煤矸石分选是煤炭开采过程中重要的环节之一，矸石含量过高会引起电厂燃效效率低下，造成严重的环境污染，通过实地考察调研发现目前大多数矿场仍采用人工选矸，选煤仓里的灰尘大、光线暗，长时间高负荷的工作不仅会对工人的身体造成危害，还会影响分选效率。因此，本课题设计了一种基于YOLOv5s的煤矸石检测算法代替人工选矸；目前，基于深度学习方法针对煤矸石检测的任务仍然存在以下问题：1、在实验室采集煤矸石图像数据并进行测试，而实验室的光线、灰尘等环境因素与实际应用场景差别较大；2、目前尚不存在一个大规模、公开可用的煤矸石图像数据集；3、小块煤与矸石的识别存在漏检与误检。

本课题首先从构建煤矸石数据集入手，建立贴合实际应用场景的煤矸石图像数据集。通过在选煤仓实地采集煤矸石图像，采用一种基于深度卷积生成对抗网络(Deep Convolutional Generative Adversarial Networks，DCGAN)的图像扩充方法来改善目前煤和矸石图像数量不充足的问题；但基础DCGAN模型生成的图像质量差、缺乏多样性，无法满足本课题的需求；通过理论与实验分析，对原始DCGAN模型做出改进。首先，在生成器网络中添加一层反卷积操作用以增大输出图像的分辨率；其次，加入特征融合模块提高生成图像的多样性；同时，损失函数采用改进的Wasserstein距离，判断生成数据与真实数据的分布情况，优化损失函数，提高生成煤矸石图像的质量。最终通过实验表明，改进后的DCGAN模型能够生成满足本课题所需的煤矸石图像需求，并且可以有效提高目标检测模型的泛化能力。

在完成煤矸石图像数据集扩充后，本课题选取YOLOv5s模型作为基础目标检测模型。通过实验发现，YOLOv5s能够满足实时检测的需求，但是对于小块煤和矸石的识别存在漏检与误检。针对此问题，本课题对基础YOLOv5s模型做出以下改进：1、在YOLOv5s的主干网络中，添加一个针对小尺度目标检测的预测头；2、采用K-means聚类算法生成针对本课题数据集的初始锚框参数；3、在主干网络的每一个普通卷积层后添加SE(Squeeze-and-Excitation)注意力模块；4、使用改进的快速归一化特征融合方法（Fast-Concat，F-Concat）加强特征提取。基础YOLOv5s模型的检测精度能够达到97.89%，而改进后的YOLOv5s模型在单张图像推理速度上牺牲了0.015s的情况下，检测精度与基础模型相比提高了1.01%，并且能够实现对小块煤矸石的准确检测，同时，仍然能够满足实时检测的需求。

Coal gangue separation is one of the important links in the process of coal mining. High gangue content will lead to low combustion efficiency of power plants and serious environmental pollution. Through field investigation and investigation, it is found that at present, most mines still use manual gangue separation. The dust in the coal preparation bunker is large and the light is dark. Long time and high load work will not only harm the workers' health, but also affect the separation efficiency. Therefore, this paper designs a coal gangue detection algorithm based on yolov5s instead of manual gangue selection; At present, there are still the following problems in the task of coal gangue detection based on the deep learning method: 1. The image data of coal gangue are collected and tested in the laboratory, but the environmental factors such as light and dust in the laboratory are quite different from the actual application scenarios; 2. At present, there is no large-scale and publicly available coal gangue image data set; 3. There are missed detection and false detection in the identification of small coal and gangue.

Firstly, starting with the construction of coal gangue data set, the coal gangue image data set fitting the actual application scene is established. By collecting the coal gangue images in the coal preparation bunker, an image expansion method based on Deep Convolutional Generative Adversarial Networks (DCGAN) is adopted to improve the problem of insufficient coal and gangue images at present; However, the image quality generated by the basic DCGAN model is poor and lack of diversity, which can not meet the needs of this subject; Through theoretical and experimental analysis, the original DCGAN model is improved. Firstly, a layer of deconvolution operation is added to the generator network to increase the resolution of the output image, and then the feature fusion module is added to improve the diversity of the generated image; At the same time, the loss function uses the improved Wasserstein distance to judge the distribution of generated data and real data, optimize the loss function and improve the quality of coal gangue image. Finally, experiments show that the improved DCGAN model can generate coal gangue images to meet the needs of this subject, and can effectively improve the generalization ability of the target detection model.

After completing the expansion of coal gangue image data set, this subject selects YOLOv5s model as the basic target detection model. Through experiments, it is found that YOLOv5s can meet the needs of real-time detection, but there are missed detection and false detection in the identification of small coal and gangue. To solve this problem, this topic makes the following improvements to the basic YOLOv5s model: 1. Add a prediction header for small-scale target detection in the backbone network of YOLOv5s; 2. The K-means clustering algorithm is used to generate the initial anchor frame parameters for the data set of this subject; 3. Add SE attention module after each ordinary volume layer of the backbone network; 4. The basic Concat feature fusion method is replaced by F-Concat fast normalized feature fusion method. The improved YOLOv5s model improves the detection accuracy by 1.01% compared with the basic model at the expense of a small reasoning speed, but it can realize the accurate detection of small coal gangue and still meet the needs of real-time detection.

