| 论文中文题名： | 基于YOLO V5卷积网络的煤矸识别系统的构建与优化 |
| 姓名： | |
| 学号： | 20213226062 |
| 保密级别： | 保密（2年后开放） |
| 论文语种： | chi |
| 学科代码： | 085700 |
| 学科名称： | 工学 - 资源与环境 |
| 学生类型： | 硕士 |
| 学位级别： | 工程硕士 |
| 学位年度： | 2023 |
| 培养单位： | 西安科技大学 |
| 院系： | |
| 专业： | |
| 研究方向： | 煤矸分选 |
| 第一导师姓名： | |
| 第一导师单位： | |
| 第二导师姓名： | |
| 论文提交日期： | 2023-06-26 |
| 论文答辩日期： | 2023-05-31 |
| 论文外文题名： | Construction and optimization of coal gangue recognition system based on YOLO V5 convolutional network |
| 论文中文关键词： | |
| 论文外文关键词： | Coal gangue identification ; Near infrared light source ; Image pre-processing ; YOLOV5 model |
| 论文中文摘要： |
煤炭资源在我国的能源当中，依然是不可缺少的一种重要来源。选煤厂在开采之后需要对原煤进行选矸处理，目前大多采用的是人工分选法，该方法不仅效率低下而且工人劳动强度较大。近年来，计算机视觉技术已在工业上得到了广泛的应用。用于煤矸识别的小目标检测依然无法满足现实的需要。小目标煤矸识别具有分辨率低，分布稠密、尺度不均衡、容易受到背景的干扰等困难，传统方法容易造成误匹配现象。基于此，本文重点研究在不同场景下如何实现对小目标的煤与矸石的有效检测，并给出相应的解决方案。本文以YOLO V5s模型为基本框架进行建模，针对煤矸特点对模型进行了优化，以提升对煤和矸石检测效果。具体工作如下： （1）数据集采集方面，设计搭建了煤和矸石数据采集装置，在装置内设置了补光器，减少外界光源对数据采集的干扰，使用白光摄像头与近红外摄像头采集两种不同光源的数据集，剔除白光源数据集的信息冗余，提高模型的抗干扰能力。考虑到实际的选矸情况，进一步将煤与矸石的粒度分成三类，测试模型的分类能力。 （2）采用灰度化、降噪、锐化以及目标标注四个步骤完成图像的预处理。灰度化使用加权平均值法，降噪使用中值滤波，对于单个噪声点、椒盐噪声的消除十分有效，锐化使用的是拉普拉斯算子，能够最大程度的保留目标轮廓，而图像标注使用的是labelimg，效果良好。 （3）鉴于基础的YOLO V5s模型对于小目标的分辨率低，对煤和矸石判识容易造成漏检，本文在YOLO V5s的Head层增加了一个检测头，提高了模型识别小目标煤和矸石能力及多尺度目标检测性能。同时，在Neck 端结合FPN+PAN网络结构的特点，对YOLO V5s的特征融合结构加以优化，新增两条横向跨尺度连接路径，丰富了煤和矸石等小目标的特征信息。 （4）模型测试分析表明：近红外光源数据集的识别精度高于白光数据集，其中白光数据集在优化前对煤的平均识别精度为87.2%，对矸石的平均识别精度为86.5%，优化后对煤的平均识别精度为88.6%，对矸石的平均识别精度为87.6%，近红外光源数据集在优化前对煤的平均识别精度为93.2%，对矸石的平均识别精度为92.8%，优化后对煤的平均识别精度为94.5%，对矸石的平均识别精度为94.8%。不同粒度条件下，100～200 mm粒度集的识别精度最高，其中白光数据集优化前对煤的识别精度为89.4%，对矸石的识别精度为88.7%，优化后对煤的识别精度为90.3%，对矸石的识别精度为89.8%，近红外数据集优化前对煤的识别精度为96.8%，对矸石的识别精度为95.4%，优化后对煤的识别精度为98.1%，对矸石的识别精度为97.5%；200～300 mm粒度集的识别精度次之，其中白光数据集优化前对煤的识别精度为87.8%，对矸石的识别精度为86.9%，优化后对煤的识别精度为88.9%，对矸石的识别精度为87.6%，近红外数据集优化前对煤的识别精度为93.4%，对矸石的识别精度为92.6%，优化后对煤的识别精度为94.3%，对矸石的识别精度为95.2%； 50～100 mm粒度集识别精度最差，其中白光数据集优化前对煤的识别精度为84.5%，对矸石的识别精度为83.9%，优化后对煤的识别精度为86.7%，对矸石的识别精度为85.6%，近红外数据集优化前对煤的识别精度为89.4%，对矸石的识别精度为90.3%，优化后对煤的识别精度为91.2%，对矸石的识别精度为91.7%。优化后的模型整体识别精度有了一定的提升，模型优化方案可行。 |
| 论文外文摘要： |
Coal resources in China's energy, is still an important source of indispensable. Coal processing plant needs to select gangue for raw coal after mining, at present, most of them adopt manual sorting method, which is not only inefficient but also labor-intensive for workers. In recent years, computer vision technology has been widely used in industry. But the small target detection applied to the gangue identification still can not meet the needs of reality. Small target gangue identification has the characteristics of low resolution, dense distribution, uneven scale, easy to be disturbed by the background and other difficulties, and the traditional method is easy to cause the phenomenon of mis-matching. Based on this, this paper focuses on how to achieve effective detection of coal and gangue of small targets in different scenarios and gives corresponding solutions. In this paper, the YOLO V5s model is modeled as the basic framework, and the model is optimized for the characteristics of coal and gangue to improve the detection of coal and gangue. The specific work is as follows: (1) In terms of data set acquisition, the coal and gangue data acquisition device is designed and built, and a light filler is