基于YOLOv5的路面病害图像识别方法研究
Research on Recognition Method of Road Surface Disease Image Based on YOLOv5
刘德坤 1刘肖亮 1张帅2
作者信息
- 1. 湖南联智科技股份有限公司,湖南长沙 410200
- 2. 湖南大学,湖南长沙 410082
- 折叠
摘要
在公路养护中,道路路面病害现有的检测方法多为自动化采集、人工识别,这极大地降低了公路的养护效率.为提高公路路面病害识别效率,提出了一种基于改进YOLOv5的路面病害图像识别算法,在YOLOv5的主干中引入CA注意力机制及SPPCSPC结构,CA提高了模型的感受野,精确定位目标的感兴趣区域,SPPCSPC结构使算法能适应不同的分辨率图像,提高识别速度;在锚框上,将YOLOv5的k-means改为k-means++,锚框更符合数据集中真实标记框大小;试验结果表明,在9种病害类型、56 879张病害图像的数据集中,所提出的方法相比于原模型在精度上提高了 8.1％,在检测速度上提高了 12.8％,与FasterR-CNN、YOLOv3等方法相比均有所提高.
Abstract
In highway maintenance,the existing methods for detecting road surface diseases mostly rely on automated data collection and manual identification,which greatly reduces the efficiency of road maintenance.To improve the efficiency of road surface disease recognition,this paper proposes an improved YOLOv5-based algorithm for road surface disease image recognition.The CA mechanism and SPPCSPC structure are introduced into the backbone of YOLOv5.The CA mechanism enhances the receptive field of the model and accurately localizes the regions of interest.The SPPCSPC structure enables the algorithm to adapt to different image resolutions and improves the recognition speed.In terms of anchor boxes,the k-means algorithm in YOLOv5 is replaced with k-means++to make the anchor boxes better fit the sizes of real annotated boxes in the dataset.Experimental results show that compared to the original model,the proposed method achieves a 8.1％improvement in accuracy and a 12.8％improvement in detection speed in a dataset consisting of 56 879 images of 9 types of road surface diseases.It also outperforms methods such as Faster R-CNN and YOLOv3.
关键词
路面病害/目标检测/YOLOv5/注意力机制Key words
road surface diseases/object detection/YOLOv5/attention mechanism引用本文复制引用
基金项目
湖南省自然科学基金项目(2022JJ30155)
出版年
2024
NETL
NSTL
万方数据