煤尘治理是煤矿安全的基础，煤尘的粒度分布(Particle Size Distribution, PSD)是检验除尘效果的支撑信息。通过手工筛分、仪器检测等方法能够精准获取煤尘粒度分布，但设备昂贵，人工成本高，难以大规模应用于煤矿场景。因此，本文在分析现有粒子分割和统计理论的基础上，对基于卷积神经网络的煤尘粒度检测方法进行研究。

首先，针对煤尘图像中颗粒数目多、轮廓不规则导致分割精度低的问题，提出M-SegNet网络模型。该模型在编码结构中，通过深度卷积和点向卷积提取煤尘特征，在运算次数和参数上减少模型运算负担。引入注意力机制获取空间和通道的权重矩阵，整合全局语义信息，优化特征提取能力。解码阶段，记录编码位置并建立位置索引，以非线性上采样的方式恢复特征图尺寸。在不同尺寸的煤尘图像中进行实验，拟合模型的最优参数，有效分割了煤尘图像的粒子。

其次，针对煤尘图像中粒子粘连引起的统计误差问题，提出基于凹点搜索的粘连粒子分割算法。该算法在判断粒子粘连的基础上，采用Gauss-Canny算子提取煤尘轮廓，滤除噪点。滑动Harris矩阵提取灰度变换明显的特征点，形成初始凹点序列，并结合凹点夹角和面积约束条件，剔除颗粒本身凹陷形成的伪凹点。最后计算凹点的最小欧式距离构建分割线，分离粘连粒子的边界。实验结果表明，本文算法在不损害粒子几何特征的前提下，准确分离了粘连粒子。

最后，对于本文提出的基于卷积神经网络的煤尘分割算法，进行粒度分布实验。首先对测试样本进行特性评估，五次重复实验统计的粒子数目平均识别精度为95.35%，错误率为2.64%。其次，定义五种粒径进行数量粒度分布实验。相较于人工筛分法，Feret径、Martin径、椭圆径、周长径和面积径频率分布的平均误差值分别为1.43%，2.49%，1.63%，1.80%和1.77%。结合累积分布曲线，表明椭圆径和Feret径可以准确评估煤尘的数量粒度分布。最后，通过煤尘几何特征构建多元线性回归方程预测厚度，对0~75μm、75~200μm和 >200μm三种粒级的煤尘进行质量粒度分布评估。Feret径、Martin径、椭圆径、周长径和面积径质量分布的平均误差值分别为4.75%、2.45%、2.05%、4.56%和5.29%，表明本文提出的质量转换模型具有可行性。

本文的研究综合了煤矿安全和图像处理学科，设计了煤尘粒度分布的整体检测方案，产生的研究成果能够丰富煤尘图像的相关理论，提高煤尘图像分割精度，进而准确评估煤尘的粒度分布信息，具有理论意义和实际意义。

Coal dust control is the foundation of coal mine safety, and particle size distribution (PSD) of coal is the supporting information to test the dust removal effect. The PSD of coal dust can be accurately obtained by manual screening, instrument detection and other methods, but the equipment is expensive and labor costs are high, making it difficult for large-scale application in coal mine scenarios. Therefore, on the basis of analyzing the existing particle segmentation and statistical theory, this paper studies the detection algorithm of coal dust PSD based on convolutional neural network.

Firstly, aiming at the low segmentation accuracy caused by the large number of particles and irregular contours in coal dust images, an M-SegNet network model is proposed. In the codec structure of the model, coal dust features are extracted through depth wise convolution and pointwise convolution, which reduces the computational burden of the model in terms of the number of operations and parameters. An attention mechanism is introduced to obtain spatial and channel weight matrices, and global semantic information is integrated to optimize feature extraction capabilities. In the decoding stage, the encoding position is recorded to establish a position index, and the feature map size is restored by nonlinear up-sampling. Experiments are carried out on coal dust images of different sizes, and the optimal parameters of the model are fitted to effectively segment the particles of coal dust images.