set in the device to reduce the interference of external light source on data acquisition, and the data sets of two different light sources are collected using white light camera and near infrared camera to eliminate the information redundancy of white light source data set and improve the anti-interference ability of the model. Considering the actual gangue selection situation, the particle size of coal and gangue is also divided into three categories to test the classification ability of the model. (2) In this paper, four steps of grayscale, noise reduction, sharpening and target labeling are used to complete the image and pre-processing. Grayscale using the weighted average method, noise reduction using the median filter, for a single noise point, pepper noise elimination is very effective, sharpening using the Laplace operator, to maximize the preservation of the target outline, and image labeling using labelimg, good results. (3) In view of the low resolution of the basic YOLO V5s model for small targets, coal and gangue recognition is easy to cause missed detection, this paper adds a P2 detection head in the Head layer of YOLO V5s, which improves the model recognition of small targets coal and gangue capability and multi-scale target detection performance At the same time, the features of FPN+PAN network structure are combined at the Neck side, and the At the same time, the feature fusion structure of YOLO V5s is optimized by combining the characteristics of FPN+PAN network structure at the Neck side, and two new horizontal cross-scale connection paths are added to enrich the feature information of small targets such as coal and gangue. (4) Model test analysis shows that the recognition accuracy of NIR light source data set is higher than that of white light data set, in which the average recognition accuracy of white light data set is 87.2% for coal and 86.5% for gangue before optimization, and 88.6% for coal and 87.6% for gangue after optimization, and the average recognition accuracy of NIR light source data set is 93.2% for coal and 92.8% for gangue before optimization. The average recognition accuracy of coal before optimization is 93.2%, the average recognition accuracy of gangue is 92.8%, the average recognition accuracy of coal after optimization is 94.5%, the average recognition accuracy of gangue is 94.8%. Under different particle size conditions, the recognition accuracy of 100-200 mm particle size set is the highest, in which the recognition accuracy of white light data set before optimization is 89.4% for coal and 88.7% for gangue, after optimization, the recognition accuracy of coal is 90.3% and 89.8% for gangue, the recognition accuracy of NIR data set before optimization is 96.8% for coal and 95.4% for gangue. 95.4%, after the optimization of the coal recognition accuracy of 98.1%, the gangue recognition accuracy of 97.5%; 200-300 mm size set of recognition accuracy is the second, which is the white light data set before optimization of the coal recognition accuracy of 87.8%, the gangue recognition accuracy of 86.9%, after the optimization of the coal recognition accuracy of 88.9%, the gangue recognition accuracy of 87.6%. The recognition accuracy of coal before optimization is 93.4%, the recognition accuracy