新型Lite YOLOv4-Tiny算法及其在裂纹智能检测中的应用

doi:10.1007/s12204-022-2504-8

摘要/Abstract

摘要： 为了顺应高层建筑裂缝检测市场需求的快速增长，本文提出了将深度网络技术集成到爬墙机器人中进行裂缝检测的想法。考虑到在树莓派Raspberry Pi等边缘设备上部署时的依赖性和硬件要求，选择了Darknet神经网络作为检测的基本框架。为了提高在边缘设备上的推理效率并避免深度网络可能出现的过早过拟合，对原始的YOLOv4-Tiny算法进行了改进，得到了Lite YOLOv4-Tiny算法，并使用Netron对其结构进行了相应的可视化。处理了从互联网上下载以及从校园建筑物拍摄的图像，形成了裂缝检测数据集，并使用AlexeyAB版本的Darknet在个人计算机上进行训练，生成权重文件。同时，将NNpack包加速的AlexeyAB版本Darknet部署在Raspberry Pi 4B上，并进行了裂缝检测实验。通过与原始YOLOv4-Tiny算法的比较，证实了Lite YOLOv4-Tiny算法具有速度快、误检率低等特点。创新点主要集中在网络结构简单、网络层数少以及特征前向传输早，从而防止过拟合。实验结果表明，新的Lite神经网络在性能上显著优于原始YOLOv4-Tiny网络。这一研究成果对于提高高层建筑裂缝检测的效率和准确性具有重要意义，有望推动智能爬墙机器人在建筑维护领域的广泛应用。

关键词: 智能检测，深度网络，边缘设备，树莓派

Abstract: Conforming to the rapidly increasing market demand of crack detection for tall buildings, the idea of integrating deep network technology into wall climbing robot for crack detection is put forward in this paper. Taking the dependence and hardware requirements when deployed on such edge devices as Raspberry Pi into consideration, the Darknet neural network is selected as the basic framework for detection. In order to improve the inference efficiency on edge devices and avoid the possible premature over-fitting of deep networks, the lite YOLOv4-tiny algorithm is then improved from the original YOLOv4-tiny algorithm and its structure is illustrated using Netron accordingly. The images downloaded from Internet and taken from the buildings in campus are processed to form crack detection data sets, which are trained on personal computer with the AlexeyAB version of Darknet to generate weight files. Meanwhile, the AlexeyAB version of Darknet accelerated by NNpack package is deployed on Raspberry Pi 4B, and the crack detection experiments are carried out. Some characteristics, e.g., fast speed and lower false detection rate of the lite YOLOv4-tiny algorithm, are confirmed by comparison with those of original YOLOv4-tiny algorithm. The innovations of this paper focus on the simple network structure, fewer network layers, and earlier forward transmission of features to prevent over-fitting, showing the new lite neural network exceeds the original YOLOv4-tiny network significantly.

Key words: intelligent detection, deep network, edge device, Raspberry Pi

中图分类号:

TP391.4

宋立博a，费燕琼b. 新型Lite YOLOv4-Tiny算法及其在裂纹智能检测中的应用[J]. J Shanghai Jiaotong Univ Sci, 2024, 29(3): 528-536.

SONG Libo^a (宋立博), FEI Yanqiong^b (费燕琼). New Lite YOLOv4-Tiny Algorithm and Application on Crack Intelligent Detection[J]. J Shanghai Jiaotong Univ Sci, 2024, 29(3): 528-536.