Secondly, aiming at the statistical error caused by adhered particles, a segmentation algorithm based on concave points search is proposed. On the basis of judging the adhered particle area, the Gauss-Canny operator is used to extract the contour of the segmented binary image and filter out noise. The feature points with obvious grayscale transformation are extracted by sliding Harris matrix to form the initial concave point sequence. Combined with the angle and area constraints of concave points, the pseudo-concave points formed by the concave particles are eliminated. Finally, the minimum Euclidean distance of the concave points is calculated to construct a dividing line to separate the boundary of the adhered particles. The experimental results show that the algorithm in this paper can accurately separate the adhered particles without damaging the geometric characteristics of the particles.

Finally, for the coal dust segmentation algorithm based on convolutional neural network proposed in this paper, particle size distribution experiments are carried out. Firstly, the characteristics of the test samples are evaluated. The results show that the average recognition accuracy of the number of particles is 95.35%, and the error rate is 2.64%. Secondly, the quantitative particle size distribution experiments are carried out by defining five particle sizes. Compared with the manual sieving method, the mean values of the frequency distributions errors of Feret diameter, Martin diameter, ellipse diameter, perimeter diameter and area diameter are 1.43%，2.49%，1.63%，1.80% and 1.77%, respectively. Combined with the cumulative distribution curve of the above particle size, it shows that the elliptical diameter and the Feret diameter can accurately evaluate the PSD of coal dust. In addition, a multiple linear regression equation is constructed to predict the thickness through the geometric characteristics of coal dust, and the mass particle size distribution of coal dust with three particle sizes of 0~75μm, 75~200μm and >200μm are evaluated. The mean values of the mass distributions errors of Feret diameter, Martin diameter, ellipse diameter, perimeter diameter and area diameter are 4.75%、2.45%、2.05%、4.56% and 5.29% respectively, indicating that the mass conversion model proposed in this paper is feasible.

In this study, we integrate coal mine safety and image processing disciplines, and design an overall detection method for coal dust particle size distribution. The research results can enrich the relevant theories of coal dust images, improve the segmentation accuracy of coal dust images, and then accurately evaluate the PSD information of coal dust, which has theoretical and practical significance.

[1] 赵志军. 基于数理统计分析的事故管控本质研究[J]. 煤炭技术, 2020, 39(12): 187-190.

[2] 张培森, 牛辉, 朱慧聪, 等. 2019—2020年我国煤矿安全生产形势分析[J]. 煤矿安全, 2021, 52(11): 245-249.

[3] 张村, 赵毅鑫, 屠世浩, 等. 颗粒粒径对采空区破碎煤体压实破碎特征影响机制[J/OL]. 煤炭学报, 2021: 1-10.

[4] 孙深深. 矿石粒度图像检测技术的研究[D]. 郑州大学, 2019.

[5] 刘清泉, 刘嫄嫄, 杜凯, 等. 煤粒筛分现象与粒度分布特性的试验研究[J]. 中国安全科学学报, 2022, 32(2): 130-138.

[6] 焦明春, 庞厚福, 孙云栋, 等. 三种微米级颗粒粒度测试方法的比较研究[J]. 济宁学院学报, 2019, 40(5): 1-5.

[7] 冯小雨. 矿石粒度图像的研究与分析[D]. 郑州大学, 2020.

[8] 宋国庆. 基于图像处理的矿石粒度在线检测与分析[D]. 辽宁科技大学, 2020.

[9] Shi L M, Zhang J H, Zhang D, et al. Developing a dust storm detection method combining Support Vector Machine and satellite data in typical dust regions of Asia[J]. Advances in Space Research, 2020, 65: 1263-1278.

[10] 周冉, 宋华珠, 刘翔. 基于Haar-PP混合特征提取的沉降颗粒识别研究[J]. 计算机工程, 2018, 44(12): 251-257.

[11] 辛登科, 张玉杰, 苏治果. 图像处理在粉末粒度在线检测系统中的应用[J]. 计算机工程与设计, 2008(13): 3510-3512.

[12] 张伟. 基于图像处理的井下煤尘在线检测技术的研究[D]. 山东大学, 2010.

[13] 刘伟华. 基于机器视觉的煤尘在线检测系统关键技术研究[D]. 山东大学, 2011.

[14] 曾珽. 矿石粘连颗粒检测与分割算法研究[D]. 江西理工大学, 2012.

