结构化区域全卷积神经网络的钢轨扣件检测方法

doi:10.3778/j.issn.1673-9418.2009065

摘要/Abstract

摘要：

现有的深度学习模型很难满足高速检测的实时性，针对性地提出了一种结构化区域全卷积神经网络（SR-FCN）。为了满足高速综合巡检车实时检测的要求，考虑到轨道图像中钢轨、扣件、轨道板等设施位置相对固定，其位置分布可以构成轨道场景特有的结构化特征，因此设定了结构化检测区域，将一幅图像中扣件小目标的检测转化为一整块具有固定结构的大目标区域检测，将扣件小目标的检测问题转化为结构化区域的定位问题，可加快网络的训练收敛速度，减少候选区域的生成个数，从而大幅提高检测速度。将铁路轨道的结构化先验信息融合到深度学习网络的各个过程中，有效提高了定位精度，保证了检测的鲁棒性。实验室离线分析和现场在线检测的结果表明，所提出的SR-FCN网络分别获得了99.99%和99.84%的检测精度，同时还保持了较快的检测速度，可以满足350 km/h的实时检测要求。

关键词: 目标检测, 深度学习, 扣件, 结构化场景, 高速巡检

Abstract:

Existing deep learning models are difficult to meet the real-time performance of high-speed detection, and a structured region fully convolutional neural network (SR-FCN) is proposed. In order to meet the task of real-time detection of high-speed comprehensive inspection vehicles, considering that the rails, rail fasteners, track plates and other facilities in the track image are relatively fixed, their position distribution can constitute a unique structural feature of the track scene, structured detection region is set, the detection of small rail fastener targets in an image is converted into a large target region detection with a fixed structure, and the detection of small rail fastener targets is transformed into a positioning problem in a structured region, which can speed up the network training convergence speed, reduce the number of candidate regions generated, thereby greatly improving the detection speed. The struc-tured prior information of the railway track is integrated into the various processes of the deep learning network, which effectively improves the positioning accuracy to ensure the robustness of the detection. The results of labor-atory offline analysis and on-site online testing show that the proposed SR-FCN obtains 99.99% and 99.84% detec-tion accuracy, respectively, and maintains a relatively fast detection speed, which can meet the 350 km/h real-time detection requirements.

Key words: target detection, deep learning, rail fasteners, structured scenes, high-speed inspection

蒋欣兰. 结构化区域全卷积神经网络的钢轨扣件检测方法[J]. 计算机科学与探索, 2020, 14(11): 1888-1898.

JIANG Xinlan. Rail Fastener Detection Method Based on Structured Region Full Convolution Neural Network[J]. Journal of Frontiers of Computer Science and Technology, 2020, 14(11): 1888-1898.

参考文献

[1] Girshick R, Donahue J, Darrell T, et al. Rich feature hierar-chies for accurate object detection and semantic segmenta-tion[C]//Proceedings of the 2014 IEEE Conference on Com-puter Vision and Pattern Recognition, Columbus, Jun 23-28, 2014. Washington: IEEE Computer Society, 2014: 580-587.
[2] Redmon J, Divvala S, Girshick R, et al. You only look once: unified, real-time object detection[J]. arXiv:1506.02640, 2015.
[3] Liu W, Anguelov D, Erhan D, et al. SSD: single shot multi-box detector[C]//Proceedings of the European Conference on Computer Vision. Cham: Springer, 2016: 21-37.
[4] He K, Zhang X, Ren S, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE Tran-sactions on Pattern Analysis and Machine Intelligence, 2015, 37(9): 1904-1916.
[5] Girshick R. Fast R-CNN[C]//Proceedings of the 2015 IEEE International Conference on Computer Vision, Santiago, Dec 7-13, 2015. Piscataway: IEEE, 2015: 1440-1448.
[6] Ren S, He K, Girshick R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[C]// Proceedings of the 28th International Conference on Neural Information Processing Systems. Cambridge: MIT Press, 2015: 91-99.
[7] Redmon J, Farhadi A. YOLO9000: better, faster, stronger[C]// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, Jul 21-26, 2017. Washington: IEEE Computer Society, 2017: 7263-7271.
[8] Redmon J, Farhadi A. Yolov3: an incremental improvement[J]. arXiv:1804.02767, 2018.
[9] Bochkovskiy A, Wang C Y, Liao H Y M. YOLOv4: optimal speed and accuracy of object detection[J]. arXiv:2004.10934, 2020.
[10] Dai J, Li Y, He K, et al. R-FCN: object detection via region-based fully convolutional networks[C]//Proceedings of the 30th International Conference on Neural Information Pro-cessing Systems. Red Hook: Curran Associates Inc., 2016: 379-387.
[11] Wang L, Guo S, Huang W, et al. Places 205-vggnet models for scene recognition[J]. arXiv:1508.01667, 2015.
[12] Sasha T, Almeida D, Lyman K. Resnet in resnet: generaliz-ing residual architectures[J]. arXiv:1603.08029, 2016.

