%0 Journal Article
%T YOLOv5-W桥梁裂纹实时检测算法<br>YOLOv5-W Bridge Crack Real-Time Detection Algorithm
%A 史鸣凤
%A 董琴
%A 郭乃瑄
%A 许铭
%J Modeling and Simulation
%P 290-303
%@ 2324-870X
%D 2024
%I Hans Publishing
%R 10.12677/MOS.2024.131028
%X 桥梁的定期检修与维护，是保证交通安全、保障桥梁使用年限的重要措施。在桥梁的诸多损伤中，桥梁裂纹是最为普遍的损伤。针对现有裂纹研究中模型计算量大、实时性较差、需要大数据集训练等问题，提出了一种基于YOLOv5改进的YOLOv5-W模型。对YOLOv5的损失函数和颈部网络的设计进行了优化，使用Wise IoU，通过离群度来衡量检测框质量，做出最合适的梯度增益分配，提高检测精度。使用轻量化颈部Slim-Neck设计缩小模型的参数量，提高检测速度。在小数据集Crack400上验证表明，改进模型准确度为98.7，均值平均精度(mean average precision, MAP)为98.5，检测速度为47.958 FPS，模型参数减少到14.5 GPLOPs。相较于原始的YOLOv5，平均精度提升3%，FPS提升12，模型大小减小1.3 GFLOPs。<br />
Regular maintenance and repair of bridges is an important measure to ensure traffic safety and to safeguard the service life of bridges. Among the many damages to bridges, bridge cracks are the most common damage. A YOLOv5-W model based on an improved YOLOv5 is proposed to address the problems of large model computation, poor real-time performance and the need for large data sets for training in existing cracking studies. The loss function of YOLOv5 and the design of the neck network are optimized, and Wise IoU is used to measure the quality of the detection frame by the outlier degree to make the most appropriate gradient gain assignment and improve the detection accuracy. A lightweight neck Slim-Neck design was used to reduce the number of parameters in the model and improve detection speed. Validation on the small dataset Crack400 shows that the im-proved model has an accuracy of 0.987, a mean average precision (mAP) of 0.985, a detection speed of 47.958 FPS and a reduction in model parameters to 14.5 GPLOPs. The model size was reduced by 1.3 GFLOPs.
%K 目标检测，桥梁裂纹，深度学习，YOLOv5<br>Object Detection
%K Bridge Crack
%K Deep Learning
%K YOLOv5
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=79163