随着深度学习的快速发展，目标检测识别技术取得了显著突破，并成为计算机视觉领域研究的热点。传统目标检测算法在矿井特殊环境中，易受到光照变化、遮挡、尺度变化等外在因素的影响，使目标检测精度降低、检测效果不理想。而深度学习模型的使用提高了目标检测精度，但随之而来的是算力成本的不断增加，导致模型无法在算力较低的设备上实现对目标的实时检测。为此本文针对矿井复杂环境，构建了一种轻量化目标检测算法，实现对矿井机电设备工作状态高精度的实时检测识别，并在真实矿井数据上进行了应用。主要研究工作如下：

(1)阐述了传统目标检测算法以及基于卷积神经网络的目标检测算法的基本原理和操作流程，并对其优势和弊端进行了研究分析。通过对四种经典算法性能的对比分析，确定矿井机电设备状态检测识别的设计策略。

(2)针对SSD模型在矿井目标检测过程中参数量过大、漏检率高以及误检的问题，在SSD的基础特征提取网络中采用MobileNet-V2模型中的深度可分离卷积和线性瓶颈结构，将原有的VGG16网络进行替换，完成对模型的剪枝，实现模型的轻量化；利用高斯加权的惩罚函数对NMS算法中的置信度衰减函数进行改进，提高模型的检测性能。实验结果表明，本算法模型参数量减少了86.6M，检测速度提高了62FPS，漏检和误检的问题得到了有效的改善。

(3)针对矿井特殊环境下模型特征提取能力不足的问题，采用FPN网络结构对SSD模型的特征提取网络进行改进，实现高层语义信息与低层空间位置信息的融合；使用K-means++算法对模型默认框的长宽比例进行优化，使模型默认框的尺寸更加符合自制数据集中真实目标的尺寸，提高模型对矿井机电设备闸刀开关状态、按钮指示灯状态、设备数字仪表显示状态、指针式仪表整体特征的学习能力及检测识别能力，并结合传统算法完成指针式仪表的读数。实验结果表明，本文设计的多尺度特征融合的SSD机电设备工作状态检测识别算法比SSD目标检测算法的mAP提高了1.5%。

本文基于SSD目标检测算法从检测精度和检测速度两个方面进行了优化，针对矿井机电设备工作状态检测识别设计了轻量化算法模型，相比于SSD算法在识别准确度和检测速度方面都有很大程度的提升。

With the rapid development of deep learning, target detection recognition technology has made significant breakthroughs and become a hot spot for research in the field of computer vision. Traditional target detection algorithms are susceptible to external factors such as light changes, occlusion and scale changes in the special environment of mines, which make the target detection accuracy lower and the detection effect unsatisfactory. The use of deep learning models has improved the target detection accuracy, but the consequent increasing cost of arithmetic power has led to the inability of the models to achieve real-time detection of targets on devices with low arithmetic power. For this reason, in this thesis, a lightweight target detection algorithm is constructed for the complex environment of mines to achieve high-precision real-time detection and recognition of the working status of mine electromechanical equipment, and it is applied on real mine data. The main research works are as follows.

(1) The basic principles and operation processes of traditional target detection algorithms and convolutional neural network-based target detection algorithms are described, and their advantages and disadvantages are studied and analyzed. Through the comparative analysis of the performance of four classical algorithms, the design strategy of mine electromechanical equipment condition detection and identification is determined.

(2) To address the problems of excessive parameter size, high leakage rate and false detection in the SSD model in the process of mine target detection, the depth-separable convolution and linear bottleneck structure in the MobileNet-V2 model are used in the base feature extraction network of SSD to replace the original VGG16 network to complete the pruning of the model and realize the lightweight of the model; the penalty function of Gaussian weighting is used The confidence decay function in the NMS algorithm is improved to improve the detection performance of the model. The experimental results show that the amount of model parameters of this algorithm is reduced by 86.6M, the detection speed is improved by 62FPS, and the problems of missed detection and false detection are effectively improved.