[1]	王迪聪，白晨帅，邬开俊. 基于深度学习的视频目标检测综述[J]. 计算机科学与探索, 2021, 15(9): 1563-1577.
[2]	张晓旭，马志强，刘志强，朱方圆，王春喻. Transformer在语音识别任务中的研究现状与展望[J]. 计算机科学与探索, 2021, 15(9): 1578-1594.
[3]	陈璠，彭力. 多层级重叠条纹特征融合的行人重识别[J]. 计算机科学与探索, 2021, 15(9): 1753-1761.
[4]	武家伟，孙艳春. 融合知识图谱和深度学习方法的问诊推荐系统[J]. 计算机科学与探索, 2021, 15(8): 1432-1440.
[5]	马煜，杜慧敏，毛智礼，张霞. 深度语义分割人群密度检测技术[J]. 计算机科学与探索, 2021, 15(8): 1469-1475.
[6]	荣欢，马廷淮. 利用收益预测与策略梯度两阶段众包评论集成[J]. 计算机科学与探索, 2021, 15(8): 1476-1489.
[7]	马玉琨，徐姚文，赵欣，徐涛，王泽瑞. 人脸识别系统的活体检测综述[J]. 计算机科学与探索, 2021, 15(7): 1195-1206.
[8]	葛轶洲，许翔，杨锁荣，周青，申富饶. 序列数据的数据增强方法综述[J]. 计算机科学与探索, 2021, 15(7): 1207-1219.
[9]	方钧婷，谭晓阳. 注意力级联网络的金属表面缺陷检测算法[J]. 计算机科学与探索, 2021, 15(7): 1245-1254.
[10]	田萱，丁琪，廖子慧，孙国栋. 基于深度学习的新闻推荐算法研究综述[J]. 计算机科学与探索, 2021, 15(6): 971-998.
[11]	能文鹏，陆军，赵彩虹. 基于关系归纳偏置的睡眠分期综述[J]. 计算机科学与探索, 2021, 15(6): 1026-1037.
[12]	吕昊远，俞璐，周星宇，邓祥. 半监督深度学习图像分类方法研究综述[J]. 计算机科学与探索, 2021, 15(6): 1038-1048.
[13]	马宇，张丽果，杜慧敏，毛智礼. 卷积神经网络的交通标志语义分割[J]. 计算机科学与探索, 2021, 15(6): 1114-1121.
[14]	汤凌燕，熊聪聪，王嫄，周宇博，赵子健. 基于深度学习的短文本情感倾向分析综述[J]. 计算机科学与探索, 2021, 15(5): 794-811.
[15]	刘靖祎，史彩娟，涂冬景，刘帅. 零样本图像分类综述[J]. 计算机科学与探索, 2021, 15(5): 812-824.