基于改进YOLO v5的混凝土桥梁裂缝轮廓提取研究  

Research on Concrete Bridge Crack Outline Extraction Based on Improved YOLO v5

作　　者：侯珂 王翔[2] 刘文江 程文党 

机构地区：[1]山东交通学院,轨道交通学院,山东 济南 [2]济南城建集团有限公司,山东 济南 [3]山东交通学院,航空学院,山东 济南

出　　处：《交通技术》2024年第1期67-72,共6页Open Journal of Transportation Technologies

摘　　要：裂缝是混凝土桥梁的常见病害之一,当前主要依靠人工实施检测,存在效率低和准确率低等问题,研究自动化、智能化的裂缝检测方法一直是行业的热点问题。文章提出了一种改进YOLO v5的深度学习方法可以解决当前桥梁裂缝检测误检率高实时性差的问题。首先,我们将YOLO v5中主干网络替换为残差网络,进而加深网络的深度,通过残差连接使网络学习更复杂的特征;然后,在颈部网络结构的最后一层使用GCNet网络结构,通过全局相关层,将全局信息编码到每个位置的特征中,可以更好地捕捉裂缝的轮廓和细节。实验表明,该方法能够快速准确地检测出裂缝,且误检率较低,精确率可达到67.1%,比原始网络提高了9.2%,在混凝土桥梁裂缝轮廓提取任务中展现了出色的性能。Cracks are a common disease in concrete bridges. Currently, crack detection mainly relies on manual inspection, which has issues such as low efficiency and accuracy. The research on automated and intelligent crack detection methods has always been a hot topic in the industry. This paper proposes an improved deep learning method based on YOLO v5 to address the high false detection rate and poor real-time performance in current bridge crack detection. Firstly, we replace the backbone network in YOLO v5 with a residual network to deepen the network and make it learn more complex features through residual connections. Secondly, we use the GCNet network structure in the last layer of the neck network structure. Through the global contextual layer, global information is encoded into the features of each position, which helps better capture the contours and details of the cracks. Experiments show that this method can quickly and accurately detect cracks with a low false detection rate. The precision rate can reach 67.1%, which is 9.2% higher than the original network, demonstrating excellent performance in the task of concrete bridge crack outline extraction.

关 键 词：桥梁裂缝 实例分割 YOLO v5s 残差网络 GCNet 

分 类 号：TU7[建筑科学—建筑技术科学]