(3) To address the problem of insufficient feature extraction capability of the model in the special environment of the mine, the feature extraction network of the SSD model is improved by using the FPN network structure to realize the fusion of high-level semantic information and low-level spatial location information; the aspect ratio of the default box of the model is optimized by using the K-means++ algorithm to make the size of the default box of the model more consistent with the size of the real target in the homemade dataset. Improve the learning ability and detection recognition ability of the model for the gate switch state of mine electromechanical equipment, button indicator state, digital meter display state of equipment, and overall characteristics of pointer meters, and complete the reading of pointer meters by combining with traditional algorithms. The experimental results show that the multi-scale feature fusion of SSD electromechanical equipment working state detection and recognition algorithm designed in this thesis improves the mAP by 1.5% over the SSD target detection algorithm.

This thesis optimizes the SSD target detection algorithm in terms of both detection accuracy and detection speed, and designs a lightweight algorithm model for mine electromechanical equipment working condition detection and recognition, which has a great degree of improvement in recognition accuracy and detection speed compared to the SSD algorithm.

[1] M. K. and Binte Maisha and Sakib Nazmus. Targeted Face Recognition and Alarm Generation for Security Surveillance using Single Shot Multibox Detector (SSD)[J]. International Journal of Computer Applications, 2019, 177(22): 8-13.

[2] 王志华, 罗齐, 刘绍廷. 基于混沌灰狼优化算法的SVM分类器研究[J]. 计算机工程与科学, 2018, 040(011): 2040-2046.

[3] Sharma Rahul and Singh Amar. An Integrated Approach towards Efficient Image Classification Using Deep CNN with Transfer Learning and PCA[J]. Advances in Technology Innovation, 2022, 7(2): 105-117.

[4] Li H, Lin Z, Shen X, et al. A convolutional neural network cascade for face detection[C]// Computer Vision and Pattern Recognition. IEEE, Boston USA, 2015:5325-5334.

[5] 刘禹欣, 朱勇, 孙结冰, 王一博. Haar-like特征双阈值Adaboost人脸检测[J]. 中国图象图形学报, 2020, 25(08): 1618-1626.

[6] Human and Robot Tracking Using Histogram of Oriented Gradient Feature[J]. JOURNAL OF PLATFORM TECHNOLOGY, 2018, 6(4): 18-25.

[7] Bin Li and Fuqiang Sun and Yonghan Zhang. Building Recognition Using Gist Feature Based on Locality Sensitive Histograms of Oriented Gradients[J]. Pattern Recognition and Image Analysis, 2019, 29(2): 258-267.

[8] Abdelmalik Moujahid, Fadi Dornaika. Multi-scale multi-block covariance descriptor with feature selection[J]. Neural Computing and Applications, 2019, 32(10): 1-12.

[9] DUAN Hua, FENG Tong, LIU Songning, ZHANG Yulin, SU Jionglong. Tumor Classification of Gene Expression Data by Fuzzy Hybrid Twin SVM[J]. Chinese Journal of Electronics, 2022, 31(01): 99-106.

[10] Danapur N, Dizaj S, Rostami V. An efficient image retrieval based on an integration of HSV, RLBP, and CENTRIST features using ensemble classifier learning[J]. Multimedia Tools and Applications, 2020, 79(33): 24463-24486.

[11] Felzenszwalb P, Girshick R, Mcallester D, et al. Visual object detection with deformable part models[J]. Communications of the Acm, 2010, 56(9): 2241-2248.

[12] Song P, Qi L, Qian X, et al. Detection of ships in inland river using high-resolution optical satellite imagery based on mixture of deformable part models[J]. Journal of Parallel and Distributed Computing, 2019, 132(OCT.): 1-7.

[13] 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.

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

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

[16] Wu Tian cheng, Wang Xiao quan, Cai Yi jun, et al. Lightweight SSD object detection method based on feature fusion[J]. CHINESE JOURNAL OF LIQUID CRYSTALS AND DISPLAYS, 2021, 36(10): 1437-1444.

[17] Howard A G, Zhu M, Chen B, et al. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications[J]. 2017, 50(17): 188-189.