[15] 龚小明, 滕奇志, 王正勇, 等. 基于中轴线的岩心3维图像孔喉分割算法[J]. 四川大学学报(工程科学版), 2016, 48(S2): 100-106.

[16] 何磊, 王正勇, 滕奇志, 等. 基于自适应H-极大值的粘连颗粒分割算法[J]. 现代计算机(专业版), 2018(33): 29-33.

[17] 张文晔. 基于机器视觉的金刚砂线颗粒检测技术的研究与应用[D]. 江苏理工学院, 2018.

[18] 吴锴. 基于边界凹凸点和神经网络的粘连颗粒图像分割算法研究[D]. 福州大学, 2018.

[19] Li Y, Duan Y Y. A Method of Ore Image Segmentation Based on Deep Learning[C]. in Intelligent Computing Methodologies 2018, (ICIC 2018) Lecture Notes in Computer Science, 2018, 10956.

[20] Duan J, Liu X, Wu X, et al. Detection and segmentation of iron ore green pellets in images using lightweight U-net deep learning network[J]. Neural Computer & Application, 2020, 32: 5775-5790.

[21] 李鹏飞. 基于卷积神经网络的煤堆烟雾检测算法研究[D]. 燕山大学, 2020.

[22] 熊祥祥. 基于深度学习的综采工作面煤岩识别方法研究[D]. 中国矿业大学, 2020.

[23] 谢涛. 基于深度学习的微细粒矿物识别研究[D]. 中国矿业大学, 2020.

[24] 姚新. 面向卷积神经网络的模型压缩方法研究[D]. 扬州大学, 2021.

[25] Long J, Shelhamer E, Darrell T. Fully Convolutional Networks for Semantic Segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 39(4): 640-651.

[26] 张志超. 改进卷积神经网络的图像语义分割研究[D]. 东北林业大学, 2021.

[27] André D, Marcelo B, Mauren L, et al. Combining max-pooling and wavelet pooling strategies for semantic image segmentation[J]. Expert Systems with Applications, 2021, 183, 115403.

[28] Tseng K K, Lin J R, Chen C M, et al. A fast instance segmentation with one-stage multi-task deep neural network for autonomous driving[J]. Computers & Electrical Engineering, 2021, 93.

[29] 薛程, 叶少珍. 多尺度定位信息增强的图像全景分割方法[J]. 福州大学学报(自然科学版), 2021, 49(3): 302-308.

[30] Badrinarayanan V, Kendall A, Cipolla R. SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(12): 2481-2495.

[31] Wang W Z, Wang S Y, Li Y, et al. Adaptive multi-scale dual attention network for semantic segmentation[J]. Neurocomputing, 2021, 460: 39-49.

[32] Fu J, Liu J, Tian H, et al. Dual attention network for scene segmentation[C]. In: Conference on Computer Vision and Pattern Recognition, 2019, 3141–3149.

[33] Khan F, Hussain S, Basal S, et al. An efficient encoder–decoder model for portrait depth estimation from single images trained on pixel-accurate synthetic data[J]. Neural Networks, 2021, 142: 479-491.

[34] Sun J D, Peng Y J, Guo Y F, et al. Segmentation of the multimodal brain tumor image used the multi-pathway architecture method based on 3D FCN[J]. Neurocomputing, 2021, 423: 34-45.

[35] Sathananthavathi V, Indumathi G. Encoder Enhanced Atrous (EEA) Unet architecture for Retinal Blood vessel segmentation[J]. Cognitive Systems Research, 2021, 67: 84-95.

[36] Zhu X L, Cheng Z Y, Wang S, et al. Coronary angiography image segmentation based on PSPNet[J]. Computer Methods and Programs in Biomedicine, 2021, 200, 105897.

[37] Zhang Z, Yang J, Su X, et al. Multi-scale image segmentation of coal piles on a belt based on the Hessian matrix[J]. Particuology, 2013, 11(5): 549–555.

[38] 张亚秋, 吴文福, 王刚. 基于逐步改变阈值方法的玉米种子图像分割[J]. 农业工程学报, 2011, 27(7): 200-204.

[39] 张建立, 叶平坤, 孙深深. 形态学图像处理下的矿石粒度的检测[J]. 机械设计与制造, 2020(3): 68-71.

