随着煤矸石分选系统对其智能化要求的逐渐提高，研究人员提出了煤矸石自动分选方法，其中包括射线法、机械洗选法、图像识别法。前两者会造成环境污染且成本过高，所以图像识别法成为了煤矸石自动分选的首选方案。该方法的难点是如何准确完成煤矸石的目标检测任务。深度学习已在各个领域展现出巨大的潜力，本文将生成对抗网络和one-stage目标检测模型应用到煤矸石目标检测中，旨在提高检测精度及速度。

目前没有公开可用的煤矸石数据集，且采集的煤矸石图像数量有限，传统的数据集扩充方法无法有效提高样本的多样性；煤矸石目标检测是实时检测任务，对检测速度要求高，two-stage目标检测模型精度较高，但速度难以满足实时检测的要求。针对以上问题，本文主要进行了以下研究工作：第一，在煤矸石数据集扩充中，对基于生成对抗网络的方法进行研究，针对原始GAN训练不稳定的问题，本文分别采用WGAN和DCGAN进行数据集扩充，并进行对比实验。第二，one-stage目标检测模型SSD在保持高准确率的同时，也能保证高检测速度，但其主干网络VGG-16参数众多，本文采用SSD-MobileNet作为煤矸石检测模型，并对其进行改进，在进行预测时去掉19×19的浅层特征层，使先验框的数量由原来的1917减少到834。

实验结果表明，DCGAN生成图像的FID分数明显低于WGAN生成图像的FID分数，说明其生成的图像质量优于WGAN生成的图像；基于DCGAN的数据集扩充相较于传统方法，明显提高了目标检测模型的mAP值。改进后的SSD-MobileNet模型mAP值仅下降了0.79%，且单幅图像的平均检测时间由原来的0.06s减少到0.03s，在保证检测精度的同时，有效减少了检测时间。

With the gradual improvement of the intelligent requirements in coal and gangue sorting system, researchers have proposed automatic coal and gangue separating methods, including ray method, mechanical washing method, and image recognition method. The former two will cause environmental pollution and cost too much, so the automatic separating of image-based coal and gangue recognition has become the first choice of researchers. The difficulty of this method is how to accurately complete the object detection task of coal and gangue. Deep learning has shown great potential in various fields, this paper applies generative Adversarial network and one-stage object detection model to coal and gangue object detection, aiming to improve detection accuracy and speed.

At present, there is no publicly available coal and gangue data set, and the number of collected coal gangue images is limited. The traditional data set expansion method cannot effectively improve the diversity of samples.Coal and gangue object detection is a real-time detection task, which requires high detection speed. The two-stage object detection model has high accuracy, but the speed is difficult to meet the requirements of real-time detection. In response to the above problems, this paper mainly conducts the following research work: First, in the expansion of coal and gangue data set, the method based on generative adversarial network is studied, aiming at the problem of the instability of the original GAN training, this paper uses WGAN and DCGAN to expand the data set, and conduct a comparative experiment. Second, the one-stage object detection model SSD can ensure high detection speed while maintaining high accuracy. However, it’s backbone network VGG-16 has many parameters. This article uses SSD-MobileNet as the coal gangue detection model and performs The improvement is that the 19×19 shallow feature layer is removed when making predictions, so that the number of a prior bounding box is reduced from 1917 to 834.

Experimental results show that the FID score of images generated by DCGAN is significantly lower than that of WGAN generated images, indicating that the quality of images generated is better than that of images generated by WGAN, compared with traditional methods, DCGAN-based data set expansion significantly improves the mAP value of the object detection model. The mAP value of the improved SSD-MobileNet model is only reduced by 0.79%, and the average detection time of a single image is reduced from 0.06s to 0.03s, which effectively reduces the detection time while ensuring the detection accuracy.

[1] 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.

[2] 袁亮.煤炭精准开采科学构想[J].煤炭学报,2017,42(1):1-6.

[3] 石焕,程宏志,刘万超.我国选煤技术现状及发展趋势[J].煤炭科学技术,2016,44(6):169-174.

[4] 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.

[5] Yang Z, Wang Z, Yan M. Performance Analysis of Natural γ-Ray Coal Seam Thickness Sensor and Its Application in Automatic Adjustment of Shearer’s Arms[J]. Journal of Electrical and Computer Engineering, 2020, 2020.

[6] 何晓明.基于X射线的煤与矸石自动识别方法研究[D].沈阳:东北大学,2013.

[7] 葛学海,白云飞,陈鹏,等.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.

[8] 王卫东,张晨,马中良,等.煤矸光电分选系统设计[J].工矿自动化,2013,39(12):5-8.

[9] 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.

[10] 谭春超,杨洁明.煤与矸石图像灰度信息和纹理特征的提取研究[J].工矿自动化, 2017,43(04):27-31.

[11] 马宪民.一种基于小波变换的煤矸石图像边缘检测方法[J].仪器仪表学报, 2006(S3):2130-2131.

[12] 张世杰,陈泽华.基于梯度下降算法的煤矸石分割识别方法[J].煤矿机械, 2014,35(08):128-130.