[18] Sandler M, Howard A, Zhu M, et al. Mobilenetv2: Inverted residuals and linear bottlenecks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 4510-4520.

[19] 林本丰, 王呈, 孙悦程. 融合LSD算法与深度学习的开关状态检测方法[J]. 计算机工程与应用, 2021, 57(17): 181-189.

[20] 高露. 基于深度学习的变电站开关状态识别系统的设计与实现[D]. 中国科学院大学. 2019.

[21] 史梦安, 陆振宇. 基于YOLO-tiny-RFB模型的电站旋钮开关状态识别[J]. 计算机应用, 2020, 40(12): 3679-3686.

[22] 周凯, 唐瞾, 胡志坚, 严利雄, 毕如玉, 邓科, 李煜磊. 基于NSST图像融合的变电站开关状态识别[J]. 高压电器, 2021,57(10): 50-58.

[23] 高琦煜, 方虎生. 多卷积特征融合的HOG行人检测算法[J]. 计算机科学, 2017, 44(S2): 199-201+232.

[24] Chen F, Müller J, Müller J, et al. Efficient feature-based motion estimation in neurosurgery using non-maximum suppression[J]. Current Directions in Biomedical Engineering, 2018, 4(1): 555-558.

[25] Tomczak, Jakub M. Learning Informative Features from Restricted Boltzmann Machines [J]. Neural Processing Letters, 2016, 44(3): 735-750.

[26] Zhao S, Yin L, Zhang J, et al. Real-time fabric defect detection based on multi-scale convolutional neural network[J]. IET Collaborative Intelligent Manufacturing, 2020, 2(4): 189-196.

[27] 谢永明, 王红蕾. 复杂背景下远距离及小尺寸行人检测改进算法[J]. 计算机工程与设计, 2021, 42(05): 1323-1330.

[28] Vinod Kumar Chauhan, Kalpana Dahiya, Anuj Sharma. Problem formulations and solvers in linear SVM: a review[J]. Artificial Intelligence Review, 2019, 52(2): 803-855.

[29] Zhenzhen Song, Longsheng Fu, Jingzhu Wu, et al. Kiwifruit detection in field images using Faster R-CNN with VGG16[J]. IFAC Papers On Line, 2019, 52(30): 76-81.

[30] 刘萍, 杨鸿波, 宋阳. 改进YOLOv3网络的海洋生物识别算法[J]. 计算机应用研究,2020, 37(S1): 394-397.

[31] Su Jian-li, Wang Hua. An improved adaptive differential evolution algorithm for single unmanned aerial vehicle multitasking[J]. Defence Technology, 2021, 17(6): 1967-1975.

[32] P. Pandiyan et al. Real-time monitoring of social distancing with person marking and tracking system using YOLO V3 model[J]. International Journal of Sensor Networks, 2022, 38(3): 154-165.

[33] Quang T N, Lee S, Song B C. Object detection using improved bi-directional feature pyramid network[J]. Electronics, 2021, 10(6): 746-746.

[34] Zhang T, Hao L Y, Guo G. A Feature Enriching Object Detection Framework with Weak Segmentation Loss[J]. Neurocomputing, 2019, 335(MAR.28): 72-80.

[35] Srinivasu P N, SivaSai J G, et al. Classification of skin disease using deep learning neural networks with MobileNetV2 and LSTM[J]. Sensors, 2021, 21(8): 2852.-2852.

[36] 侯冰基, 杨明辉, 孙晓玮. 基于改进Faster RCNN的毫米波图像实时目标检测[J]. 激光与光电子学进展, 2019, 56(13): 127-133.

[37] Feng Y, Tao X, Lee E J. Classification of Shellfish Recognition Based on Improved Faster R-CNN Framework of Deep Learning[J]. Mathematical Problems in Engineering, 2021, 2021(28): 1-10.

[38] 苏钰桐, 杨炜毅, 李俊霖. 基于YOLO v3的煤岩钻孔图像裂隙智能识别方法[J]. 煤矿安全, 2021, 52(4): 156-161.

