融合注意力机制和旋转边界框的遥感影像桥梁自动检测  

Automatic Bridge Detection in Remote Sensing Image by Combining Attention Mechanism and Rotating Boundary Frame

作　　者：单韵宇 黄亮[1,3] SHAN Yunyu;HUANG Liang(Faculty of Land Resource Engineering,Kunming University of Science and Technology,Kunming 650093,China;Kunming Real Estate Registration Survey Center,Kunming 650200,China;Key Laboratory of Plateau Remote Sensing,Yunnan Provincial Department of Education,Kunming 650093,China)

机构地区：[1]昆明理工大学国土资源工程学院,昆明650093 [2]昆明市不动产权籍调查中心,昆明650200 [3]云南省教育厅高原遥感重点实验室,昆明650093

出　　处：《遥感信息》2024年第6期122-129,共8页Remote Sensing Information

基　　金：国家自然科学基金(42361054);云南省基础研究计划项目(202201AT070164)。

摘　　要：桥梁对国家发展、百姓生活起着举足轻重的作用,采用高分辨率遥感影像对桥梁目标进行自动检测具有重要意义。但现有基于YOLOv5的检测算法存在水平边界框重叠、包含无关信息、识别精度低等问题。针对这些问题,文章提出了一种改进YOLOv5的桥梁检测方法,通过旋转边界框精确地描述桥梁的形状和姿态,并结合Transformer、BiFPN算法提高网络精度,最终形成一种具有方向性、高精度的遥感影像桥梁自动检测算法。首先,在水平边界框的基础上增加了θ角度参数,得到旋转边界框;然后,在YOLOv5的backbone中添加Transformer模块,通过多头注意力层使得在复杂场景下有效捕捉目标信息;接着,将YOLOv5的特征融合网络修改为BiFPN结构,更好地挖掘影像信息;最后,采用精确度(precision)、召回率(recall)、平均精度值(mean average precision,mAP)、帧率(frames per second,FPS)评估实验结果。该方法实验结果较目前常见深度学习方法表现更为优异,mAP达到了99.47%,较YOLOv5提高了近8%,较YOLOv8提高了近1%;FPS达到了56,较YOLOv8提升了14,证明了所提方法的有效性。Bridges play a pivotal role in national development and people’s lives.It is of great significance to use high-resolution remote sensing images for automatic detection of bridge.However,the existing detection algorithms based on YOLOv5 have some problems such as horizontal bounding box overlapping,containing irrelevant information and low recognition accuracy.To solve these problems,this paper proposes an improved YOLOv5 detection method,which accurately describes the shape and attitude of the bridge by rotating the boundary frame,and combines Transformer and BiFPN algorithms to improve the network accuracy.Finally,a directional and high-precision bridge automatic detection algorithm is formed which is more suitable for remote sensing images.Firstly,θ angle parameter is added to the horizontal bounding box to get the rotating bounding box.Then,a Transformer module is added to the backbone of YOLOv5 to effectively capture target information in complex scenes through multiple attention layers.Next,the feature fusion network of YOLOv5 is modified to BiFPN structure to better mine image information.Finally,precision,recall and mean average precision(mAP)are used to evaluate the experimental results.The experimental results of the proposed method achieve 99.47%mAP,which is nearly 8%higher than that of YOLOv5 and 1%higher than that of YOLOv8.FPS is 56,up 14 from YOLOv8.The performance is better than that of common deep learning methods at present,which proves the effectiveness of the proposed method.

关 键 词：深度学习 目标检测 YOLOv5 桥梁检测 TRANSFORMER 锚框 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]