参考文献

[1] TANG D L, LONG Z Y, TANG Y J, et al. Complete coverage path planning of oil tank detection wall climbing robot [J]. Chinese Journal of Engineering Design, 2020, 27(2): 162-171 (in Chinese).
[2] YANG Z L, WAN Y, WANG Y, et al. Analysis of safety and stability of wall-climbing robot [J]. Science Technology and Engineering, 2020, 20(15): 6113-6121 (in Chinese).
[3] SHU J J, WANG Q, HE Y Y, et al. Design of weld identification and inspection wall-climbing robot for large spherical tank container [J]. Hot Working Technology, 2020, 49(11): 127-131 (in Chinese).
[4] XIAO R H, CHENG Y X, JIANG Z Z, et al. Design of adsorption mechanism for rust-removing wall-climbing robot on the oil tank’s inner wall [J]. Journal of Machine Design, 2019, 36(7): 21-26 (in Chinese).
[5] LYU Z, ZHANG C, ZHONG G, et al. Kinematics analysis and simulation of a quadruped magnetic adsorption wall-climbing robot [J]. Advanced Engineering Sciences, 2020, 52(2): 121-129 (in Chinese).
[6] WANG W, GAO C, ZHANG X, et al. Analysis and simulation of permanent magnet adsorption characteristics of lightweight wall-climbing robot [J]. Manufacturing Automation, 2020, 42(12): 23-27 (in Chinese).
[7] LIU Y W, WANG L M, LIU S W, et al. Design and analysis of an inchworm-inspired wall-climbing robot [J]. Journal of Mechanical Transmission, 2019, 43(8): 87-91 (in Chinese).
[8] ZHUANG Y, TENG H, XU T, et al. Obstacle avoidance control of wall-climbing robot based on degraded fuzzy algorithms [J]. Science Technology and Engineering, 2020, 20(19): 7729-7736 (in Chinese).
[9] LU J, ZHU H, LIANG J, et al. Global path planning for a biped wall-climbing robot in 3D wall environment [J]. Journal of Harbin Institute of Technology, 2020, 52(1): 148-155 (in Chinese).
[10] TENG H, ZHUANG Y, DENG S, et al. Doubleloop trajectory tracking control of wall-climbing robot based on global stability [J]. Computer Engineering and Design, 2020, 41(9): 2636-2642 (in Chinese).
[11] REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection [C]//2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 779-788.
[12] SUN X, HAO H, LIU Y, et al. Application of YOLOv4 in power inspection target detection [J]. Modern Information Technology, 2020, 4(20): 115-117 (in Chinese).
[13] TANG X, HUANG J, FENG J, et al. Image segmentation and defect detection of insulators based on U-net and YOLOv4 [J]. Journal of South China Normal University (Natural Science Edition), 2020, 52(6): 15-21 (in Chinese).
[14] DU X, CHEN D, LIU H, et al. Real-time hand tracking based on YOLOv4 model and Kalman filter [J]. The Journal of China Universities of Posts and Telecommunications, 2021, 28(3): 86-94.
[15] BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv4: Optimal speed and accuracy of object detection [DB/OL]. (2020-04-23). https://arxiv.org/abs/2004.10934.
[16] LI B, WANG C, WU J, et al. Surface defect detection of aeroengine components based on improved YOLOv4 algorithm [J]. Laser & Optoelectronics Progress, 2021, 58(14): 406-415 (in Chinese).
[17] XIE Y, ZHANG P. Small target detection of transmission line based on improved YOLOv4 [J]. Foreign Electronic Measurement Technology, 2021, 40(2): 47-51 (in Chinese).
[18] CAI S, SUN Z, LIU H, et al. Real-time detection methodology for obstacles in orchards using improved YOLOv4 [J]. Transactions of the Chinese Society of Agricultural Engineering, 2021, 37(2): 36-43 (in Chinese).
[19] ZHANG K, LUO X, SUN Z, et al. Generator stator surface defect detection algorithm based on lightweight YOLOv4 [J]. Computer and Digital Engineering, 2021, 49(4): 686-691 (in Chinese).
[20] WANG B, LE H, LI W, et al. Mask detection algorithm based on improved YOLO lightweight network [J]. Computer Engineering and Applications, 2021, 57(8): 62-69 (in Chinese).
[21] CAO Y, GAO Y. Lightweight beverage recognition network based on Ghostnet residual structure [J]. Computer Engineering, 2022, 48(3): 310-314 (in Chinese).