结合三维无参数注意力机制的隧道裂缝检测方法
Tunnel Crack Detection Method Based on Three-Dimensional Parameter-Free Attention Mechanisms
武斌 1于双玲 1陈杨杨 1赵洁1
作者信息
- 1. 天津城建大学计算机与信息工程学院,天津 300384
- 折叠
摘要
为进一步研究现有公路隧道裂缝检测算法中对裂缝特征提取不充分、抗干扰能力弱而导致漏检以及检测速度慢等问题,提出一种基于改进YOLOv5 的隧道裂缝检测方法.首先,在主干网络中结合高效的三维无参数注意力机制,引入残差模块C3SM,用于增强深浅层特征信息之间的交互,在优化计算复杂度的同时增强网络特征提取能力;其次,在特征金字塔中采用一种新的特征融合网络结构,整合相邻层的特征图,能较好地保留裂缝边缘信息,在保证语义信息不受损失的同时加快模型的检测速度;最后,采用位置损失函数WIoU优化遮挡和重叠目标的检测效果.为验证该方法的有效性,在Tunnel-crack和湖州隧道裂缝数据集上进行大量试验,结果表明:所提出检测方法的精度与速度分别达到88.7%、103.5 帧/s和85.1%、99.4 帧/s,相比大多数高性能目标检测器具有更高的识别准确率,并能满足隧道裂缝检测的要求.
Abstract
The existing crack detection algorithms for highway tunnels exhibit drawbacks such as inadequate feature extraction and weak anti-interference ability,leading to missed detections and slow detection speeds.Therefore,a tunnel crack detection method based on improved YOLOv5 is proposed.First,the residual module C3SM is introduced into the backbone network along with an efficient three-dimensional parameter-free attention mechanism to enhance the interaction between deep and shallow feature information.This enhances the feature extraction capability of the network while optimizing computational complexity.Second,a new feature fusion network structure is employed in the feature pyramid to integrate the feature maps of neighboring layers,better preserving crack edge information and accelerating model detection without loss of semantic information.Finally,a position loss function,WIoU,is used to optimize the detection of occluded and overlapped targets.To validate the effectiveness of this method,extensive experiments are conducted on the Tunnel-crack and Huzhou tunnel crack datasets.The results show that the proposed method achieves 88.4%and 103.5 frames/s and 85.1%and 99.4 frames/s in terms of accuracy and speed,respectively.These results demonstrate a higher crack detection accuracy than most high-performance target detectors.
关键词
隧道工程/裂缝检测/注意力机制/YOLOv5/WIoUKey words
tunnel engineering/crack detection/attention mechanism/YOLOv5/WIoU引用本文复制引用
基金项目
天津市企业科技特派员项目(19JCTPJC47200)
天津市重点研发计划科技支撑重点项目(19YFZCGX00130)
天津市研究生科研创新项目(2021YJSS351)
出版年
2024
NETL
NSTL
万方数据