[13] 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(01):118-123.

[14] 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.

[15] 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.

[16] 陈雪梅,张晞,徐莉莉,等.煤与矸石分形维数的差异研究[J].煤炭科学技术, 2017,45(07):196-199.

[17] 潘越,曾哲,张恩瑜.基于MATLAB和图像灰度值对X射线探测煤矸识别的研究[J].煤炭技术,2017,036(011):307-309.

[18] 余乐,郑力新,杜永兆,等.采用部分灰度压缩扩阶共生矩阵的煤和煤矸石图像识别[J].华侨大学学报(自然科学版),2018,39(06):906-912.

[19] 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.

[20] 郭永存,于中山,卢熠昌.基于PSO优化NP-FSVM的煤矸光电智能分选技术研究[J].煤炭科学技术,2019,47(04):18-24.

[21] 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,2020,039(001):112-116.

[22] 曹现刚,薛祯也.基于迁移学习的GoogLenet煤矸石图像识别[J].软件导刊,2019,18(12):183-186.

[23] Zarie M, Jahedsaravani A, Massinaei M. Flotation froth image classification using convolutional neural networks[J]. Minerals Engineering, 2020, 155: 106443.

[24] 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.

[25] Krizhevsky A, Sutskever I, Hinton G. ImageNet Classification with Deep Convolutional Neural Networks[J]. Advances in neural information processing systems, 2012, 25: 1097-1105.

[26] Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition[J]. arXiv preprint arXiv:1409.1556, 2014.

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

[28] 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.

[29] 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. 2014: 580-587.

[30] Girshick R. Fast r-cnn[C]//Proceedings of the IEEE international conference on computer vision. 2015: 1440-1448.

[31] Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks[J]. arXiv preprint arXiv:1506.01497, 2015.

[32] Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real-time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 779-788.

[33] Liu W, Anguelov D, Erhan D, et al. SSD: Single Shot MultiBox Detector[C]// European Conference on Computer Vision. Springer, Cham, 2016:21-37.

[34] Goodfellow I J, Pouget-Abadie J, Mirza M, et al. Generative Adversarial Networks[J]. IEEE Advances in Neural Information Processing Systems, 2014,3(32):2672-2680.

[35] Rezende D J, Mohamed S, Wierstra D. Stochastic backpropagation and approximate inference in deep generative models[C]//International conference on machine learning. PMLR, 2014: 1278-1286.

[36] Mao X, Li Q, Xie H, et al. Least squares generative adversarial networks[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2794-2802.

[37] Lotter W, Kreiman G, Cox D. Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning[J].arXiv: Machine Learning, 2016,15(27):895-906.

[38] Mirza M, Osindero S. Conditional generative adversarial nets[J]. arXiv preprint arXiv:1411.1784, 2014.

[39] Radford A, Metz L, Chintala S. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks[J]. Computer Science, 2015,16(23):718-

726.

[40] Arjovsky M, Chintala S, Bottou L. Wasserstein GAN[J]. IEEE Computer Vision,2017,34

(22):2672-2680.

[41] Gulrajani I, Ahmed F, Arjovsky M. Improved Training of Wasserstein GANs[J]. IEEE Advances in Neural Information Processing Systems, 2017,36(18): 5769-5779

[42] Zhu J Y, Park T, Isola P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2223-2232.

[43] Isola P, Zhu J Y, Zhou T, et al. Image-to-image translation with conditional adversarial networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1125-1134.

[44] 肖旭.基于深度学习的图像语义分割研究[D].南昌:南昌航空大学,2017.

[45] Miyato T, Koyama M. cGANs with Projection Discriminator[C]// International Conference on Learning Representations. 2018:276-284.

[46] 陈锟,乔沁,宋志坚.生成对抗网络在医学图像处理中的应用[J].生命科学仪器,2018,016(004):71-80,91.

[47] Perez L, Wang J. The Effectiveness of Data Augmentation in Image Classification using Deep Learning[J]. IEEE Computer Vision and Pattern Recognition, 2017,11(24):385-393.

[48] Shrivastava A, Pfister T, Tuzel O, et al. Learning from simulated and unsupervised images through adversarial training[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 2107-2116.

[49] Madani A, Moradi M, Karargyris A, et al. Chest x-ray generation and data augmentation for cardiovascular abnormality classification[C]//Medical Imaging 2018: Image Processing. International Society for Optics and Photonics, 2018, 10574: 105741M.

[50] Galbusera F, Niemeyer F, Seyfried M, et al. Exploring the potential of generative adversarial networks for synthesizing radiological images of the spine to be used in in silico trials[J]. Frontiers in bioengineering and biotechnology, 2018, 6-53.

[51] Lucic M, Kurach K, Michalski M, et al. Are GANs Created Equal? A Large-Scale Study[J]. IEEE Advances in Neural Information Processing Systems,. 2018,25(08): 700-709.

[52] 王坤峰,苟超,段艳杰,等.生成式对抗网络GAN的研究进展与展望[J]. 自动化学报, 2017,043(003):321-332.