤炭工程, 2017(1):74-76.

[6] 徐洁琼. 重介质选煤技术工艺与管理[J]. 煤炭技术, 2011, 30(8):137-139.

[7] 蒋林龙. X射线智能选矸技术在姚桥选煤厂的应用[J]. 煤炭加工与综合利用, 2021(07):37-39.

[8] 饶中钰,吴景涛,李明. 煤矸石图像分类方法[J]. 工矿自动化, 2020, 46(03):69-73.

[9] 徐春云. 井下煤矸石液压分选技术研究[J]. 液压气动与密封, 2014, 34(02):5-7.

[10] Zhang N, Liu C. Radiation characteristics of natural gamma-ray from coal and gangue for recognition in top coal caving[J]. Scientific reports, 2018, 8(1): 1-9.

[11] 王仁宝, 欧阳名三, 王爽. 基于DSP与改进边缘检测算法的煤矸石自动分选系统[J]. 电子技术应用, 2011 ,37(02):20-22+25.

[12] 吴开兴, 宋剑. 基于灰度共生矩阵的煤与矸石自动识别研究[J]. 煤炭工程, 2016, 48(02): 98-101.

[13] 曹现刚, 吴旭东, 王鹏, 李莹, 刘思颖, 张国祯, 夏护国. 面向煤矸分拣机器人的多机械臂协同策略[J]. 煤炭学报, 2019, 44(S2):763-774.

[14] Singh V, Singh T N, Singh V. Image processing applications for customized mining and ore classification[J]. Arabian Journal of Geosciences, 2011, 4(7):1163-1171.

[15] Chowdhury O, Tripathy D P. Design of haul road illumination system for an opencast coal mining project—a case study[J]. LEUKOS, 2014, 10(3): 133-143.

[16] 陈立. 基于小波变换的煤矸石自动分选方法[J]. 工矿自动化, 2018, 44(12):64-68.

[17] Zou H, Jia R. Visual Positioning and Recognition of Gangues Based on Scratch Feature Detection[J]. Traitement du Signal, 2019, 36(2):147-153.

[18] 王锐, 桂志国, 刘祎, 张鹏程. 基于X射线和结构光相机的煤矸石分拣方法研究[J]. 中北大学学报(自然科学版), 2021, 42(02):123-128+134.

[19] Hu F, Zhou M, Yan P, et al. Multispectral imaging: A new solution for identification of coal and gangue[J]. IEEE Access, 2019, 7:169697-169704.

[20] 田冬艳, 丁苏凡, 郭星歌. 基于图像处理的煤矸识别方法[J]. 煤炭技术, 2022, 41(03): 201-204.21

[21] 赵明辉. 一种煤矸石优化识别方法[J]. 工矿自动化, 2020, 46(07):113-116.

[22] 孙立新. 基于卷积神经网络的煤矸石识别方法研究[D]. 河北工程大学, 2020.

[23] 曹现刚, 刘思颖, 王鹏, 许罡, 吴旭东. 面向煤矸分拣机器人的煤矸识别定位系统研究[J]. 煤炭科学技术, 2022, 50(01):237-246.24

[24] Wei Y, Tian Q, Guo J, et al. Multi-vehicle detection algorithm through combining Harr and HOG features[J]. Mathematics and Computers in Simulation, 2019, 155:130-145.

[25] Chen S, Zhong S, Xue B, et al. Iterative scale-invariant feature transform for remote sensing image registration[J]. IEEE Transactions on Geoscience and Remote Sensing, 2020, 59(4):3244-3265.

[26] Zhao H, Zhan Z H, Lin Y, et al. Local binary pattern-based adaptive differential evolution for multimodal optimization problems[J]. IEEE transactions on cybernetics, 2019, 50(7):3343-3357.

[27] 杨笑, 王志章, 周子勇, 等. 基于参数优化AdaBoost算法的酸性火山岩岩性分类[J]. 石油学报, 2019, 40(4):457-467.

[28] Seeja R D, Suresh A. Deep learning based skin lesion segmentation and classification of melanoma using support vector machine (SVM)[J]. Asian Pacific journal of cancer prevention: APJCP. 2019, 20(5):1555.

[29] Le Cun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11):2278-2324.

[30] Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[C]//Advances in neural information processing systems. 2012: 1097-1105.

[31] Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition[J], Computer Science. 2014:1-14.

[32] Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. IEEE, 2015:1-9.

[33] He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE conference on computer vision and pattern recognition. 2016:770-778.

[34] Jie H, Li S, Gang S, et al. Squeeze-and-excitation networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020 42(8): 2011-2023.

[35] Iandola F N, Han S, Moskewicz M W, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size[J]. arXiv:1602.07360, 2016.

[36] Howard A G, Zhu M, Chen B, et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications[J]. arXiv:1704.04861, 2017.

[37] Sandler M , Howard A , Zhu M , et al. MobileNetV2: Inverted Residuals and Linear Bottlenecks[C]. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2018:4510-4520.

[38] Howard A, Sandler M, Chu G, et al. Searching for mobilenetv3[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. IEEE, 2019:1314-1324.

[39] Zhang X, Zhou X, Lin M, et al. Shufflenet: An extremely efficient convolutional neural network for mobile devices[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018:6848-6856.

[40] Han K, Wang Y H, Tian Q, et al. GhostNet: more features from cheap operations[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2020:1580-1589.

[41] Redmon J, Farhadi A. YOLO9000: better, faster, stronger[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition(CVPR). IEEE, 2017:7263-7271.

[42] Redmon J, Farhadi A. YOLOv3:An incremental improvement[J]. Optics Laser Technology, 2018, 23(03):44-57.

[43] Bochkovskiy A., Wang C. Y., Liao H Y M. YOLOv4: Optimal Speed and Accuracy of Object Detection [J]. arXiv: 2020, 2004:10934

[44] Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector[C]//European conference on computer vision. Springer, Cham, 2016: 21-37.

[45] Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. IEEE, 2014:580-587.

[46] Girshick R. Fast R-CNN[C]// IEEE International Conference on Computer Vision. IEEE Computer Society. IEEE, 2015:1440-1448.

[47] Ren S, He K, Girshick R, et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6):1137-1149.

[48] Law H, Deng J. Cornernet: Detecting objects as paired keypoints[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 734-750.

[49] Duan K, Bai S, Xie L, et al. Centernet: Keypoint triplets for object detection[C]//Proceedings of the IEEE/CVF international conference on computer vision. 2019: 6569-6578.

[50] Z. Tian, C. Shen, H. Chen, et al. FCOS: Fully convolutional one-stage object detection[C]// IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2019:9627-9636.

[51] Zhang S, Chi C, Yao Y, et al. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 9759-9768.

[52] Redmon J., Divvala S., Girshick R. et al. You Only Look Once: Unified, Real-Time Object Detection[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016:779-788.

[53] Lin T, Goyal P, Girshick R, He K, Dollár P. Focal Loss for Dense Object Detection[C]. 2017 IEEE International Conference on Computer Vision(ICCV). IEEE,2017:2999-3007.

[54] Everingham M, Eslami S M A, Gool L V, et al. The Pascal Visual Object Classes Challenge: A Retrospective[J]. International Journal of Computer Vision, 2015, 111(1):98-136.

[55] Lin T Y, Maire M, Belongie S, et al. Microsoft coco: Common objects in context[C] //European Conference on Computer Vision. Springer, Cham, 2014: 740-755.

[56] S. Yun, D. Han, S. Chun, et al. Cutmix: Regularization strategy to train strong classifiers with localizable features[C]// IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2019:6022-6031.