[1] 谢和平,吴立新,郑德志.2025年中国能源消费及煤炭需求预测[J].煤炭学报2019,44(07):1949-1960.

[2] 马骏,郁钟铭,舒仕海,等.煤矸石对矿区的环境危害及治理措施[J].煤炭工程,2015,47(10):70-73.

[3] 袁亮.煤炭精准开采科学构想[J].煤炭学报,2017,42(01):1-7.

[4] 郭彦霞,张圆圆,程芳琴.煤矸石综合利用的产业化及其展望[J].化工学报2018,65(7):2443-2453.

[5] 程健,王东伟,杨凌凯,等.一种改进的高斯混合模 型煤矸石视频检测方法[ J]. 中南大学学报:自然科 学版,2018,49(1):118-123.

[6] 张振红.我国干法选煤技术发展现状与应用前景[J].选煤技术,2019(01):43-47+52.

[7] 王卫东,张晨,马中良,曹春晓,王彤彤,张璐琪,任春醒,梁静.煤矸光电分选系统设计[J].工矿自动化,2013,39(12):5-8.

[8] 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.

[9] 孔力,李红,徐恕宏,徐琦.双能γ射线透射法煤矸石在线识别与分选系统[J].华中理工大学学报,1997(10):108-109+113.

[10] 孟繁明,郭科伟.基于双能γ射线透射技术的煤矸石识别与分选系统研究[J].黑龙江科技信息,2013(36):83.

[11] Wang W, Lv Z, Lu H. Research on methods to differentiate coal and gangue using image processing and a support vector machine[J].International Journal of 4Coal Preparation and Utilization, 2018: 1-14.

[12] 申利飞,王文清,白林绪.基于加权方差的煤矸石X射线图像分形维数最优估计算法[J].计算技术与自动化,2020,39(01):112-116.

[13] Zhao Y D , He X . Recognition of Coal and Gangue Based on X-Ray[J]. Applied Mechanics and Materials, 2013, 275-277:2350-2353.

[14] 葛学海,白云飞,陈鹏,张立功.煤与矸石分离系统中X射线探测器的设计[J].选煤技术,2017(06):64-67.

[15] 杨慧刚,乔志敏.基于X射线和机器视觉的煤与矸石分选系统设计[J].工矿自动化,2017,43(03):85-89.

[16] 杨慧刚,乔志敏,高绘彦,刘宇,赵一丁.煤与矸石分选系统设计[J].工矿自动化,2018,44(08):91-95.

[17] 刘富强,钱建生,王新红,宋金铃.基于图像处理与识别技术的煤矿矸石自动分选[J].煤炭学报,2000(05):534-537.

[18] Hobson D M, Carter R M, Yan Y, et al. Differentiation between coal and stone through image analysis of texture features[C]. 2007 IEEE International Workshop on Imaging Systems and Techniques, 2007: 1-4.

[19] Tripathy D P, Reddy K G R. Novel methods for separation of gangue from limestone and coal using multispectral and joint color-texture features[J]. Journal of The Institution of Engineers (India): Series D, 2017, 98(1): 109-117.

[20] 丁泽海,薛斌,窦东阳.图像处理在煤矸石分选系统中的应用[J].煤矿机械,2017,38(03):173-175.

[21] 梁澈. 粗糙集理论在煤矸石图像识别技术中的应用[D].西安科技大学,2020.

[22] 于国防,邹士威,秦聪.图像灰度信息在煤矸石自动分选中的应用研究[J].工矿自动化,2012,38(02):36-39.

[23] 王家臣,李良晖,杨胜利.不同照度下煤矸图像灰度及纹理特征提取的实验研究[J].煤炭学报,2018,43(11):3051-3061.

[24] 吴景涛. 机器视觉技术在煤矸石智能识别中的应用研究[D].中国矿业大学,2019.

[25] 佘杰. 基于图像的煤岩识别方法研究[D].中国矿业大学（北京）,2014.

[26] 洪惠超. 基于机器视觉的煤矸石分选算法的研究[D].华侨大学,2018.

[27] YD Zhao, MF Sun. Image Processing and Recognition System Based on Da Vinci Technology for Coal and Gangue[J]. Applied Mechanics & Materials, 2011 , 130-134 : 2107-2110.

[28] YU Le. A New Method for Image Recognition of Coal and Coal Gangue[J]. Modern Computer, 2017(17):66-70.

[29] R Wang, Z Liang. Automatic Separation System of Coal Gangue Based on DSP and Digital Image Processing[J].Photonics & Optoelectronics , 2011 :1-3.

