张 勇a,c，郭永存b,c，陈 伟b,c，王 爽b,c，程 刚b,c
（安徽理工大学 a. 电气与信息工程学院；b. 机械工程学院；c. 深部煤矿采动响应与灾害防控国家重点实验室，安徽 淮南 232001）
DOI:10.13732/j.issn.1008-5548.2023.01.007
收稿日期： 2022-05-22，修回日期：2022-11-10，在线出版时间：2022-12-02 16:20。
基金项目：国家自然科学基金项目，编号：51904007；安徽省科技重大专项资助项目，编号：18030901049；安徽省高校协同创新项目，编号：GXXT-2021-076。
第一作者简介：张勇(1996—)，男，硕士研究生，研究方向为煤矸智能分选技术。E-mail：15955942745@163.com。
通信作者简介：郭永存(1965—)，男，教授，博士，博士生、 硕士生导师，研究方向为煤矿智能化技术。E-mail： guoyc@aust.edu.cn。
摘要：针对传统轻量型卷积神经网络模型复杂度高，移动端识别速度慢，小样本数据集上训练、识别效果差的等问题，提出一种高效的改进后的移动端煤矸识别方法；分析卷积神经网络模型轻量化的方法，并从注意力机制、激活函数和分类头3个方面对MobileNetv3网络进行改进，通过模型量化压缩网络在移动端部署模型，分析改进网络量化前、后的空间存储容量，浮点运算次数，推理时间和识别准确率；在移动端煤矸识别实验装置中训练、部署和测试模型的识别效果。结果表明：改进后网络经过20次的训练后模型即收敛，收敛速度较快，训练和验证准确率均大于99%;改进后模型经量化压缩后模型存储容量较小，仅为原网络的24.64%,模型复杂度大幅度下降；移动端推理时间仅为77 ms,识别准确率达到99.7%;利用实验装置实时采集的煤和矸石图像的识别效果较好，识别方法可靠。
关键词：煤矸识别；网络轻量化；模型压缩；注意力机制；小样本数据集；移动端
Abstract：An efficient and improved mobile terminal coal and gangue recognition method(E-MobileNetv3) was proposed, aiming at the problems such as high complexity of traditional lightweight convolutional neural network model, slow recognition speed of mobile terminal, poor training and recognition effect on small sample data sets. The lightweight method of convolutional neural network model was analyzed, and the MobileNetv3 network was improved from the three aspects of attention mechanism, activation function and classification head. The spatial storage capacity, floating point operation number, inference time and recognition accuracy of the improved network before and after quantization were analyzed by the model quantization compression network deployment model on the mobile devices. The model was trained, deployed and tested in an experimental equipment for coal and gangue recognition at mobile device. The results show that the improved network model converges after 20 times of training, the convergence speed is fast, and the accuracy of training and verification is greater than 99%. After quantization and compression, the storage capacity of the improved model is smaller, only 24.64% of that of the original network, and the complexity of the model is greatly reduced. The inference time of mobile device is only 77 ms, and the recognition accuracy reaches 99.7%. The identification effect of coal and gangue images collected in real time by the experimental device is good, which verifies the reliability of the identification method.
Keywords：recognition of coal and gangue; network lightweight; model compression; attention mechanisms; small dataset; mobile device
参考文献（References）：
[1]MA D, DUAN H Y, LIU J F, et al. The role of gangue on the mitigation of mining-induced hazards and environmental pollution: an experimental investigation[J]. Science of the Total Environment, 2019, 664: 436-448.
[2]娄德安, 杨康. SKT跳汰机的发展与应用[J]. 选煤技术, 2016(2): 90-92.
[3]SHAHBAZI B, S CHELGANI C. Modeling of fine coal flotation separation based on particle characteristics and hydrodynamic conditions[J]. International Journal of Coal Science and Technology, 2016, 3(4): 429-439.
[4]BAHRAMI A, GHORBANI Y, MIRMOHAMMADI M, et al. The beneficiation of tailing of coal preparation plant by heavy-medium cyclone[J]. International Journal of Coal Science and Technology, 2018, 5(3): 374-384.
[5]鲁恒润, 王卫东, 徐志强, 等. 基于机器视觉的煤矸特征提取与分类研究[J]. 煤炭工程, 2018, 50(8): 137-140.
[6]DOU D Y, WU W Z, YANG J G, et al. Classification of coal and gangue under multiple surface conditions via machine vision and relief-SVM[J]. Powder Technology, 2019, 35(6): 1024-1028.
[7]HOU W. Identification of coal and gangue by feed-forward neural network based on data analysis[J]. International Journal of Coal Preparation and Utilization, 2019, 39(1): 33-43.
[8]徐志强, 吕子奇, 王卫东, 等. 煤矸智能分选的机器视觉识别方法与优化[J]. 煤炭学报, 2020, 45(6): 2207-2216.
[9]LV Z Q, WANG W D, XU Z Q, et al. Fine-grained object detection method using attention mechanism and its application in coal-gangue detection[J]. Applied Soft Computing Journal, 2021, 113: 107891-107904.
[10]郜亚松, 张步勤, 郎利影. 基于深度学习的煤矸石识别技术与实现[J]. 煤炭科学技术, 2021, 49(12): 202-208.
[11]杜京义, 史志芒, 郝乐, 等. 轻量化煤矸目标检测方法研究[J]. 工矿自动化, 2021, 47(11): 119-125.
[12]LI G, TONG N, ZHANG Y, et al. Moving target detection classifier for airborne radar using squeezenet[J]. Journal of Physics: Conference Series, 2021, 1883(1): 012003-012009.
[13]HAN S, LIU X Y, MAO H Z, et al. EIE: efficient inference engine on compressed deep neural network[J]. Computer Architecture News, 2016, 44(3): 243-254.
[14]樊景超. 基于MobileNets的果园害虫分类识别模型研究[J]. 天津农业科学, 2018, 24(9): 11-13,26.
[15]LIU J, WANG X. Early recognition of tomato gray leaf spot disease based on MobileNetv2-YOLOv3 model[J]. Plant Methods, 2020, 16(1): 1-16.
[16]ABD E M, DAHOU A, ALSALEH N A, et al. Boosting COVID-19 image classification using MobileNetV3 and aquila optimizer algorithm[J]. Entropy, 2021, 23(11)： 1383.
[17]THACHAN SOPHANYOULY. 基于ShuffleNet的人脸识别[D]. 杭州: 浙江大学, 2019.
[18]彭红星, 徐慧明, 刘华鼐. 基于改进ShuffleNet V2的轻量化农作物害虫识别模型[J]. 农业工程学报, 2022, 38(11): 161-170.
[19]HU J, LI S, GANG S. Squeeze-and-excitation networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 42(8): 2011-2023.
[20]QING Y, LIU W. Hyperspectral image classification based on multi-scale residual network with attention mechanism[J]. Remote Sensing, 2021, 13(3): 335-352.
[21]LANGER S. Analysis of the rate of convergence of fully connected deep neural network regression estimates with smooth activation function[J]. Journal of Multivariate Analysis, 2021, 182: 104695-104708.