of gangue is 92.6%, the recognition accuracy of coal after optimization is 94.3%, the recognition accuracy of gangue is 95.2%; the recognition accuracy of 50-100 mm size set is the worst, including the recognition accuracy of coal before optimization is 84.5%, the recognition accuracy of gangue is 83.9%, after optimization The recognition accuracy of coal is 86.7%, the recognition accuracy of gangue is 85.6%, the recognition accuracy of coal is 89.4%, the recognition accuracy of gangue is 90.3% before the optimization of NIR data set, the recognition accuracy of coal is 91.2%, the recognition accuracy of gangue is 91.7% after the optimization. The overall recognition accuracy of the optimized model has been improved to a certain extent, and the model optimization scheme is feasible. |
| 参考文献： |
[1] 杨昊睿, 宁树正, 丁恋, 等. 新时期我国煤炭产业现状及对策研究 [J]. 中国煤炭地质, 2021, 33(S1): 44-48. [3] 我国第二个年产量超两亿吨煤炭企业诞生[J]. 中国煤炭工业, 2020, (2): 15. [4] 谭春超. 基于图像处理技术的煤矸识别与分选技术研究 [D].太原理工大学, 2017. [5] 郑欢. 中国煤炭产量峰值与煤炭资源可持续利用问题研究 [D]. 西南财经大学, 2014. [6] 姚国华,高知睿,李小赛. 中国煤炭资源承载能力评价 [J]. 中国矿业, 2020, 29(8): 1-7. [10] 段雍. 基于图像的煤矸识别和定位方法研究与实现 [D]. 西安科技大学, 2020. [12] 张明. 基于压缩感知理论的杂草种子分类识别研究 [D].西北农林科技大学, 2012. [15] Klempner A. Optimal Thresholds in Ra-diometric Coal Separation [J], 1982, 18(6): 550-553. [16] 张万枝,王增才. 基于视觉技术的煤岩特征分析与识别 [J], 2014, 33(10): 272-274. [17] 杨庆福. 国外光电选矿及其进展 [J], 1986: 50-54. [19] 韩成石, 董长双, 周西军, 等. 煤和矸石 γ 射线分选系统的研究 [J], 1997, 2: 157-161. [20] 孔力, 李红. 基于双能 γ 射线的煤,矸石区域分割识别方法 [J]. 华中科技大学学报(自然科学版), 1998, (1): 39-40. [21] 程葳, 孔力, 袁树风, 等. 基于模糊神经网络的煤矸石模式识别方法研究 [J]. 华中科技大学学报(自然科学版), 1999, 27(6): 30-32. [22] 徐琦, 孔力, 程晶晶. 基于生态遗传算法的煤矸石自适应模糊模式识别 [J]. 华中科技大学学报（自然科学版）, 2003, 31(12): 22-24. [23] 刘文中, 孔力. 煤矸石自适应神经元网络识别方法研究 [J]. 选煤技术, 2002, (3): 22-24. [24] 李文斌, 杨剑瑜, 文建国. 光电选矸识别系统的研制与测试 [J]. 仪器仪表学报, 2001, 22(3): 265-268. [25] 刘富强, 钱建生, 王新红, 等. 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, 25(5): 534-537. [26] 马宪民, 田红, 龚尚福. 煤矸石在线识别与自动分选系统的研究 [C]//中国智能自动化会议, 2001: 66-68. [27] 任云鹏. 基于 CAN 控制器的煤矸石自动分选系统的设计 [J]. 工矿自动化, 2005, (3): 1-3. [28] 李祥瑞. 机器视觉研究进展及工业应用综述_李祥瑞 [J]. 数字通信世界, 2021, (11): 79-80, 146. [29] 马宪民, 宋晓茹. 基于 ARM 核和 CAN 总线的煤矸石分选系统 [J]. 仪器仪表学报, 2005, 26(z1): 305-307. [30] 臧洁. 小波分析在煤矸石图像识别 ARM 分选控制系统中的应用 [D]. 西安:西安科技大学, 2008. [33] 廖阳阳. 基于 BP 网络和图像处理的煤矸石的动态识别[J]. 工业控制计算机, 2015, (7): 119-120. [37] 曹现刚, 费佳浩, 王鹏, 等. 基于多机械臂协同的煤矸分拣方法研究 [J]. 煤炭科学技术, 2019, 47(4): 7-12. [39] 李炳臻, 姜文志, 顾佼佼, 等. 基于卷积神经网络的目标检测算法综述 [J]. 计算机与数字工程, 2022, 50(5): 1010-1017. [43] 张硕. 近红外复杂场景成像仿真关键技术研究 [D]. 北京理工大学, 2018. [44] 陈宇. 近红外单像素成像系统及算法研究 [D]. 哈尔滨工业大学, 2019. [45] 陈海峰, 丁丽丽. 二值化图像的灰度处理算法研究_陈海峰 [J]. 电脑与电信, 2019, (7): 34-38. [46] 庞尚钟. 基于机器视觉的煤矸目标识别与定位方法研究 [D]. 太原理工大学, 2020. [51] 徐少平, 李芬, 陈孝国, 等. 一种利用改进深度图像先验构建的图像降噪模型 [J]. 电子学报, 2022, 50(7): 1573-1578. [53] 王野, 李钦奉, 齐继阳. 基于改进自适应中值滤波的图像降噪方法 [J]. 舰船电子工程, 2022, 42(6): 112-115. [55] 洪亮, 刘洁, 王心茹. 图像锐化在Photoshop和Matlab中的应用 [J]. 广东印刷, 2023, (1): 15-16. [59] 邵延华, 张铎, 楚红雨, 等. 基于深度学习的YOLO目标检测综述 [J]. 电子与信息学报, 2022, 44(10): 3697-3708. [60] 赵天宇. 基于YOLO-V4的输电线路多故障检测与跟踪 [D]. 辽宁工程技术大学, 2021. [61] 耿创, 宋品德, 曹立佳. YOLO算法在目标检测中的研究进展 [J]. 兵器装备工程学报, 2022, 43(9): 162-173. [62] 吴士好, 刘艳华. 基于MATLAB/GUI的文字图片识别系统设计 [J]. 电子产品世界, 2023, 30(4): 18-19, 42. [63] 冯桂尔. 基于Python的GUI(Tkinter)实例开发 [J]. 信息与电脑(理论版), 2023, 35(1): 175-178 |
| 中图分类号： | TD94 |
| 开放日期： | 2025-06-26 |
谷歌
火狐