[30] Su Lingling, Cao Xiangang, Ma Hongwei, et al. Research on coal gangue identification by using convolutional neural network [C].The 2nd IEEE Advanced Information Management, Communicates, Electronic and Automation Control Conference, Xi’an, 2018:810-814.

[31] 单鹏飞,孙浩强,来兴平,朱兴攀,杨建辉,高健铭.基于改进Faster R-CNN的综放煤矸混合放出状态识别方法[J].煤炭学报,2022,47(03):1382-1394.

[32] 郜亚松. 基于深度学习的煤矸石在线识别定位研究与设计[D].河北工程大学,2020.

[33] 来文豪,周孟然,胡锋,卞凯,宋红萍.基于多光谱成像和改进YOLO v4的煤矸石检测[J].光学学报,2020,40(24):72-80.

[34] Goodfellow I,Pouget-Abadie J,Mirza M,et al.Generative adversarial nets//Proceedings of the Advances in Neural Information Processing Systems.Montrel,Canada,2014:2672-2680.

[35] 孙钰,李林燕,叶子寒,等.多层次结构生成对抗网络的文本生成图像方法[J].计算机应用,2019,39(11):3204-3209.

[36] WANG W, HUANG Q, YOU S, et al. Shape inpainting using 3d generative adversarial network and recurrent convolutional networks[C]//Proceedings of the IEEE international conference on computer vision, 2017: 2298- 2306.

[37] MA J, YU W, CHEN C, et al. Pan-GAN: An unsupervised pan-sharpening method for remote sensing image fusion[J]. Information Fusion, 2020, 62: 110-120.

[38] ZHANG H, XU T, LI H, et al. Stackgan: Text to photo-realistic image synthesis with stacked generative ad-versarial networks[C]//Proceedings of the IEEE international conference on computer vision. 2017: 59.

[39] MIRZA M, OSINDERO S. Conditional generative adversarial nets[J]. arXiv preprint arXiv:1411.1784, 2014.

[40] JAISWAL A, ABDALMAGEED W, WU Y, et al. Bidirectional conditional generative adversarial networks[C]//Asian Conference on Computer Vision. 2018: 216-232.

[41] PERARNAU G, VAN DE WEIJER J, RADUCANU B, et al. Invertible conditional GANs for image editing[J]. arXiv preprXiv:1611.06355, 2016.

[42] Radford A,Luke M,Soumith C.Unsupervised representation learning with deep convolutional generative adversarial networks.arXiv preprint arXiv:1511.06434,2015.

[43] ARJOVSKY M, CHINTALA S, BOTTOU L. WassersteinGAN[J]. arXiv preprint arXiv:1701.07875, 2017.

[44] GULRAJANI I, AHMED F, ARJOVSKY M, et al. Improved training of WassersteinGANs[C]//Advances in Neural Information Processing Systems. 2017: 5767-5777.

[45] SANDFORT V, YAN K, PICKHARDT P J, et al. Data augmentation using generative adversarial networks (CycleGAN) to improve generalizability in CT segmentation tasks[J]. Scientific reports, 2019, 9(1): 1-9.

[46] HAN C, RUNDO L, MURAO K, et al. MADGAN: unsupervised Medical Anomaly Detection GAN using multiple adjacent brain MRI slice reconstruction[J]. arXiv preprint arXiv:2007.13559, 2020.

[47] TANG Y, CAI J, LU L, et al. CT image enhancement using stacked generative adversarial networks and transfer learning for lesion segmentation improvement[C]//International Workshop on Machine Learning in Medical Imaging. Springer, Cham, 2018: 46-54.

[48] Krizhevsky A,Sutskever I,Hinton GE.ImageNet Classification with Deep Convolutional Neural Networks[C].Washington.International Conferenceon Neural Information Processing Systems.Curran Associates Inc.2012:1097-1105.

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

[50] TAN M,PANG R,LE Q V.Efficientdet:Scalable and Efficient Object Detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.[S.I.]:IEEE,2020:10781-10790.

[51] 梅国新,姚庆华,陈瑶,刘华.一种边缘环境下基于EfficientDet的施工人员安全帽检测方法[J].数字通信世界,2020(09):77-78.

[52] 何江,张科星.基于机器视觉的AlexNet网络煤矸石检测系统[J].煤炭技术,2022,41(03):205-208.

[53] 张晓峰,吴刚.基于生成对抗网络的数据增强方法[J].计算机系统应用,2019,28(10):201-206.

[54] 任盼飞,宋晓茹,陈超波,高嵩.基于改进DCGAN的水下图像生成方法研究[J].自动化与仪表,2021,36(02):44-49.

[55] TOROUS J,LIPSCHITZ J,NG M,et al.Dropout rates in clinical trials of smartphone apps for depressive symptoms:a systematic review and meta-analysis[J].Journal of Affective Disorders,2020,263:413-419.

[56] 朱登柯,侯兴松,要晓迪.基于EfficientDet网络的输电线路故障检测[J].国外电子测量技术,2021,v.40;No.319(06):144-151.