[39] Buslaev A, Iglovikov V I, Khvedchenya E, et al. Albumentations: Fast and Flexible Image Augmentations[J]. Information (Switzerland), 2020, 11(2):125-125.

[40] Rampasek L, Goldenberg A. TensorFlow: Biology's Gateway to Deep Learning?[J]. Cell Systems, 2016, 2(1): 12-14.

[41] Wang Xi and Yu Ming and Ren Hong e. Remote sensing image semantic segmentation combining UNET and FPN[J]. CHINESE JOURNAL OF LIQUID CRYSTALS AND DISPLAYS, 2021, 36(3): 475-483.

[42] 喻钧, 康秦瑀, 陈中伟, 等. 基于全卷积神经网络的遥感图像海面目标检测[J]. 弹箭与制导学报, 2020, 40(5): 15-19+23.

[43] 高梅鹃, 刘达, 庄建斌, 等. 基于图像识别技术的变电站视频监控系统设计[J]. 电子设计工程, 2021, 29(22): 175-179.

[44] 卢红阳, 刘且根, 熊娇娇, 等. 基于最大加权投影求解的彩色图像灰度化对比度保留 算法[J]. 自动化学报, 2017, 43(5): 843-854.

[45] 袁立, 周伟. 一种改进的基于边缘检测的中值滤波图像去噪算法[J]. 电子制作, 2020(13): 53-55.

[46] Aggarwal Alankrita, Singh Shweta, Sharma Ayush. Performance Comparison of Image Processing Techniques on Various Filters: A Review[J]. International Journal of Security and Privacy in Pervasive Computing (IJSPPC), 2021, 13(3): 34-42.

[47] 吴杰,吴怀宇,陈洋.基于轮廓拟合与径向分割的指针式仪表自动读数方法[J]. 高技术通讯, 2019, 29(09): 905-913.

[48] 盖荣丽, 蔡建荣, 王诗宇, 等. 卷积神经网络在图像识别中的应用研究综述[J]. 小型微型计算机系统, 2021, 42(9): 1980-1984.

[49] 宋晓宇, 袁帅, 郭寒冰, 等. 基于自适应阈值区间的广义Hough变换图形识别算法[J]. 仪器仪表学报, 2014, 35(5): 1109-1117.

[50] 高建龙,郭亮,吕耀宇,吴清文,母德强.改进ORB和Hough变换的指针式仪表识读方法[J].计算机工程与应用,2018,54(23):252-258.

[51] 李建森, 项偲. 基于随机采样的随机Hough变换快速圆检测算法[J]. 科技创新与应用, 2021, 11(29): 128-130.

[52] 闫勇, 陈立夫, 郭坤坤, 江朝晖, 饶元. Lab色彩模型下梯度Hough圆变换的成熟草莓识别[J]. 安徽农业大学学报, 2020, 47(03): 488-493.

[53] 张国福, 王呈. 基于改进概率霍夫直线检测的电梯门状态检测方法[J]. 南京理工大学学报, 2020, 44(2): 162-170.

[54] 王兴涛, 李文永, 陈亮. 用于电力巡检机器人的挂空异物智能检测方法[J]. 机械与电子, 2021, 39(12): 76-80.

[55] 赵乐, 王先培, 代荡荡, 等. 复杂背景下电力线自动提取算法[J]. 高电压技术, 2019, 45(1): 218-227.

[56] 孙晓朋, 侯立群, 渠怀胜. 基于卷积神经网络的渐进式指针表自动读数方法[J]. 传感技术学报, 2021, 34(10): 1326-1333.

[57] 汤亮, 何稳, 李倩, 张董洁, 汪威, 胡新宇. 基于空间变换的指针式仪表读数识别算法研究[J]. 电测与仪表, 2018,55(06): 116-121.

[58] 高露, 马元婧. 基于Faster R-CNN的设备故障检测与识别[J]. 计算机系统应用, 2019, 28(04): 170-175.

[59] 胡鑫, 欧阳华, 侯卓成, 等. 特征增强的SSD算法及在仪表读数识别上的应用[J]. 电子测量技术, 2021, 44(17): 154-159.