[40] Zhang L, Zou L, Wu C Y, et al. Method of famous tea sprout identification and segmentation based on improved watershed algorithm[J]. Computers and Electronics in Agriculture, 2021, 184: 0168-1699.

[41] 王瑞, 杨朝旭, 荣海军, 等. 阈值自适应的极线距离变换算法及其在立体匹配中的应用[J]. 西安交通大学学报, 2021, 55(7): 132-140.

[42] 高辉, 甄彤, 李智慧. 粘连颗粒图像的分割方法综述[J/OL]. 中国粮油学报, 2021:1-14.

[43] 潘少伟, 王蒙歌, 王亚玲, 等. 改进Sobel算子对岩心图像的边缘检测研究[J/OL]. 数学的实践与认识, 2021: 1-11.

[44] 代文征, 杨勇. 基于改进高斯—拉普拉斯算子的噪声图像边缘检测方法[J]. 计算机应用研究, 2019, 36(8): 2544-2547+2555.

[45] 李静, 陈桂芬, 丁小奇. 基于改进Canny算法的图像边缘检测方法研究[J]. 计算机仿真, 2021, 38(4): 371-375.

[46] Harris C, Stephens M. A Combined Corner and Edge Detector[C]. In Proceedings of the 4th Alvey vision conference. United Kingdom: BMVA, 1988: 147-151.

[47] 王小鹏, 姚丽娟, 文昊天, 等. 形态学多尺度重建结合凹点匹配分割枸杞图像[J].农业工程学报, 2018, 34(2): 212-218.

[48] Sun A Y, Jia W B, Hei D Q, et al. Application of concave point matching algorithm in segmenting overlapping coal particles in X-ray images[J]. Minerals Engineering, 2021, 171, 107096.

[49] He L, Guo Y C, Wang S, et al. Multi-scale Coal and Gangue Dual-energy X-ray Image Concave Point Detection and Segmentation Algorithm[J]. Measurement, 2022, 111041.

[50] Bai F Y, Fan M Q, Yang H L, et al. Image segmentation method for coal particle size distribution analysis[J]. Particuology, 2021, 56: 163-170.

[51] 蔡改贫, 刘占, 汪龙, 等. 基于形态学优化处理的标记符分水岭矿石图像分割[J]. 科学技术与工程, 2020, 20(23): 9497-9502.

[52] 燕红文, 崔清亮. 基于分水岭分割算法对燕麦籽粒识别中粘连区域的分割[J]. 计算机时代, 2020, 01: 23-25.

[53] 张文飞, 韩建海, 郭冰菁, 等. 改进的分水岭算法在粘连图像分割中的应用[J]. 计算机应用与软件, 2021, 38(6): 243-248.

[54] Zhang Z, Yang J, Su X, et al. Analysis of large particle sizes using a machine vision system[J]. Physicochem. Probl. Min. Process, 2013, 49: 397–405.

[55] Slobodan D, Nataša S, Joakim L. Estimation of Feret’s diameter from pixel coverage representation of a shape[J]. Pattern Recognition Letters, 2016, 80: 37-45.

[56] Loeb P A. An intuitive approach to the Martin boundary[J]. Indagationes Mathematicae, 2020, 31(5): 879-884.

[57] Mihal D, Zorana A, Darko J, et al. Optimal pixel resolution for sand particles size and shape analysis[J]. Powder Technology, 2016, 302: 177-186.

[58] 张立志, 王潇, 张召强. 综采工作面煤尘粒度分布特性分析[J]. 煤矿现代化, 2017(2): 138-140.

[59] Vallebuona G, Arburo K, Casali A. A Procedure to Estimate Weight Particle Distributions from Area Measurements[J]. Minerals Engineering, 2003, 16(4): 323-329.

[60] Larry B, Cheng K, John Z. Estimation of limestone particle mass from 2D images[J]. Powder Technology, 2003, 132: 184-189.

[61] Andersson T, Thurley M, Carlson J. A Machine Vision System for Estimation of Size Distributions by Weight of Limestone Particles[J]. Minerals Engineering, 2016, 25(1): 38-46.

[62] Zhang Z L, Liu Y, Hu Q, et al. Multi - information online detection of coal quality based on machine vision[J]. Powder Technology, 2020, 374: 205-262.